# Science Applications of Generative Neural Networks

Machine learning is a common tool used in all areas of science. Applications range from simple regression models used to explain the behavior of experimental data to novel applications of deep learning. One area that has emerged in the last few years is the use of generative neural networks to produce synthetic samples of data that fit the statistical profile of real data collections. Generative models are among the most interesting deep neural networks and they abound with applications in science. The important property of all generative networks is that if you train them with a sufficiently, large and coherent collection of data samples, the network can be used to generate similar samples. But when one looks at the AI literature on generative models, one can come away with the impression that they are, at best, amazing mimics that can conjure up pictures that look like the real world, but are, in fact, pure fantasy. So why do we think that they can be of value in science? There are a several reasons one would want to use them. One reason is that the alternative method to understand nature may be based on a simulation that is extremely expensive to run. Simulations are based on the mathematical expression of a theory about the world. And theories are often loaded with parameters, some of which may have known values and others we can only guess at. Given these guesses, the simulation is the experiment: does the result look like our real-world observations? On the other hand, generative models have no explicit knowledge of the theory, but they do an excellent job of capturing the statistical distribution of the observed data. Mustafa Mustafa from LBNL states,

“We think that when it comes to practical applications of generative models, such as in the case of emulating scientific data, the criterion to evaluate generative models is to study their ability to reproduce the characteristic statistics which we can measure from the original dataset.” (from Mustafa, et. al arXiv:1706.02390v2 [astro-ph.IM] 17 Aug 2018)

Generated models can be used to create “candidates” that we can use to test and fine-tune instruments designed to capture rare events. As we shall see, they have also been used to create ‘feasible’ structures that can inform us about possibilities that were not predicted by simulations. Generative models can also be trained to generate data associated with a class label and they can be effective in eliminating noise. As we shall see this can be a powerful tool in predicting outcomes when the input data is somewhat sparse such as when medical records have missing values.

Flavors of Generative Models

There are two main types of GMs and, within each type, there are dozens of interesting variations. Generalized Adversarial Networks (GANs) consist of two networks, a discriminator and a generator (the bottom part of Figure 1 below). Given a training set of data the discriminator is trained to distinguish between the training set data and fake data produced by the generator. The generator is trained to fool the discriminator. This eventually winds up in a generator which can create data that perfectly matches the data distribution of the samples. The second family are autoencoders. Again, this involved two networks (top in figure below). One is designed to encode the sample data into a low dimensional space. The other is a decoder that takes the encoded representation and attempts to recreate it. A variational autoencoder (VAEs) is one that forces the encoded representations to fit into a distribution that looks like the unit Gaussian. In this way, samples from this compact distribution can be fed to the decoder to generate new samples.

Figure 1.

Most examples of generative networks that are commonly cited involve the analysis of 2-D images based on the two opposing convolutional or similar networks.  But this need to be the case. (see “Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space” by Anh Nguyen, et. al. arXiv:1612.00005v2  [cs.CV]  12 Apr 2017).

One fascinating science example we will discuss in greater detail later is by Shahar Harel and Kira Radinsky.  Shown below (Figure 2), it is a hybrid of a variational autoencoder with a convolutional encoder and recurrent neural network decoder for generating candidate chemical compounds.

Figure 2.  From Shahar Harel and Kira Radinsky have a different approach in “Prototype-Based Compound Discovery using Deep Generative Models” (http://kiraradinsky.com/files/acs-accelerating-prototype.pdf ).

## Physics and Astronomy

In statistical mechanics, Ising models provide a theoretical tool to study phase transitions in materials. The usual approach to study the behavior of this model at various temperatures is via Monte Carlo simulation. Zhaocheng Liu, Sean P. Rodrigues and Wenshan Cai from Georgia Tech in their paper “Simulating the Ising Model with a Deep Convolutional Generative Adversarial Network” (arXiv: 1710.04987v1 [cond-mat.dis-nn] 13 Oct 2017). The Ising states they generate from their network faithfully replicate the statistical properties of those generated by simulation but are also entirely new configurations not derived from previous data.

Astronomy is a topic that lends itself well to applications of generative models. Jeffrey Regier et. al. in “Celeste: Variational inference for a generative model of astronomical images” describe a detailed multi-level probabilistic model that considers both the different properties of stars and galaxies at the level of photons recorded at each pixel of the image. The purpose of the model is to infer the properties of the imaged celestial bodies. The approach is based on a variational computation similar to the VAEs described below, but far more complex in terms of the number of different modeled processes. In “Approximate Inference for Constructing Astronomical Catalogs from Images, arXiv:1803.00113v1 [stat.AP] 28 Feb 2018”, Regier and collaborators take on the problem of building catalogs of objects in thousands of images. For each imaged object there are 9 different classes of random variables that must be inferred. The goal is to compute the posterior distribution of these unobserved random variables conditional on a collection of astronomical images. They formulated a variational inference (VI) model and compared that to a Markov chain monte carlo (MCMC) method. MCMC proved to be slightly more accurate in several metrics but VI was very close. On the other hand, the variational method was 1000 times faster. It is also interesting to note that the computations were done on a Cori, the DOE supercomputer and the code was written in Julia.

Cosmological simulation is used to test our models of the universe. In “Creating Virtual Universes Using Generative Adversarial Networks” (arXiv:1706.02390v2 [astro-ph.IM] 17 Aug 2018) Mustafa Mustafa, et. al. demonstrate how a slightly-modified standard GAN can be used generate synthetic images of weak lensing convergence maps derived from N-body cosmological simulations. The results, shown in Figure 3 below, illustrate how the generated images match the validation tests. But, what is more important, the resulting images also pass a variety of statistical tests ranging from tests of the distribution of intensities to power spectrum analysis. They have made the code and data available at http://github.com/MustafaMustafa/cosmoGAN . The discussion section at the end of the paper speculates about the possibility of producing generative models that also incorporate choices for the cosmological variable that are used in the simulations.

Figure 3.  From  Mustafa Mustafa, et. al. “Creating Virtual Universes Using Generative Adversarial Networks” (arXiv:1706.02390v2 [astro-ph.IM] 17 Aug 2018

## Health Care

Medicine and health care are being transformed by the digital technology. Imaging is the most obvious place where we see advanced technology.  Our understanding of the function of proteins and RNA has exploded with high-throughput sequence analysis. Generative methods are being used here as well. Reisselman, Ingraham and Marks in “Deep generative models of genetic variation capture mutation effects” consider the problem of how mutations to a protein disrupt it function. They developed a version of a variational autoencoder they call DeepSequence that is capable if predicting the likely effect of mutations as they evolve.

Another area of health care that is undergoing rapid change is health records. While clearly less glamourous than RNA and protein analysis, it is a part of medicine that has an impact on every patient. Our medical records are being digitized at a rapid rate and once in digital form, they can be analyzed by many machine learning tools. Hwang, Choi and Yoon in “Adversarial Training for Disease Prediction from Electronic Health Records with Missing Data” (arXiv:1711.04126v4 [cs.LG] 22 May 2018) address two important problems. First, medical records are often incomplete. They have missing value because certain test results were not correctly recorded. The process of translating old paper forms to digital artifacts can introduce additional errors. Traditional methods of dealing with this are to introduce “zero” values or “averages” to fill the gaps prior to analysis, but this is not satisfactory. Autoencoders have been shown to be very good at removing noise from data (see https://towardsdatascience.com/how-to-reduce-image-noises-by-autoencoder-65d5e6de543). Hwang and his colleagues applied this to medical records. The second thing they have done is to use a GAN to predict the disease from the “corrected” record. The type of GAN they use is an “AC-GAN” (see https://arxiv.org/pdf/1610.09585.pdf) which incorporates a class label with each training item. This allows a class label along with the random latent variable as input to force the generator to create an output similar to training elements of that class. A byproduct is a discriminator that can tell if an input has the correct class label. In their case the they are interested in if a given medical record may predict the occurrence of a tumor or not. Of course, this is far from usable as a sole diagnostic in a clinical setting, but it is a very interesting technology.

## Drug Design

One exciting application of these techniques is in the design of drugs. The traditional approach is high throughput screening in which large collections of chemicals are tested against potential targets to see if any have potential therapeutic effects. Machine learning techniques have been applied to the problem for many years, but recently various deep learning method have shown surprisingly promising results. One of the inspirations for the recent work has been the recognition that molecular structures have properties similar to natural language (see Cadeddu, A, et. al.. Organic chemistry as a language and the implications of chemical linguistics for structural and retrosynthetic analyses. Angewandte Chemie 2014, 126.) More specifically, there are phrases and grammar rules in chemical compounds that have statistical properties not unlike natural language. There is a standard string representation called SMILES that an be used to illustrate these properties. SMILES representations describe atoms and their bonds and valences based on a depth-first tree traversal of a chemical graph. In modern machine learning, language structure and language tasks such as machine natural language translation are aided using recurrent neural networks. As we illustrated in our book, an RNN trained with lots of business news text is capable of generating realistic sounding business news headlines from a single starting word. However close inspection reveals that the content is nonsense. However, there is no reason we cannot apply RNNs to SMILES string to see if they can generate new molecules. Fortunately, there are sanity tests that can be applied to generated SMILES string to filter out the meaningless and incorrectly structured compounds. This was done by a team at Novartis (Ertl et al. Generation of novel chemical matter using the LSTM neural network, arXiv:1712.07449) who demonstrated that these techniques could generate billions of new drug-like molecules. Anvita Gupta, Alex T. Muller, Berend J. H. Huisman, Jens A. Fuchs, Petra Schneid and Gisbert Schneider applied very similar ideas to “Generative Recurrent Networks for De Novo Drug Design”. They demonstrated that if they started with fragments of a drug of interest they could use the RNN and transfer learning to generate new variations that can may be very important. Another similar result is from Artur Kadurin, et. al. in “druGAN: An Advanced Generative Adversarial Autoencoder Model for de Novo Generation of New Molecules with Desired Molecular Properties in Silico.

Shahar Harel and Kira Radinsky have a different approach in “Prototype-Based Compound Discovery using Deep Generative Models”. There model is motivated by a standard drug discovery process which involves start with a molecule, called a prototype, with certain known useful properties and making modifications to it based on scientific experience and intuition. Harel and Radinsky designed a very interesting Variational Autoencoder shown in figure 2 above. As with several others the start with a SMILES representation of the prototype. The first step is an embedding space is generated for SMILES “language”. The characters in the prototype sequence are imbedded and fed to a layer of convolutions that allow local structures to emerge as shorter vectors that are concatenated, and a final all-to-all layer is used to generate sequence of mean and variance vectors for the prototype. This is fed to a “diversity layer” which add randomness.

The decoder is an LSTM-based recurrent network which generates the new molecule. The results they report are impressive. In a one series of experiments they took as prototypes compounds from drugs that were discovered years ago, and they were able to generate more modern variations that are known to be more powerful and effective. No known drugs were used in the training.

## Summary

These are only a small sample of the research on the uses of Generative Neural networks in science.   We must now return to the question posed in the introduction:  When are these applications of neural networks advancing science?    We should first ask the question what is the role of ‘computational science’?  It was argued in the 1990s that computing and massive computational simulation had become the third paradigm of science because it was the only way to test theories for which it was impossible to design physical experiments.   Simulations of the evolution of the universe is a great example.    These simulations allowed us to test theories because they were based on theoretical models.  If the simulated universe did not look much like our own, perhaps the theory is wrong.   By 2007 Data Science was promoted as the fourth paradigm.   By mining the vast amounts of the data we generate and collect, we can certainly validating or disproving scientific claims.    But when can a network generating synthetic images qualify as science?  It is not driven by theoretical models.   Generative models can create statistical simulations that are remarkable duplicates of the statistical properties of natural systems.   In doing so they provide a space to explore that can stimulate discovery.   There are three classes of why this can be important.

• The value of ‘life-like’ samples. In “3D convolutional GAN for fast Simulation” F. Carminati, G.  Khattak, S.  Vallecorsa make the argument that designing and testing the next generation of sensors requires test data that is too expensive to compute with simulation.  But a well-tuned GAN is able to generate the test cases that fit the right statistical model at the rate needed for deployment.
• Medical records-based diagnosis. The work on medical records described above by Hwang shows that using a VAE to “remove noise” is statistically superior to leaving them blank or filling in averages.   Furthermore their ability to predict disease is extremely promising as science.
• Inspiring drug discovery. The work of Harel and Radinsky show us that a VAE can expand the scope of potential drug for further study.   This is an advance in engineering if not science.

Can it replace simulation for validating models derived from theory?  Generative neural networks are not yet able to replace simulation.   But perhaps theory can evolve so that it can be tested in new ways.

# Part 2. Generative Models Tutorial

Generative Models are among the most interesting deep neural networks and they abound with applications in science. There are two main types of GMs and, within each type, several interesting variations. The important property of all generative networks is that if you train them with a sufficiently, large and coherent collection of data samples, the network can be used to generate similar samples. The key here is the definition of ‘coherent’. One can say the collection is coherent if when you are presented with a new example, it should be a simple task to decide if it belongs to the collection or not. For example, if the data collection consists entirely of pictures of cats, then a picture of a dog should be, with reasonably high probability, easily recognized as an outlier and not a cat. Of course, there are always rather extreme cats that would fool most casual observers which is why we must describe our collect of objects in term of probability distributions. Let us assume our collection c is naturally represented embedded in $R^m$ for some m. For example, images with m pixels or other high dimensional instrument data. A simple way to think about a generative model is a mathematical device that transforms samples from a multivariant normal distribution $\mathcal{N}^k (0,1)$  into so that they look like they come from the distribution $p_c(X)$ for our collection c. Think of it as a function

$Gen: v \sim \mathcal{N}^k (0,1) \to R^m$

Another useful way to say this is to build another machine we can call a discriminator

$Disc: R^m \to [0,1]$

such that for $X \in R^m, Disc(X)$   is probability that X is in the collection c. To make this more “discriminating” let us also insist that $Disc(Gen(v)=0$.  In other word, the discriminator is designed to discriminate between the real c objects and the generated ones. Of course, if the Generator is really doing a good job of imitating  then the discriminator with this condition would be very hard to build.  In this case we would expect $Disc(Gen(v)=1$.

were introduced by Goodfellow et, al (arXiv:1406.2661) as a way to build neural networks that can generate very good examples that match the properties of a collection of objects.  It works by designed two networks:  one for the generator and one for the discriminator. Define  to be the distribution of latent variables that the generator will map to the collection space. The idea behind the paper is to simultaneously design the discriminator and the generator as a two-player min-max game.

The discriminator is being trained to recognize object from c (thereby reducing   for  ) and pushing  to zero for .   The resulting function

Represents the min-max objective for the Discriminator.

On the other hand, the generator wants to push  to 1 thereby maximizing
.   To do that we minimize

.

There are literally dozens of implementations of GANs in Tensorflow or Karas on-line.   Below is an example from one that works with 40×40 color images.   This fragment shows the step of setting up the training optimization.

#These two placeholders are used for input into the generator and discriminator, respectively.
z_in = tf.placeholder(shape=[None,128],dtype=tf.float32) #Random vector
real_in = tf.placeholder(shape=[None,40,40,3],dtype=tf.float32) #Real images
Gz = generator(z_in) #Generates images from random z vectors
Dx = discriminator(real_in) #Produces probabilities for real images
Dg = discriminator(Gz,reuse=True) #Produces probabilities for generator images
#These functions together define the optimization objective of the GAN.
d_loss = -tf.reduce_mean(tf.log(Dx) + tf.log(1.-Dg)) #This optimizes the discriminator.
g_loss = -tf.reduce_mean(tf.log(Dg)) #This optimizes the generator.
tvars = tf.trainable_variables()
#The below code is responsible for applying gradient descent to update the GAN.

#Only update the weights for the discriminator network.
#Only update the weights for the generator network.



We tested this with a very small collection of images of galaxies found on the web.  There are three types: elliptical, spiral and barred spiral.  Figure 4 below shows some high-resolution samples from the collection.

(Note:  the examples in this section use pictures of galaxies, but , in terms of the discussion in the previous part of this article, these are illustrations only.  There are no scientific results; just algorithm demonstrations. )

Figure 4.  Sample high-resolution galaxy images

We reduced the images to 40 by 40 and trained the GAN on this very small collection.  Drawing samples at random from the latent z-space we can now generate synthetic images.  The images we used here are only 40 by 40 pixels, so the results are not very spectacular.  As shown below, the generator is clearly able to generate elliptical and spiral forms.  In the next section we work with images that are 1024 by 1024 and get much more impressive results.

Figure 5.   Synthetic Galaxies produced by the GAN from 40×40 images.

## Variational Autoencoders

The second general category generative models are based on variational autoencoders. An autoencoder transforms our collection of object representations into a space of much smaller dimension in such a way so that that representation can be used to recreate the original object with reasonably high fidelity. The system has an encoder network that creates the embedding in the smaller space and a decoder which uses that representation to regenerate an image as shown below in Figure 6.

Figure 6. Generic Autoencoder

In other words, we want  to approximate  for each i in an enumeration of our collection of objects.  To train our networks we simply want to minimize the distance between   and  for each i.   If we further set up the network inputs and outputs so that they are in the range [0, 1] we can model this as a Bernouli distribution so cross entropy is a better function to minimize.  In this case the cross entropy can be calculated as

(see http://www.godeep.ml/cross-entropy-likelihood-relation/  for a derivation)

A variational autoencoder differs from a general one in that we want the generator to create an embedding that is very close to a normal distribution in the embedding space.  The way we do this is to make the encoder force the encoding into a representation consisting of a mean and standard deviation.  To force it into a reasonably compact space we will force our encoder to be as close to   as possible. To do that we need a way to measuree how far a distribution p is from a Gaussian q. That is given by the Kullback-Leibler divergence which measures now many extra bits (or ‘nats’) are needed to convert an optimal code for distribution q into an optimal code for distribution p.

If both p and q are gaussian this is easy to calculate (thought not as easy to derive).

In terms of probability distributions we can think of our encoder as  where x is a training image. We are going to assume   is normally distributed and let  be  parameterized by     .  Computing   is now easy. We call this the Latent Loss and it is

We now construct our encoder to produce  and  .  To sample from this latent space, we simply draw from and transform it into the right space.   Our encoder and decoder networks can now be linked as follows.

the loss function is now the sum of two terms:

Note: there is a Baysian approach to deriving this.  see https://jaan.io/what-is-variational-autoencoder-vae-tutorial   for an excellent discussion.

One of the interesting properties of VAEs is that they do not require massive data sets to converge.   Using our simple galaxy photo collection we trained a simple VAE.  The results showing the test inputs and the reconstructed images are illustrated below.

Figure 7.   test input and reconstruction from the galaxy image collection.   These images are 1024×1024.

Using encodings of five of the images we created a path through the latent space to make the gif movie that is shown below.  While not every intermediate “galaxy” looks as good as some of the originals, it does present many reasonable “synthetic” galaxies that are on the path between two real ones.

Figure 8.  image from the “movie”

The notebook for this autoencoder is available as html (see https://s3.us-east-2.amazonaws.com/a-book/autoencode-galaxy.html) and as a jupyter notebook (see https://s3.us-east-2.amazonaws.com/a-book/autoencode-galaxy.ipynb )  The compressed tarball of the galaxy images is here: https://s3.us-east-2.amazonaws.com/a-book/galaxies.tar.gz.

## acGANs

The generative networks described above are just the basic variety.    One very useful addition is the Auxiliary Classifier GAN.    An acGAN allows you to incorporate knowledge about the class of the objects in your collection into the process.   For example, suppose you have labeled images such as all pictures of dogs are labeled “dog” and all pictures of cats have the label “cat”.    The original paper on this subject “Conditional Image Synthesis with Auxiliary Classiﬁer GANs” by Oden, Olah and Shlens  shows how a GAN can be modified so that the generator can be modified so that it takes a class label in addition to the random latent variable so that it generates a new element similar to the training examples of that class. The training is augmented with an additional loss term that models the class of the training examples.

There are many more fascinating examples.   We will describe them in more detail in a later post.

# Building a “ChatBot” for Scientific Research

We can use Alexa, Cortana and Siri to check the weather, stream our favorite music and lookup facts like “who directed Casablanca?” But if I want to find all the research on quantum entanglement in the design of topological quantum computers, these services will fall short.   If, in addition, I want these articles cataloged in my personal archive and the citations automatically pulled I need a much more robust digital assistant.  Raj Reddy talks about a broader and more pervasive future for digital assistant he calls Cognition Amplifiers and Guardian Angels.  In this post we look at chatbots and their limitations and show how to build a simple, voice-driven scientific metasearch tool we call the research assistant.  Finally, we discuss the next phase of research assistant.

# Smart Speakers and Chatbots.

The revolution in “smart speaker” digital assistants like Siri, Amazon Echo, Google Home is showing us the power of voice to provide input to smart cloud services.   These assistants can take notes, tell us the weather conditions, place on-line orders for us and much more.   I even allow Microsoft’s Cortana to read my email.  If I send the message “I’ll get back to you tomorrow” to a friend, Cortana will remind me the next day that a response is needed.  Amazon allows people to add “skill” (additional capabilities) to there Alexa system.  These smart speakers are designed around open-ended question-response scenario.  These assistants leverage very powerful speech-to-text technology and semantic analysis systems to extract a query or command.  The query answers are derived from web-data analysis and the commands are to a limited number of native services (or external skills).

A chatbot is a system that engages the user in a dialog.   They go beyond the question answering smart speakers and are usually designed to help people interact with the services of a specific company or agency.   The Google, IBM, Amazon and Microsoft have all introduced cloud services to help anybody build a chatbot.   These services guide you through the process of building and training a chatbot.   A good example is Google’s Dialogflow.   Using this tool to create a bot, you specify three things:

• Intents – which are mapping between what the user says and how you wish the system to respond.
• Entities – that are the categories of subjects that your bot understands. For example, if you bot is a front end to your clothing store, one category of entity my be product type: shoes, dress, pants, hats,  and another entity might be size: large, xl, small, medium,  and another might be color.
• Context – This is knowledge obtained by the bot during the course of the conversation. For example, the name of the user, or the users favorite shoe color.

The goal of the bot is to extract enough information from the user to take a meaning action such as fulfilling an order.   The hard part is designing the intents so that a dialog can lead to the desired conclusion.  The resulting sysgtem is called an agent and the flow of action is illustrated in Figure 1.  It might start with ‘Good morning, how can I help you?’   and end with a summary of the discussion.  As the programmer you need to supply as many possible variations on possible user questions and responses as possible.  And you must annotate these with makers where your entities can be extracted.    Your examples are used in a training phase that maps your intents and entities together and builds a model that will learn variations on the input and not repeat the same sequence of responses each time, so it seems less robotic to the user.

Amazon’s Lex provides a similar service that also integrates with their lambda services.  Microsoft has the Azure Bot Service and IBM has Watson assistant chatbot tools.

Figure 1. Google Dialogflow Agent architecture.

These tools are all designed to help you build a very focused conversation with the user about a very narrow universe such as the on-line part of a business.   But this raises the question, can one build a chat bot that can carry out an open-ended conversation.  Perhaps one that could pass the Turing test?   The research literature on the subject is growing and deep learning tools like recurrent and convolutional neural networks have been applied to the problem (see https://arxiv.org/pdf/1606.07056.pdf , https://arxiv.org/pdf/1605.05110.pdf and more).   Unfortunately chatbots designed to engage in open-ended conversation have only had a limited success.   Xiaoice is one that interacts with users on the Chinese micro blogging service Weibo.  The problem is that while it sounds like conversation, it is mindless.  Microsoft’s Tay was an English language version that operated on Twitter until it was taken down after only 16 hours because of the unfortunate language it had learned.  A successor Zo seems to be working better, but it does not produce conversations with intellectual content.

There is an excellent pair of articles by Denny Britz about the role of deep learning for conversational assistants.   He make the point that for open-ended conversations (he calls them open-domain) the challenges are large compared to the fixed domain chatbots because so much more world knowledge is required.

# Cognition Amplifiers and the Research Assistant.

In the spring of 2018 Raj Reddy gave the keynote presentation at the IEEE services congress.  His topic was one he has addressed before and it clearly resonated with the audience.   He described Cognition Amplifiers and Guardian Angels.  He defined a Cognition Amplifier (COG) as a Personal Enduring Autonomic Intelligent Agent that anticipates what you want to do and help you to do it with less effort. A Guardian Angel (GAT) is a Personal Enduring Autonomic Intelligent Agent that discovers and warns you about unanticipated events that could impact your safety, security and happiness.

Consider now the application of the Cognition Amplifier to scientific research.   If you are working on writing a research paper, you may wish your autonomic research assistant to provide a fast and accurate search of the scientific literature for a specific list of scientific concepts. In fact, as you write the paper, the assistant should be able to pick up the key concepts or issues and provide a real-time bibliography of related papers and these should be stored  and indexed in a private space on the cloud.  Extracting key phrases from technical documents is already a heavily research field so applying this technology to this problem is not a great leap.   However, key phrase extraction is not the whole challenge.   Take sentence “It seems that these days investors put a higher value on growth than they do on profitability”.  The categorical topic is value of growth vs profitability – investor opinions which is not simply a key phrase, but a concept and we need the research assistant to look for concepts.    Your research assistant should always understand and track the context of your projects.

Finally, a good research assistant for science should be able to help with the analytical part of science.  For example, it should help locate public data in the cloud related to experiments involving your topics of interest.   The assistant should be able to help formulate, launch and monitor data analysis workflows. Then coordinate and catalog the results.

And, of course, if your autonomous research assistant is also a Guardian Angel, it will also keep you informed of grant reporting deadlines and perhaps pull together a draft quarterly report for your funding agency.

I fully expect that it is possible to build such an agent in the years ahead.   However, the remainder of this article is a simple demo that is a far cry from the full research assistant agent described above.

# The Research Assistant Metasearch Tool.

In the following paragraphs we describe a very simple voice-driven agent that can be used to look for research articles about scientific topics.  We also show how such a system can be assembled from various simple devices and cloud services.   The system we describe is not very sophisticated.  In fact it is not much better than Cortana at finding things given English input sentence.  However we feel it does illustrate the architecture of a voice-driven agent that can be built by gluing together easy to use cloud services.

Our scenario is a researcher sitting near the device and asking about very specific research topics such as “physical models of immune response” or “programming a topological quantum computer”.    We assume the user wants a spoken response if that response is simple, but we also realize that this is impractical if the system is creating a list research journal papers.  To address this issue, the system also has a display in a web browser.  (We note that the Cortana and Google assistant do the same if the response is a list.)

Figure 2 illustrates the basic architecture of the system.

Figure 2.  The architecture of the research assistant.

The components of the system are:

1. The voice-to-text translator. Here we use a simple voice kit from google.   This consists of some special hardware and a raspberry pi 2 computer all packaged in an elegant cardboard box.  You wake the system up by pressing the botton on top and speak.   The audio is captured and sent to the google voice service for transcription and it is returned as a text string.
2. The next step is to parse the text string into the components that will allow us to extract the topics of the query.  This is another cloud service call.  This time it is to Algorithmia.com and the service is called key phrases.  (we wrote about this in a previous article.)  The service takes English sentences and invoked Googles ParsyMcParseface (another Algorithmia.com AI service) and returns a list composed of three types of phrases: subject (s), actions (a) and objects (o).  It also flags prepositional phrases with a “/” character.   So for example, “I am interested in physical models of immune response” returns

[‘s: I ‘, ‘a: am ‘,  ‘o: interested /in physical models /of immune response.’]

1. The analysis engine is a 500-line Python Dash-based web server that extracts the topics and a few other items of interest and decides how to search and display the results on the browser. There are three web services used for the search: Wikipedia,  Bing and Cornell’s ArXiv service[1].  To see how this works, consider the example the sentence “research papers by Michael Eichmair about the gannon-lee singularity are of interest“. The analysis engine detects the topic as the gannon-lee singularity and Michael Eichmar as the author.  The fact that research papers are of interest indicates that the we should look in the Cornell ArXiv repository of papers.   (The results for this query are at the end of this section).    (Truth in advertising: our parser and analysis are far from perfect.   For example, “tell me about Deep Learning”   vs “tell me about deep learning” yield two different parses.  The first yields

[‘a: tell ‘, ‘o: me /about Deep Learning ‘]

which is fine. But the second gives us

[‘a: tell ‘, ‘o: me about deep ‘, ‘a: learning ‘]

which causes the analysis to fail. )

1. Finally, we use the Amazon Lex services to generate the audio reading of the Wikipedia results. If you have an aws account, the Python API is easy to use.

## Examples

Figure 3 illustrates the interface.  We have started with the statement “I am interested in physical models of immune response.”

Figure 3.   The interface provides a switch to allow the Wikipedia response to read aloud.   In this case we have typed the English statement of intent into the query box and hit the “Translate and Search” button.

We respond with the phrase “look for it in Wikipedia” and get the result in figure 4.    Following that response, we say “how about research papers” and we get the response in figure 5.

Figure 4.   The response to “look for it in wikipedia”.  A short summary from Wikipedia is extracted along with related images found on the subject.  The spoken response is controlled by the audio control  at the top of the response.

Figure5.   The mention of research papers suggest that we should consult the Cornell library arXiv.    Shown above is only the first result of 10 listed on the page.

Returning to the example mentioned above “research papers by Michael Eichmair about the gannon-lee singularity are of interest” we get the following results.   You will notice that the Wikipedia result is a default hit for “big bang singularity” and not directly related to the precise query.  The Bing results and the ArXiv hits are accurate.

Figure 6.  Results for the query “research papers by Michael Eichmair about the gannon-lee singularity are of interest”.  (This page was slightly edited to shorten the list of Bing results.)

The system has a limited capability to pull together concepts that are distributed over multiple sentences.  For example the input string “what are anyons?   How do they relate to topological quantum computation?” will build the topic “anyons topological quantum computation”.

If you are interested in trying to use the system point your browser here.  I will try keep it up and running for a few months. There is no voice input because that requires a dedicated Google voice kit on your desk.   You need to decide if you want to have a playback of the audio for Wikipedia summaries.   If you don’t want it, simply press the “Read Aloud” button.  Then enter a query and press the “Translate and Search” button.   Here are some samples to try:

1. what did they say about laughter in the 19th century?
2. are there research papers about laughter by Sucheta Ghosh?
3. what can you tell me about Quantum Entanglement research at Stanford? (this one fails!)
4. what can you tell me about research on Quantum Entanglement at Stanford?
5. what are anyons? How do they relate to topological quantum computation?
6. Who was Winslow Homer? (this one give lots of images)
7. I am interested in gravitational collapse. (respond with web, Wikipedia or arxiv)

As you experiment, you will find MANY errors.  This toy is easily confused.   Please email me examples that break it.  Of course, feedback and suggestions are always welcome.  I can make some of the source code available if there is interest. However, this is still too unreliable for public github.

[1] There are other arguably superior sources we would like to have used.  For example, Google Scholar would be perfect, but they have legal restrictions on invoking that service from an application like ours.  Dblp is also of interest but it is restricted to computer science.

# Introduction

This post is based on a talk I prepared for the Scientific Cloud Computing Workshop at HPDC 2018.

Two years ago, Ian Foster and I started writing  “Cloud Computing for Science and Engineering”.   That book covers fundamental cloud tools and computational models, but there are some topics we alluded to but did not explore fully because they were still on the horizon.  In other cases, we were simply clueless about changes that were coming. Data center design and cloud services have evolved in some amazing ways in just two years and many of these changes represent opportunities for using cloud technology in scientific research.

## Whose cloud?

Any  discussion of cloud computing in science leads to the question of definition.  What defines a cloud for science?   For example, the European Open Science Cloud (EOSC) is a European-wide virtual environment for data sharing and collaboration.  That project will involve multiple data archives, research labs and HPC centers, commercial service providers and EU agencies and funding.  It is truly an important project.  However, my focus here is on the fundamental technologies that are driving hardware and software innovation, and these tend to come from a combination of academic, open source and commercial providers.   The most ubiquitous commercial clouds are:

• Amazon Web Services (AWS) – 40% of all commercial cloud resources on the planet,
• Microsoft Azure – about 50% of AWS but growing,
• Google Cloud – a solid and growing third place,
• IBM Bluemix – growing very fast and in some measures bigger now that Google.

There are many more, smaller or more specialized providers: Salesforce, DigitalOcean, Rackspace, 1&1, UpCloud, CityCloud, CloudSigma, CloudWatt, Aruba, CloudFerro, Orange, OVH, T-Systems.

There are a number of smaller cloud systems that have been deployed for scientific research.  They  include Aristotle, Bionimbus, Chameleon, RedCloud, indigo-datacloud, EU-Brazil Cloud,  and the NSF JetStream.  The advantage of these research clouds is that they can be optimized for use by a specific user community in ways not possible in a commercial cloud.  For example, Chameleon is funded by the US NSF to support basic computer systems research at the foundational level which is not possible when the foundation is proprietary.

## Are these clouds of any value to Science?

When the first commercial clouds were introduced in 2008 the scientific community took interest and asked if there was value there.  In 2011 the official answer to this question seemed to be  “no”.  Two papers (see end node 1) described research experiments designed to address this question.   The conclusion of both papers was that these systems were no match for traditional supercomputers for running MPI-based simulation and modeling.   And, in 2010, they were correct.   Early cloud data centers were racks of off-the-shelf PCs and the networks had terrible bisection bandwidth and long latencies.   They were no match for a proper HPC cluster or supercomputer.

Over the last few years, others have recognized a different set of roles for the cloud in science that go beyond traditional supercomputer simulation.   The biology community was quick to adopt cloud computing especially when it is necessary to do large scale analysis on thousands of independent data samples.  These applications ranged from metagenomics to protein folding.   These computations could each fit on a single server, so network bandwidth is not an issue and, using the scale of the cloud, it is easy to launch thousands of these simultaneously.   Hosting and sharing large public scientific data collections is another important application.   Google, AWS, Microsoft and other have large collections and they also are also providing new ways to host services to explore this data.

However, there are at least three additional areas where the cloud is a great platform for science.

### Analysis of streaming data

Microsoft’s AI for earth project (Figure 1) looks at the application of streaming data from sources such as satellites to do land cover analysis,  sensors on and above farm land to improve agriculture outcomes and crowd sourced data to help understand biodiversity.

Figure 1.  Applications of streaming include land cover analysis, using sensor data for improving agriculture and biodiversity.   From  https://www.microsoft.com/en-us/aiforearth

The internet of things is exploding with sensor data as are on-line experiments of all types.  This data will be aggregated in edge computing networks that do initial analysis with results fed  to the cloud systems for further analysis.   Urban Informatics is a topic that has emerged as critical to the survival of our cities and the people who live in them.  Sensors of all types are being deployed in cities to help understand traffic flow, microclimates, local pollution and energy waste.  Taken together this sensor data can paint a portrait of the city that planners can use to guide its future.  Streaming data and edge computing is a topic that will involve the growing capabilities and architecture of the cloud.  We will return to this later in this document.

### Interactive big data exploration

Being able explore and interact with data is a critical component of science.   If it fits on our laptop we can use tools like Matlab, Excel or Mathematica to conduct small computational experiments and visualize the results.  If the data is too big it must be stored on something bigger than that laptop.   Traditional supercomputers are up to the task of doing the analysis, but because they are designed around batch computing there are not well suited to interactive use.   The cloud exists to host services that can be used by thousands of simultaneous users.   In addition, there is a new generation of interactive data analysis tools that are cloud-ready for use on very large data collections.  This collection of tools includes Spark and Python Dask.   In both cases these tools can be driven by the open-source Jupyter studio which provides a powerful, interactive compute and visualization tool.  The commercial providers have adapted Jupyter and its interactive computational model into their offerings.   Google has Cloud Datalab (Figure 2), Amazon uses Jupyter with its SageMaker Machine Learning platform and Microsoft provide a special data science virtual machine that runs Jupyter Hub so that teams of users can collaborate.

Figure 2.  Google’s Cloud Data lab integrates SQL-like queries to be combined with Python code and visualization to a Jupyter based web interface. https://cloud.google.com/datalab/

Being able to interact with data at scale is part of the power of the cloud.   As this capability is combined with advanced cloud hosted machine learning tools and other services, some very promising possibilities arise.

### The quest for AI and an intelligent assistant for research

The commercial clouds were originally built to  host web search engines.   Improving those search engines led to a greater engagement of the tech companies with machine learning.   That work led to deep learning which enabled machine language translation,  remarkably strong spoken language recognition and generation and image analysis with object recognition.  Many of these  capabilities rival humans in accuracy and speed.  AI is now the holy grail for the tech industry.

One outcome of this has been the proliferation of voice-driven digital assistants such as Amazon’s Echo, Microsoft’s Cortana, Apple’s Siri and Google Assistant.   When first introduce these were novelties, but as they have improved their ability to give us local information, do web searching, keep our calendars has improved considerably.   I believe there is an opportunity for science here.

Ask the question “what would it take to make Alexa or Cortana help with my research?”   The following use cases come to mind.

1. Provide a fast and accurate search of the scientific literature for a given specific scientific concept or topic and not just a keyword or specific author. Then ask who is working on or has worked on this topic?  Is there public data in the cloud related to experiments involving this topic?  Translate and transcribe related audio and video.
2. Understand and track the context of your projects.
3. Help formulate, launch and monitor data analysis workflows. Then coordinate and catalog results.  If state-space search is involved, automatically narrow the search based on promising findings.
4. Coordinate meetings and results from collaborators.

If I ask Alexa to do any of this now, she would politely say “Sorry.  I can’t help you with that.”  But with the current rate of change in cloud AI tools, ten years seems like a reasonable timeframe.

Figure 3.  Siri’s science geek grandson.

# Technical Revolutions in the Cloud

Two of the three scenarios above are here now or very close.   The third is some ways off.   There have been three major changes in cloud technology in the past five years and some aspects of these changes are true game-changers for the industry.    The first, and most obvious is the change in scale of cloud deployments. The two leaders, AWS and Azure are planetary in scale.  This is illustrated in Figure 4 below.

Figure 4.  A 2016 map of cloud coverage from Atomia.com.  https://www.atomia.com/2016/11/24/ comparing-the-geographical-coverage-of-aws-azure-and-google-cloud/ There is some inaccuracy here because AWS and Azure define regional data centers differently, so counting the dots is not a good comparison.   In addition, data centers are now under construction in South Africa and the Middle East.

This map does not include all the data centers run by the smaller cloud providers.

# Cloud Services

A major departure from the early days of the cloud, where scientists focused on storage and servers, has been an explosion in pay-by-the-hour cloud hosted services.  In addition to basic IaaS the types of services available now are:

• App services: Basic web hosting, mobile app backend
• Streaming data: IoT data streams, web log streams, instruments
• Security services: user authentication, delegation of authorization, privacy, etc.
• Analytics: database, BI, app optimization, stream analytics
• Integrative: networking, management services, automation

In addition, the hottest new services are AI machine learning services for mapping, image classification, voice-to-text and text-to-voice services and text semantic analysis.   Tools to build and train voice activated bots are also now widely available.   We will take a look at two examples.

### A planet scale database

The Azure Cosmos DB is a database platform that is globally distributed.   Of course, distributing data across international boundaries is a sensitive topic, so the Cosmos platform allows the database creator to pick the exact locations you want copies to reside.   When you create an instance of the database you use a map of Azure data centers and select the locations as shown in Figure 5.

Figure 5.  Cosmos DB.  A database created in central US and replicated in Europe, South India and Brazil.

The database can support 4 modes: Documents, key-value Graph and NoSQL.  In addition, there are five different consistency models the user can select: eventual, consistent prefix, session, bounded stateless and strong consistency all with 99.9% guarantee of less than 15ms latency.  My own experiments validated many these claims.

### Cloud AI Services

The commercial clouds are in a race to see who can provide the most interesting and useful AI services on their cloud platform.   This work began in the research laboratories in universities and companies over the past 25 years, but the big breakthroughs came when deep learning models trained on massive data collections began to reach levels of human accuracy.  For some time now, the public cloud companies have provided custom virtual machines that make it easy for technically sophisticated customers to use state of the art ML and neural network tools like TensorFlow, CNTK and others.  But the real competition is now to provide services for building smart applications that can be used by developers lacking advanced training in machine learning and AI. We now have speech recognition, language translation, image recognition capabilities that can be easily integrated into web and mobile applications.

We gave this a try with services that use a technique called Transfer Learning to make it possible to re-train a deep neural network to recognize objects from a narrow category using a very small training set.   We chose images of galaxies and used the services of IBM Watson, Azure and Amazon.   Figure 6 illustrates the results from IBM’s tool.  The results were surprisingly good.

Figure 6.  IBM’s Watson recognizing previously unseen images of three different galaxies.  The details of this study are here: https://esciencegroup.com/2018/02/16/cloud-services-for-transfer-learning-on-deep-neural-networks/

# The Revolution in Cloud Service Design

Making all of these services work, perform reliably and scale to thousands of concurrent users forced a revolution in cloud software design.    In order to support these applications, the tech companies needed a way to design them so that they could be scaled rapidly and updated easily.   They settled on a design pattern that based on the idea of breaking the applications into small stateless components with well defined interfaces.   Statelessness meant that a component could be replaced easily if it crashed or needed to be upgraded.   Of course, not everything can be stateless, so state is saved in cloud hosted databases.   Each component was a “microservice” and it could be built from containers or functions.  This design pattern is now referred to as “cloud native” design.   Applications built and managed this way include Netflix, Amazon, Facebook, Twitter, Google Docs, Azure CosmosDB, Azure Event hub, Cortana, Uber.

Figure 7.  Conceptual view of microservices as stateless services communicating with each other and saving needed state in distribute databases or tables.

To manage applications that required dozens to hundreds of concurrently running microservice you need a software foundation or container orchestration system to monitor the services and schedule them on available resources.  Several candidates emerged and are used. Siri, for example, is composed of thousands of microservices running on the Apache Mesos system.   Recently cloud providers have settled on a de-facto standard container orchestrator built by Google and released as open source called Kubernetes.   It is now extremely easy for any customer to use Kubernetes on many cloud deployments to launch and manage cloud native applications.

## Serverless Functions

The next step in the cloud software evolution was the introduction of “serverless functions”.   The original idea of cloud computing involved launching and managing a virtual machine.  However, suppose you want to have a cloud-based application whose sole job is to wait for an event to trigger some action.  For example, monitor a file directory and wait for a change such as the addition of a new file.  When that happens, you want to send email to a set of users alerting them of the change.   If this is a rare event, you don’t want to have to pay for an idle virtual machine that is polling some service looking for a change.  Amazon was the first to introduce the concept of a function-as-a-service.   With AWS Lambda, you only need to describe the function in terms of the trigger event and the output actions it takes when the even appears.  As illustrated in Figure 8, there are many possible triggers and corresponding output channels.

Figure 8.   From Amazon AWS.   The triggers and outputs of a lambda function.

In addition to AWS Lambda, Azure Functions, Google Functions, IBM OpenWhisk are similar systems.  OpenWhisk is now open source.  Another open source solution is Kubeless that allow you to deploy a lambda-like system on top of your Kubernetes cluster.   These serverless systems let you scale up and down extremely rapidly and automatically.   You can have hundreds of instances responding to events at once.  And the cost is based on charge-by-use models.  AWS has  introduced AWS Fargate which allows any containerized application to run in serverless mode.

## The Edge and the Fog

The frontier of cloud computing is now at the edge of the network.  This has long been the home of content distribution systems where content can be cached and access quickly, but now that the Internet-of-Things (IoT) is upon us, it is increasingly important to do some computing at the edge.  For example, if you have a thousand tiny sensors in a sensitive environment or farm and you need to control water from sprinklers or detect severe weather conditions, it is necessary to gather the data, do some analysis and signal an action.   If the sensors are all sending WIFI messages they may be routable to the cloud, but a more common solution is to provide local computing that can do some event preprocessing and response while forwarding summary data to the cloud.  That local computing is called the Edge, or if a distributed systems of edge servers, the Fog.

If serverless functions are designed to respond to signals, that suggests that it should be possible to extend them to run in the edge servers rather than the cloud.  AWS was the first to do this with a tool called GreenGrass that provides a special runtime system that allows us to push/migrate lambda functions or microservices from the data center to the edge.   More recently Microsoft has introduced Azure IoT Edge which is built on open container technologies.  Using an instance open source Virtual Kubelet deployed on the edge devices we can run our Kubernetes containers to run on the edge.  You can think of a Kubelet as the part of Kubernetes that runs on a single node. This enables Kubernetes clusters to span across the cloud and edge as illustrated in Figure 9.

Figure 9.  shows a sketch of migrating containerized functions to edge function.   That way our IOT devices can communicate with locally deployed microservices.  These microservices can communicate with cloud-based services.  The edge function containers can also be updated and replaced remotely like any other microservice.

# The Evolution of the Data Center

As mentioned at the beginning of this paper, the early days (2005) of cloud data center design systems were based on very simple server and networks were designed for outgoing internet traffic and not bisectional bandwidth for parallel computing.   However, by 2008 interest in performance began to grow.  Special InfiniBand sub-networks were being installed at some  data centers.  The conventional dual-core servers were being replaced by systems with up to 48 cores and multiple GPU accelerators.  By 2011 most of the commercial clouds and a few research clouds had replaced traditional network with software defined networking.  To  address the demand of some of its customers, in 2017 Microsoft added  Cray® XC™  and Cray CS™ supercomputers to a few data centers and then acquired the company cycle computing.

From 2016 we have seen progress focused on performance and parallelism.   The driver of this activity has been AI and, more specifically, the deep neural networks (DNNs) driving all the new services.  There are many types of DNNs but two of the most common are convolutional, which look like a linear sequence of special filters, and recurrent networks which, as the name implies, are networks with a feedback component.   And there are two phases to neural network design and use.  The first is the training phase which requires often massive parallelism and time.  But it is usually an off-line activity.  The second phase is called inference and it refers to the activity of evaluating the trained network on classification candidates.  In both the convolutional and recurrent network inference boils down to doing a very large number of matrix-vector and matrix-matrix multiplies where the coefficients of the matrix are the trained model and the vector represent the classification candidates.      To deliver the performance at scale that was needed by the AI services it was essential to do these operations fast.  While GPUs were good, more speed was needed.

In 2015 Google introduced the Tensor Processing Unit (TPU) and the TPU2 in 2017.

Figure 10.  Above, Google data center.  https://www.nextplatform.com/2017/05/22/hood-googles-tpu2-machine-learning-clusters/  Below architecture of Google Tensor Processing Unit TPU. From “In-Datacenter Performance Analysis of a Tensor Processing Unit”, Norman P. Jouppi et al.​  ISCA 2017  https://ai.google/research/pubs/pub46078

Figure 10 illustrates several racks of TPU equipped servers and the functional diagram of the TPU.  One of the key components is the 8-bit matrix multiply capable of delivering 92 TeraOps/second (TOPS).   (It should be noted that DNNs can be well trained on less than IEEE standard floating-point standards and floating point systems with small mantissa are popular.)  The multiply unit uses a systolic algorithm like those proposed for VLSI chips in the 1980s.

## Microsoft’s Brainwave

In 2011 a small team in Microsoft Research led by Doug Burger began looking at the use of FPGAs to accelerate the search result ranking produced by Bing.   Over several iterations they arrived at a remarkable design that allowed them to put the FPGA between the network and the NIC so that the FPGA could be configured into separate plane of computation that can be managed and used independently from the CPU (see Figure 11).   Used in this way groups of FPGAs could be configured into a subnetwork to handle tasks such as, database queries and  inference stage of deep learning in addition to Bing query  optimization.

Figure 11.  The Brainwave architecture.

Figure 12. The brainwave software stack for mapping a DNN to one or more FPGAs. From BrainWave_HOTCHIPS2017.pptx, Eric Chung, et. al., https://vneetop.wordpress.com/ 2017/10/28/accelerating-persistent-neural-networks-at-datacenter-scale/

The team also built a software stack that could really make this possibility a reality.   What makes much of this possible is that the models for DNNs are all based on flow graphs which describe sequences of tensor operations.  As shown in Figure 12 above, the flow graphs can be compiled to a graph internal representation that can be split and partitioned across one or more FPGAs.  They refer to the result as a hardware microservice.   Just recently Mary Wall [see endnote 2] wrote a nice blog about the teams work on using this idea to do deep learning inference on land cover maps.  Each compiled inference hardware microservice is mapped to a single FPGA, but they used 800 of the inference instances in parallel with 80 VMs to process 20 terabytes of aerial imagery into land cover data for the entire United States.   It took only about 10 minutes for a total cost of $42. [see endnote3] Mary Wall’s code is in the blog and available in Github. # Conclusion Cloud data centers are getting high performance networks (with latencies of only a few microseconds in the case of Azure Brainwave) and immense computing capacity such as the tensor processing capability of Google’s TPU. At the same time designers of supercomputers are having to deal with more failure resilience and complexity in the design of the first exascale supercomputers. For the next generation exascale systems the nodes will be variations on a theme of multicore and GPU-style accelerators. Observed from a distance, one might conclude the architectures of cloud data centers and the next generation of supercomputers are converging. However, it is important to keep in mind that the two are designed for different purposes. The cloud is optimized for fast response for services supporting many concurrent globally distributed clients. Supers are optimized for exceptionally fast execution of programs on behalf of a small number of concurrent users. However, it may be the case that an exascale system may be so large that parts of it can run many smaller parallel jobs at once. Projects like Singularity provide a solution for running containerized application on supercomputers in a manner similar to the way microservices are run on a cloud. ## Possible Futures ### The continuum: edge+cloud+supercomputer There are interesting studies showing how supercomputers are very good at training very large, deep neural networks. Specifically, NERSC scientists have show the importance of this capability in many science applications[4]. However, if you need to perform inference on models that are streamed from the edge you need the type of edge+cloud strategy described here. It not hard to imagine scenarios where vast numbers of instrument streams are handled by the edge and fed to inference models on the cloud and those models are being continuously improved on a back-end supercomputer. ### A data garden In the near future, the most important contribution clouds can make to science is to provide access to important public data collections. There is already reasonable start. AWS has an opendata registry that has 57 data sets covering topics ranging from astronomy to genomics. Microsoft Research has a Data Science for Research portal with a curated collection of datasets relating to human computer interaction, data mining, geospatial, natural language processing and more. Google cloud has a large collection of public genomics datasets. The US NIH has launch three new cloud data and analytics projects. They include the Cancer Genomics Cloud led by the Institute for Systems Biology with Google’s cloud, FireCloud from the Broad Institute also using Google’s cloud and Cancer Genomics Cloud (CGC), powered by Seven Bridges. These NIH facilities also provide analytics frameworks designed to help research access and effective use the resources. I am often asked about research challenges in cloud computing that student may wish to undertake. There are many. The fact that the IEEE cloud computing conference being held in San Francisco in July received nearly 300 submissions shows that the field is extremely active. I find the following topics very interesting. 1. Find new ways to extract knowledge from the growing cloud data garden. This is a big challenge because the data is so heterogeneous and discovery of the right tool to use to explore it requires expert knowledge. Can we capture that community knowledge so that non-experts can find their way? What are the right tools to facility collaborative data exploration? 2. There are enormous opportunities for systems research in the edge-to-cloud-to-supercomputer path. How does one create a system to manage and optimize workflows of activities that span this continuum? Is there a good programming model for describing computations involving the edge and the cloud? Can a program be automatically decomposed into the parts that are best run on the edge and the parts on cloud? Can such a decomposition be dynamically adjusted to account for load, bandwidth constraints, etc.? 3. Concerning the intelligent assistant for research, there are a number of reasonable projects short of build the entire thing. Some may be low hanging fruit, and some may be very hard. For example, ArXiv, Wikipedia and Google search and Bing are great for discovery but in different ways. Handling complex queries like “what is the role of quantum entanglement in the design of a quantum computer?” should lead to a summary of the answer with links. There is a lot of research on summarization and there are a lot of sources of data. Another type of query is “How can I access data on genetic indicators related to ALS?” Google will go in the right direction, but it takes more digging to find data. These are rather broad topics, but progress on even the smallest part may be fun. [1] L. Ramakrishnan, P. T. Zbiegel, S. Campbell, R. Bradshaw, R. S. Canon, S. Coghlan, I. Sakrejda, N. Desai, T. Declerck, and A. Liu. Magellan: Experiences from a science cloud. In 2nd International Workshop on Scientiﬁc Cloud Computing, pages49–58., ACM, 2011. P. Mehrotra, J. Djomehri, S. Heistand, R. Hood, H. Jin, A. Lazanoﬀ, S. Saini, and R. Biswas. Performance evaluation of Amazon EC2 for NASA HPC applications. In 3rd Workshop on Scientiﬁc Cloud Computing, pages 41–50. ACM, 2012 [3] https://blogs.microsoft.com/green/2018/05/23/achievement-unlocked-nearly-200-million-images-into-a-national-land-cover-map-in-about-10-minutes/ from Lucas Joppa – Chief Environmental Scientist, Microsoft # Algorithmia™: A Cloud Marketplace for Algorithms and Deep Learning One area of great frustration encountered by application developers involves the challenge of integrating new algorithms into a code base. There are many reasons for this. For example, the algorithm may be described in a journal article where many details of the implementation are omitted or it is available only in a programming language different from the one being used. The code may have software dependencies that are hard to resolve. The new algorithm may also have hardware dependencies, such as reliance on a GPU to get performance and you may not have access to this hardware. On the other hand, if you are the author of a great new algorithm you may be disappointed that your new invention is not being used for these very same reasons. About 18 months ago a company called Algorithmia™ was founded in Seattle that provides an elegant solution to these problems. They provide a very simple multi-language API that can be used to invoke any of their catalog of 3,500 different cloud-based algorithms. While we may be getting tired of reading about X-as-a-Service for different versions of X, there is one binding for X that has been around for a while in various forms and, as much as it pains me to do so, it begs to be called Algorithms as a Service. And AaaS is just one of the things Algorithmia provides. AaaS is indeed not a new idea. Jack Dongarra and his ICL team at the University of Tennessee created NetSolve/GridSove in 2003 to provide scientists and engineers with access to state-of-the-art numerical algorithms running on a distributed network of high performance computers. As cool as NetSolve is, Algorithmia goes several steps beyond this concept. One of Algorithmia’s cofounders and CEO, Diego Oppenheimer has a deep background in building business intelligence tools. While working on that he developed an appreciation of the power of being able to call out to powerful algorithms from inside a user facing application. This capability allows the application to have access to deeper knowledge and more powerful computational resources than available on the user’s device. A key insight from this experience is that algorithms must be discoverable an invokable from any user application runtime. These ideas are all central to Algorithmia. In the following paragraphs we will look at Algoritmia’s marketplace, explore building a new algorithm and discuss a bit of the system microservice architecture. # Algorithmia is a marketplace. There are over 50,000 developers that use Algorithmia services and the platform encourages these developers to contribute new algorithms to the collection. Invoking an algorithm is dead simple and it can be done from any programming language that can formulate a JSON doc and send a REST message. We will provide some detailed illustrations at the end of this document. To use it, you need to set up an account. Doing so will get you a starter award of 5000 or so “credits”. When you invoke an algorithm, credits are deducted from your account. Typically, there is a “royalty” cost of about 10 credits and then the cost is usually around one credit per second of execution. A fun example from their library of deep learning collection is an image colorizer. Input is a PNG file of a black and white image and the returned value is a link to the output colorized image. We took a color image from a visit to Red Square a few years ago. We converted it to a grayscale image and gave that to the colorizer. The result is shown illustrated below. The original is on the left, grayscale in the middle and the colorized image on the right. While it is not as good as the best hand-colored photos, it is not too bad. It lost the amazing color of St. Bazil’s Cathedral which is not too surprising, but it was great with sky and skin tones of those people in foreground. (It seemed to think the bricks of the square would look better with some grass color.) The Python code to upload the grayscale image and invoke the service was incredibly simple. import Algorithmia client = Algorithmia.client(‘youruserkeyfromaccountrecation’) input = bytearray(open("path_to_grayscale.png", "rb").read()) result = client.algo("deeplearning/ColorfulImageColorization/1.1.6") .pipe(input).result path_to_local_copy_of_result_image= client.file(result[‘output’]).getFile()  The cost in credits was 154. The exchange rate for credits is 1$ = 10,000 credits (approximately) so this invocation would have cost about 1.5 cents.

This algorithm is from their extensive machine learning and AI collection.  A related algorithm is one that computes the salience of objects in an image.  Salience is the degree to which an object in the image attracts the attention of the viewer’s eye.   The algorithm is called SalNet and it is based on ideas from the paper, Shallow and Deep Convolutional Networks for Saliency Prediction by Pan et. al.  (see arXiv:1603.00845v1).

As with the colorizer, salnet it is easy to invoke.

input = { "image": "data://.algo/deeplearning/SalNet/perm/an-uploaded-image.png" }
result2 = client.algo("deeplearning/SalNet/0.2.0").pipe(input).result


Note that in this case we have loaded the image from one that we uploaded to Algorithmia’s data cloud.  In fact, it is the same grayscale image of red square.  As you can see below, the algorithm picks out the woman in the foreground and also notices the church in the background.

Salience computation can be very helpful in identifying and labeling objects in an image.   Image tagging is also something that Algorithmia supports.   Running the same image through their tagger returned the observations that the image was “safe” and that there were multiple boys and multiple girls and sky and clouds and it seem to be near a palace.

There are many other AI related image algorithms such as nudity detection, character recognition, face detection and a very impressive car make and model recognition algorithm.   A quick look at https://algorithmia.com/use-cases will show many other fascinating use cases.

Another very cool capability of Algorithmia is its ability to host your trained machine learning model.  Suppose you have a model you have built with MsXNet, TensorFlow, Scikit-Learns, CNTK or any of the other popular ML frameworks, you can upload your model to Algorithmia so that it can be available as a service.   This is explained in here. We will explore this capability in a later post.

While the main emphasis and attraction of the Algorithmia collection is machine learning and AI, there are many more algorithm categories represented there.  For example, there is an excellent collection of utilities for managing data and making certain programming tasks extremely easy: such as extracting text from web pages, Wikipedia search tools, computing the timezone and elevation from lat, lon coordinates.

There is also a large collection of time series analysis algorithms.   These include forecasting, outlier detection, Fourier filters, auto-correlation computation and many more.

# Algorithmia is cloud of microservices

In an excellent talk at the 2017 Geekwire cloud summit, Oppenheimer described some key elements of Algorithmia’s architecture.  In this talk he makes the critically important observation that two phases of machine learning,  training and prediction, if used in production require very different execution environments.   Training is often done on a dedicated system consuming many hours of compute and as much memory as is available.   The result of training is a model codified as data.   Prediction (also called Inference) uses the model to make predictions or inferences about a sample case.   Prediction can be done on the same hardware platform that was used for the training, but if the model is to be used to make predictions concerning thousands of cases for thousands of concurrent users,  one need a completely different design.

Their approach to the scale problem for predictions (and for any high demand algorithm in their collection) is based on serverless microservices.    They use a Kubernetes microservice foundation with algorithms deployed in Docker containers.  Requests from remote client applications are load balanced across API servers who dispatch requests to container instances for the requested function.  The challenge is making the latency from request to reply very low.  If a container for an algorithm is already in system memory, it requires very little time to spawn a new instance on Kubernetes.  Another technique they use it to dynamically load algorithms into running containers.  (We don’t know the exact mechanism Algorithmia uses here, but we expect it is exploiting these facts.)

They have made some very interesting optimizations.   For example, if the data used in the computation is stored in one of their cloud regions, the docker instance will be instantiated nearby.   Just as important, if an algorithm invokes another algorithm they will attempt to co-locate the two containers and reduce the inter-process latency.  Composability of algorithms is one of their guiding concepts.

Turning your own algorithm into a microservice

The process of turning your own algorithm into a microservice is remarkably simple.   From the Algorithmia portal there is a “+” symbol in the upper right-hand corner.   This give you a dialog box to fill out.   You provide a name of your algorithm, the programming language you are using (from a long list .. but sorry, no Fortran or Julia but there are lots of alternatives), and several other choices including: your source license policy, does your function invoke other Algorithmia functions, does your function invoke things on the open internet?

Answering these questions causes Algorithmia to create a nice GitHub repo for your function.   Your next step is to install the Algorithmia command line interface and then you can clone your functions GitHub repo.  Once you have done that you can edit the function so that it does what you want.   The basic skeleton is already there for you in the “src” directory.   Here is the basic skeleton in Python rendered as a hello world function.

import Algorithmia
# API calls will begin at the apply() method,
# with the request body passed as 'input'
# For more details, see algorithmia.com/developers/algorithm-
# development/languages
def apply(input):
return "hello {}".format(input)


You can edit the function directly from an editor built into the Algorithmia portal or, now that you have a clone of the repo you can use your own tools to transform this skeleton into your algorithm.   If you have done this work on your clone you need to use the Github commands to push your code back to the master.

We tried this with a small experiment.   We built a function called KeyPhrases that takes English language text as input and breaks it down into subjects (s), actions (a) which are like verb clauses and objects (o).   The algorithm is not very useful or sophisticated.   In fact, it uses another Algorithmia microservice  called Parsey McParseface which was originally released by Goolge (see https://arxiv.org/ pdf/1603.06042v1.pdf) .   This is truly a deep parser that build a very sophisticated tree.  For example the figure below illustrates the tree for a pars of the sentence

Einstein’s general theory of relativity explains gravity in terms of the curvature of spacetime.

Parsey McParseface tree output.

Our function KeyPhrases walks the tree and groups the terms, subjects(s), objects(o) and actions (a) and returns a JSON document with the original string and the list of phrases.  It also breaks out separate subphrases with “/” marks.  In this case it returns

{"phrases":[
"s: Einstein's general theory /of relativity ",
"a: explains ",
"s: gravity /in terms /of the curvature /of spacetime. "
],
"text":"Einstein's general theory of relativity explains gravity in terms of the curvature of spacetime."
}


A more complex example is

Facebook Incs chief security officer warned that the fake news problem is more complicated to solve than the public thinks.

The phrase output is

['s: Facebook Incs chief security officer ',
'a: warned ',
'o: that the fake news problem ',
'a: is more ',
'o: complicated and dangerous /to solve /than the public thinks ']


This is clearly not as rich in detail as the Parsey output, but it does extract some useful key phrases.

To complete the creation of the microservice for this algorithm one need only issue the git commands

$git add src/KeyPhrases.py$ git commit -m "added src mods"
$git push origin master The last push causes a compile step to happen and the microservice is now created. Algorithmia also provides an easy template to add documentation and instructions about how to invoke your function. From the Algorithmia editor there is a function that allows you to “publish” your algorithm. After pushing that button, the KeyPhrase example was put in their library. You can see it here: https://algorithmia.com/algorithms/dbgannon/KeyPhrases (If you use it, remember it has not been tested very well, so it may break.) # Algorithmia as an enterprise platform The Algorithmia serverless microservice platform is robust enough that they offer it as an enterprise product. This allows enterprises to host their own version on one of the public clouds or on their own clusters or across multiple cloud in a hybrid system. This allows their own internally used algorithm to be hosted and invoked by their in-house analytics tools and pipelines in a totally scalable way. This enterprise version comes with a management dashboard and monitoring tools. # Conclusions Algorithmia is a fascinating company with very interesting products. It is extremely easy to sign up for a free account and it is fun to use. The team was extremely helpful when we had questions. A Jupyter Notebook with some of the examples mentioned above will be posted very soon. We found experimenting with the various algorithms from an interactive notebook was a pleasure. Creating the hosted version of the KeyPhrases algorithm took less than an hour after the original python code was debugged. In our next experiment we will explore hosting deep learning models with Algorithmia. # CNTK Revisited. A New Deep Learning Toolkit Release from Microsoft In a pair of articles from last winter (first article, second article) we looked at Microsoft’s “Computational Network Toolkit” and compared it to Google’s Tensorflow. Microsoft has now released a major upgrade of the software and rebranded it as part of the Microsoft Cognitive Toolkit. This release is a major improvement over the initial release. Because these older articles still get a fair amount of web traffic we wanted to provide a proper update. There are two major changes from the first release that you will see when you begin to look at the new release. First is that CNTK now has a very nice Python API and, second, the documentation and examples are excellent. The core concepts are the same as in the initial release. The original programming model was based on configuration scripts and that is still there, but it has been improved and renamed as “Brain Script”. Brain Script is still an excellent way to build custom networks, but we will focus on the Python API which is very well documented. Installing the software from the binary builds is very easy on both Ubuntu Linux and Windows. The process is described in the CNTK github site. On a Linux machine, simply download the gziped tared binary and execute the installer. $wget https://cntk.ai/'BinaryDrop/CNTK-2-0-beta2-0-Linux-64bit-CPU-Only.tar.gz’
$tar -xf CNTK-2-0-beta2-0-Linux-64bit-CPU-Only.tar.gz$cd cntk/Scripts/linux
$./install-cntk.sh  This will install everything including a new version of Continuum’s Anaconda Python distribution. It will also create a directory called “repos’’. To start Jupyter in the correct conda environment do the following. $source “your-path-to-cntk/activate-cntk"
$cd ~/repos/cntk/bindings/python/tutorials$Jupyter notebook


A very similar set of commands will install CNTK on your Windows 10 box. (If you are running Jupyter on a virtual machine or in the cloud you will need additional arguments to the Jupyter notebook command such as “-ip 0.0.0.0 –no browser” and then then you can navigate you host browser to the VM ip address and port 8888. Of course, if it is a remote VM you should add a password. ) What you will see is an excellent set of tutorials as shown in Figure 1.

Figure 1.   CNTK tutorial Jupyter notebooks.

## CNTK Python API

CNTK is a tool for building networks and the Python and Brain Script bindings are very similar in this regard.   You use the Python program to construct a network of tensors and then train and test that network through special operations which take advantage of underlying parallelism in the hardware such as multiple cores or multiple GPUs.   You can load data into the network through Python Numpy arrays or files.

The concept of constructing a computation graph for later execution is not new.   In fact, it is an established programming paradigm used in Spark, Tensorflow, and Python Dask.   To illustrate this in CNTK consider the following code fragment that creates two variables and a constructs a trivial graph that does matrix vector multiplication and vector addition.  We begin by creating three tensors that will hold the input values  to the graph and then tie them to the matrix multiply operator and vector addition.

import numpy as np
import cntk
X = cntk.input_variable((1,2))
M = cntk.input_variable((2,3))
B = cntk.input_variable((1,3))
Y = cntk.times(X,M)+B


In this X is a 1×2 dimensional tensor, i.e. a vector of length 2 and M is a matrix that is 2×3 and B is a vector of length 3. The expression Y=X*M+B yields a vector of length 3. However, no computation has taken place. We have only constructed a graph of the computation. To invoke the graph we input values for X, B and M and then apply the “eval’’ operator on Y. We use Numpy arrays to initialize the tensors and supply a dictionary of bindings to the eval operator as follows

x = [[ np.asarray([[40,50]]) ]]
m = [[ np.asarray([[1, 2, 3], [4, 5, 6]]) ]]
b = [[ np.asarray([1., 1., 1.])]]

print(Y.eval({X:x, M: m, B: b}))
array([[[[ 241.,  331.,  421.]]]], dtype=float32)


CNTK has several other important tensor containers but two important ones are

• Constant(value=None, shape=None, dtype=None, device=None, name=”): a scalar, vector or other multi-dimensional tensor.
• Parameter(shape=None, init=None, dtype=None, device=None, name=”): a variable whose value is modified during network training.

There are many more tensor operators and we are not going to go into them here.   However, one very important class is the set of operators that can be used to build multilevel neural networks.   Called the “Layers Library’’ they form a critical part of CNTK.    One of the most basic is the Dense(dim) layer which creates a fully connected layer of output dimension dim. As shown in Figure 2.

Figure 2.   A fully connected layer created by the Dense operator with an implicit  3×6 matrix and a 1×6 vector of parameters labeled here M and B.   The input dimension is taken from the input vector V.  The activation here is the default (none), but it could be set to ReLu or Sigmod or another function.

There are many standard layer types including Convolutional, MaxPooling, AveragePooling and LSTM. Layers can also be stacked with a very simple operator called “sequential’’. Two examples taken directly from the documentation is a standard 4 level image recognition network based on convolutional layers.

with default_options(activation=relu):
conv_net = Sequential ([
# 3 layers of convolution and dimension reduction by pooling
# 2 dense layers for classification
Dense(64),
Dense(10, activation=None)
])


The other fun example is a slot tagger based on a recurrent LSTM network.

tagging_model = Sequential ([
Embedding(150),         # embed into a 150-dimensional vector
Recurrence(LSTM(300)),  # forward LSTM
Dense(labelDim)         # word-wise classification
])


The Sequential operator can be thought of as a concatenation of the layers that in the given sequence.   In the case of the slot tagger network, we see two additional important operators: Embedding and Recurrence.

Embedding is used for word embeddings where the inputs are sparse vectors of size equal to the word vocabulary (item i = 1 if the word is the i-th element of the vocabulary and 0 otherwise) and the embedding matrix is of size vocabulary-dimension by, in this case, 150.     The embedding matrix may be passed as a parameter or learned as part of training.

The Recurrence operator is used to wrap the correct LSTM output back to the input for the next input to the network.

## A Closer Look at One of Tutorials.

The paragraphs above are intended to give you the basic feel of what CNTK looks like with its new Python interface.  The best way to learn more is to study the excellent example tutorials.

### CNTK 203: Reinforcement Learning Basics

CNTK version 1 had several excellent tutorials, but version 2 has the Python notebook versions of these plus a few new ones.  One of the newest demos is an example of reinforcement learning.   This application of Neural Nets was first described in the paper Human-level control through deep reinforcement learning, by the Google DeepMind group.  This idea has proven to be very successful in systems that learn to play games.  This topic has received a lot of attention, so we were happy to see this tutorial included in CNTK. The example is a very simple game that involves balancing a stick.   More specifically they use the cart-pole configuration from OpenAI.   As shown in figure 3, the system state can be described by a 4-tuple: position of the cart, its velocity, the angle of the pole and the angular velocity.   The idea of the game is simple.  You either push the cart to the left or the right and see if you can keep the stick vertical.   If you drift too far off course or the pole angle goes beyond an angle of 15 degrees, the game is over.   Your score is the total number of steps you take before failure. The full example is in the github repository and we are not going to go through all the code here.  The Jupyter notebook for this example is excellent, but if you are new to this topic you may find some additional explanation of value in case you decide to dig into it.

Figure 3. Cart-Pole game configuration.

The part of reinforcement learning used here is called a Deep Q-Network. It uses a neural network to predict the best move when the cart is in a given state. This is done by implicitly modeling a function Q(s,a) which is the optimal future reward given state s and the action is a and where the initial reward is r. They approximate Q(s,a) using the “Bellmann equation” which describes how to choose action a in a given state s to maximize the accumulated reward over time based inductively on the same function applied to the following states s’.

The parameter gamma is a damping factor that guarantees the recurrence converges. (Another excellent reference for this topic is the blog by Jaromír Janisch.) The CNTQ team approached this problem as follows. There are three classes.

• Class Brain.    This hold our neural net and trainer.  There are three methods
• Create() which is called at initialization.   It creates the network.   There are two tensor parameters: observation, which is used to hold the input state and q_target which is a tensor used for training.   The network is nice and simple:
l1 = Dense(64, activation=relu)
l2 = Dense(2)
unbound_model = Sequential([l1, l2])
model = unbound_model(observation)


The training is by the usual stochastic gradient descent based on a loss measure.

loss = reduce_mean(square(model - q_target), axis=0)
meas = reduce_mean(square(model - q_target), axis=0)
learner = sgd(model.parameters, lr,
trainer = Trainer(model, loss, meas, learner)

• Train(x, y)  which calls the trainer for batches of states x and predicted outcomes y which we will describe below
• Predict(s) which invokes the model for state ‘s’ and returns a pair of optimal rewards given a left or right move.
• Class Memory. This hold a record of recent moves.   This is used by the system to create training batches.  There are two methods
• Add(sample configuration)  – adds a four tuple consisting of a starting state, an action and a result and a resulting  state tuple to a memory.
• Sample(n) returns a random sample of n configurations samples from the memory.
• Class Agent which is the actor that picks the moves and uses the memory to train the network.  There are three methods here.
• Act(state) returns a 0 or 1 (left move or right move) that will give the best reward for the given state.     At first it just makes random guesses, but as time passes it uses the Predict method of the Brain class to select the best move for the given state.
• Observe(sample configuration) records a configuration in the memory and keeps track of the time step and another parameter used by act.
• Replay() is the main function for doing the learning.    This is the hardest part to understand in this tutorial. It works by grabbing a random batch of memorized configurations from memory.   What we will do is use the current model to predict an optimal outcome and use that as the next step in training the model.  More specifically for each tuple in the batch we want to turn it into a training sample so that the network behaves like the Bellmann equation.  A tuple consists of the start state, the action, the reward and the following state.   We can apply our current model to predict the award for the start state and also for the result state.  We can use this information to create a new reward tuple for the given action and start state that models the Bellmann recurrence.   Our training example is the pair consisting of the start state and this newly predicted reward.  At first this is a pretty poor approximation, but amazingly over time it begins to converge. The pseudo code is shown below.
x = numpy.zeros((batchLen, 4)).astype(np.float32)
y = numpy.zeros((batchLen, 2)).astype(np.float32)

for i in range(batchLen):
s, a, r, s_ = batch[i]
# s = the original state (4 tuple)
# a is the action that was taken
# r is the reward that was given
# s_ is the resulting state.
# let t = the reward computed from current network for s
# and let r_ be the reward computed for state s_.
# now modify t[a] = r + gamma* numpy.amax(r_)
# this last step emulated the bellmann equation
x[i] = s
y[i] = t
self.brain.train(x,y)


The final part of the program is now very simple. We have an environment object that returns a new state and a done flag for each action the agent takes. We simply run our agent until it falls out of bounds (the environment object returns done=True). If the step succeeded, we increment our score. The function to run the agent and to keep score is shown below.

def run(agent):
s = env.reset()
R = 0
while True:
a = agent.act(s)
s_, r, done, info = env.step(a)
if done: # terminal state
s_ = None
agent.observe((s, a, r, s_))
agent.replay() #learn from the past
s = s_
R += r
if done:
return R


Each time we run “run” it learns a bit more.   After about 7000 runs it will take over 600 steps without failure.

The text above is no substitute for a careful study of the actual code in the notebook.  Also, as it is a notebook, you can have some fun experimenting with it.  We did.

## Final Thoughts

CNTK is now as easy to use as any of the other deep learning toolkits.   While we  have not benchmarked its performance they claim it is extremely fast and it make good use of multiple GPUs and even a cluster of servers.    We are certain that the user community will enjoy using and contributing to its success.

## Citation.

The team that first created CNTK should be cited.   I know there are likely many others that have contributed to the open source release in one way or another, but the following is the master citation.

Amit Agarwal, Eldar Akchurin, Chris Basoglu, Guoguo Chen, Scott Cyphers, Jasha Droppo, Adam Eversole, Brian Guenter, Mark Hillebrand, T. Ryan Hoens, Xuedong Huang, Zhiheng Huang, Vladimir Ivanov, Alexey Kamenev, Philipp Kranen, Oleksii Kuchaiev, Wolfgang Manousek, Avner May, Bhaskar Mitra, Olivier Nano, Gaizka Navarro, Alexey Orlov, Hari Parthasarathi, Baolin Peng, Marko Radmilac, Alexey Reznichenko, Frank Seide, Michael L. Seltzer, Malcolm Slaney, Andreas Stolcke, Huaming Wang, Yongqiang Wang, Kaisheng Yao, Dong Yu, Yu Zhang, Geoffrey Zweig (in alphabetical order), “An Introduction to Computational Networks and the Computational Network Toolkit“, Microsoft Technical Report MSR-TR-2014-112, 2014.

# A Brief Look at Google’s Cloud Datalab

Google recently released a beta version of a new tool for data analysis using the cloud called Datalab.  In the following paragraphs we take a brief look at it through some very simple examples.  While there are many nice features of Datalab, the easiest way to describe it would be to say that it is a nice integration of the IPython Jupyter notebook system with Google’s BigQuery data warehouse.  It also integrates standard IPython libraries such as graphics and scikit-learn and Google’s own machine learning toolkit TensorFlow.

The Google public data sets that are hosted in the BigQuery warehouse are fun to explore.  They include

• The names on all US social security cards for births after 1879.  (The table rows contain only the year of birth, state, first name, gender and number as long as it is greater than 5.  No social security numbers.),
• The New York City Taxi trips from 2009 to 2015,
• All stories and comments from “Hacker News”,
• The US Dept of Health weekly records of diseases reported from each city and state from 1888 to 2013,
• The public data from the HathiTrust and the Internet Book Archive,
• The global summary of the day’s (GSOD) weather from the national oceanographic and atmospheric administration from 9000 weather stations between 1929 and 2016.

And more, including the 1000 genome database.

To run Datalab on your laptop you need to have Docker installed.   Once Docker is running then and you have created a Google cloud account and created a project, you can launch Datalab with simple docker command as illustrated in their quick-start guide.  When the container is up and running you can view it at http://localhost:8081.  What you see at first is shown in Figure 1.  Keep in mind that this is beta release software so you can expect it will change or go away completely.

Figure 1.  Datalab Top level view.

Notice the icon in the upper right corner consisting of a box with an arrow.   Clicking this allows you to login to the Google cloud and effectively giving your authorization to allow you container to run on your gcloud account.

The view you see is the initial notebook hierarchy.   Inside docs is a directory called notebooks that contain many great tutorials and samples.

## A Few Simple Examples of Using Datalab

As mentioned above, one of the public data collections is the list of first names from social security registrations.   Using Datalab we can look at a sample of this data by using one of the built-in Bigquery functions as shown in Figure 2.

Figure 2.   Sampling the names data.

In modern America there is a movement to “post-gender” names.   Typical examples cited on the web are “Dakota”, “Skyler” and  “Tatum”.   A very simple SQL query can be formulated to see how the gender breakdown for these names show up in the data.  In Datalab, we can formulate the query as shown in Figure 3.

Figure 3.   Breakdown by gender of three “post-gender” names.

As we can see, this is very nearly gender balanced.  A closer inspection using each of the three names separately show that “Skyler” tends to be ‘F’ and “Tatum” tends to ‘M’. On the other hand, “Dakota” does seem to be truly post-gender with 1052 ‘F’ and 1200 ‘M’ occurrences.

We can also consider the name “Billy” which, in the US, is almost gender neutral.   (Billy Mitchel was a famous World Work I general and also a contemporary Jazz musician.  Both male. And Billy Tipton and Billy Halliday were female musicians though Billy Halliday was actually named Billie and Billy Tipton lived her life as a man, so perhaps they don’t count.   We can ask how often Billy was used as a name associated with gender ‘F’ in the database?  It turns out it is most common in the southern US. We can then group these by state and create a count and show the top five.   The SQL command is easily inserted into the Datalab note book as shown in Figure 4.

Figure 4.   Search for Billy with gender ‘F’ and count and rank by state of birth.

## Rubella in Washington and Indiana

A more interesting data collection is Center for Disease Control and Prevention dataset concerning diseases reported by state and city over a long period.   An interesting case is Rubella, which is virus also known as the “German measles”.   Through our vaccination programs it has been eliminated in the U.S. except for those people who catch it in other countries where it still exists.  But in the 1960s it was a major problem with an estimated 12 million cases in the US and a significant number of newborn deaths and birth defects.   The vaccine was introduced in 1969 and by 1975 the disease was almost gone.   The SQL script shown below is a slight modified version of one from the Google Bigquery example.   It has been modified to look for occurrences of Rubella in two states, Washington and Indiana, over the years 1970 and 1971.

%%sql --module rubella
SELECT
*
FROM (
SELECT
*, MIN(z___rank) OVER (PARTITION BY cdc_reports_epi_week) AS z___min_rank
FROM (
SELECT
*, RANK() OVER (PARTITION BY cdc_reports_state ORDER BY cdc_reports_epi_week ) AS z___rank
FROM (
SELECT
cdc_reports.epi_week AS cdc_reports_epi_week,
cdc_reports.state AS cdc_reports_state,
COALESCE(CAST(SUM((FLOAT(cdc_reports.cases))) AS FLOAT),0)
AS cdc_reports_total_cases
FROM
[lookerdata:cdc.project_tycho_reports] AS cdc_reports
WHERE
(cdc_reports.disease = 'RUBELLA')
AND (FLOOR(cdc_reports.epi_week/100) = 1970
OR FLOOR(cdc_reports.epi_week/100) = 1971)
AND (cdc_reports.state = 'IN'
OR cdc_reports.state = 'WA')
GROUP EACH BY
1, 2) ww ) aa ) xx
WHERE
z___min_rank <= 500
LIMIT
30000


We can now invoke this query as part of a python statement so we can capture its result as a pandas data frame and pull apart the time stamp fields and data values.

rubel = bq.Query(rubella).to_dataframe()
rubelIN = rubel[rubel['cdc_reports_state']=='IN']
.sort_values(by=['cdc_reports_epi_week'])
rubelWA = rubel[rubel['cdc_reports_state']=='WA']
.sort_values(by=['cdc_reports_epi_week'])
epiweekIN = rubelIN['cdc_reports_epi_week']
epiweekWA = rubelWA['cdc_reports_epi_week']
rubelINval = rubelIN['cdc_reports_total_cases']
rubelWAval = rubelWA['cdc_reports_total_cases']


At this point a small adjustment must be made to the time stamps.  The CDC reports times in epidemic weeks and there are 52 weeks in a year.    So the time stamps for the first week of 1970 is 197000 and the time stamp for the last week is 197051.  The next week is 197100.  To make these into timestamps that appear contiguous we need to make a small “time compression” as follows.

realweekI = np.empty([len(epiweekIN)])
realweekI[:] = epiweekIN[:]-197000
realweekI[51:] = realweekI[51:]-48


Doing the same thing with epiweekWA we now have the basis of something we can graph.  Figure 5 shows the progress of rubella in Washington and Indiana over two years.  Washington is the red line and Indiana is blue.   Note that the outbreaks occur about the same time in both states and that by late 1971 the disease is nearly gone.

Figure 5.   Progress of Rubella in Washington (red) and Indiana (blue) from 1970 through 1971.

Continuing the plot over 1972 and 1973 show there are flare-ups of the disease each year but their maximum size is diminishes rapidly.

(Datalab has some very nice plotting functions, but we could not figure out how to do a double plot, so we used the mathplot library with the “fivethirtheight” format.)

## A Look at the Weather

From the national oceanographic and atmospheric administration we have the global summary of the day’s (GSOD) weather from the from 9000 weather stations between 1929 and 2016.   While not all of these stations were operating during that entire period, there is still a wealth of weather data here.   To illustrate it, we can use another variation on one of Google’s examples.  Let’s find the hottest spots in the state of Washington for 2015.   This was a particularly warm year that brought unusual droughts and fires to the state. The following query will list the hottest spots in the state for the year.

%%sql
SELECT
max, (max-32)*5/9 celsius, mo, da, state, stn, name
FROM (
SELECT
max, mo, da, state, stn, name
FROM
[bigquery-public-data:noaa_gsod.gsod2015] a
JOIN
[bigquery-public-data:noaa_gsod.stations] b
ON
a.stn=b.usaf
AND a.wban=b.wban
WHERE
state="WA"
AND max

The data set ‘gsod2015’ is the table of data for the year 2015.  To get a list that also shows the name of the station we need to do a join with the ‘station’ table over the corresponding station identifiers.  We order the results descending from the warmest recordings.    The resulting table is shown in Figure 6 for the top 10.

Figure 6.   The top 10 hottest spots in Washington State for 2015

The results are what we would expect.   Walla Walla, Moses Lake and Tri Cities are in the eastern part of the state and summer was very hot there in 2015.   But  Skagit RGNL is in the Skagit Valley near Puget Sound.   Why is it 111 degrees F there in September?   If it is hot there what was the weather like in the nearby locations?   To find out which stations were nearby we can look at the stations on a map.   The query is simple but it took some trial and error.

%%sql --module stationsx
DEFINE QUERY locations
SELECT FLOAT(lat/1000.0) AS lat, FLOAT(lon/1000.0) as lon, name
FROM [bigquery-public-data:noaa_gsod.stations]
WHERE state="WA" AND name != "SPOKANE NEXRAD"


It seems that the latitude and longitude for the Spokane NEXRAD station are incorrect and resolve to some point in Mongolia.  By removing it we get a good picture of the nearby stations as shown in Figure 7.

Figure 7.   Location of weather stations in western Washington using the Bigquery chart map function.

This is an interactive map, so we can get the names of the nearby stations.   There is one only a few miles away  called PADILLA BAY RESERVE and the next closest is BELLINGHAM INTL.   We can now compare the weather for 2015 at these three locations.

To get the weather for each of these we need the station ID.   We can do that with a simple query.

%%sql
SELECT
usaf, name
FROM [bigquery-public-data:noaa_gsod.stations]
WHERE
name="BELLINGHAM INTL" OR name="PADILLA BAY RESERVE" OR name = "SKAGIT RGNL"


Once we have our three station IDs we can use the follow to build a parameterized Bigquery expression.

qry = "SELECT max AS temperature, \
TIMESTAMP(STRING(year) + '-' + STRING(mo) + \
'-' + STRING(da)) AS timestamp \
FROM [bigquery-public-data:noaa_gsod.gsod2015] \
WHERE stn = '%s' and max /< 500 \
ORDER BY year DESC, mo DESC, da DESC"

stationlist = ['720272','727930', '727976']

dflist = [bq.Query(qry % station).to_dataframe() for station in stationlist]



We can now render an image of the weather for our three stations as shown in Figure 8.

Figure  8.  Max daily temperatures for Skagit (blue), Padilla Bay (red) and Bellingham (yellow)

We can clearly see the anomaly for Skagit in September and it is also easy to spot another problem in March where the instruments seemed to be not recording.   Other than that there is close alignment of the readings.

## Conclusions

There are many features of Datalab that we have not demonstrated here.   The documentation gives an example of using Datalab with Tensorflow and the charting capabilities are more extensive than demonstrated here.  (The Google maps example here was not reproducible in any other notebook beyond the demo in the samples which we modified to run the code here.)  It is also easy to upload your own data to the warehouse and analyze it with Datalab.

Using Datalab is almost addictive.  For every one of the data collections we demonstrated here there were many more questions we wanted to explore.  For example, where and when did the name “Dakota” start being used and how did its use spread?   Did the occurrence of Rubella outbreaks correspond to specific weather events?  Can we automate the process of detecting non-functioning weather instruments over the years where records exist?  These are all relatively standard data mining tasks, but the combination of Bigquery and IPython in the notebook format makes it fun.

It should be noted that Datalab is certainly not the first use of the IPython notebook as a front-end to cloud hosted analysis tools.  The IPython notebook has been used frequently with Spark as we have previously described.  Those interested in an excellent overview of data science using Python should look at “Python Data Science Handbook”  by Jake VanderPlas which makes extensive use of IPython notebooks.     There are a variety of articles about using Jupyter on AWS  and Azure for data analytics.  A good one is by Cathy Ye about deep learning using Jupyter in the cloud where she gives detailed instruction for how to install Jupyter on AWS and deploy Caffe there.

.

# Kubernetes and the Google Cloud Container Service: Fun with Pods of Celery.

In a previous post I talked about using Mesosphere on Azure for scaling up many-tasks parallel jobs and I promised to return to Kubernetes when I figured out how to bring it up.   Google just made it all very simple with their new Google Cloud container services.   And, thanks to their good tutorials, I learned about a very elegant way to do remote procedure calls using another open source tool called Celery.

So let me set the stage with a variation on an example I have used in the past.   Suppose we have 10000 scientific documents that are stored in the cloud.   I would like to use a simple machine learning method to classify each of these by topic.    I would like to do this quickly as possible and, because the analysis of each document is independent of the others, I can try to process as many as possible in parallel.  This is the basic “many task” parallel model and one of the most common uses of the cloud for scientific computing purposes.      To do this we will use the Celery distributed task queue mechanism to take a list of our documents and send each one to a work queue where the tasks will be parceled out to workers who will do the analysis and respond.

## The Google Cloud Container Service and a few words about Kubernetes.

Figure 1.   Creating a Google cloud cluster and connecting the cloud shell to it.

Interacting with Kubernetes, which is now running on our small cluster, is through command lines which can be entered into the cloud shell.   Kubernetes has a different, and somewhat more interesting architectures than other container management tools.   The basic unit of scheduling in Kubernetes is  launching pods. A pod consists of a set of one or more Docker-style containers together with a set of resources that are shared by the containers in that pod.  When launched a pod resides on a single server or VM.   This has several advantages for the containers in that pod.   For example, because the containers in a pod are all running  on the same VM, the all share the same IP and port space so the containers can find each other through conventional means like “localhost”.   They can also share storage volumes that are local to the pod.

To start let’s consider a simple single container pod to run the Jupiter notebook.  There is a standard Docker container that contains Jupyter and the scipy software stack.  Using the Kubernetes control command kubectl we can launch Jupyter and expose its port 8888 with the following statement.

$kubectl run jupyter --image=jupyter/scipy-notebook --port=8888  To see that it is up and running we can issue the command “kubectl get pods” which will return the status of all of our running pods. Though we have launched jupyter it is still not truly visible. To do that we will associate a load balancer with the pod. This will expose the port 888 to the open Internet. $ kubectl expose deployment jupyter --type=LoadBalancer


Once that has been run you can get the IP address for jupyter from the “LoadBalancer Ingress:” field of the service description when you run the following.   If it doesn’t appear, try again.

$kubectl describe services jupyter  One you have verified that it is working at that address on port 8888, you should shut it down immediately because, as you can see, there is no security with this deployment. Deleting a deployment is easy. $ kubectl delete deployment jupyter


There is another point that one must be aware of when building containers that need to directly interact with the google cloud APIs.  To make this work you will need to get application default credentials to run in your container.   For example if you container is going to interact with the storage services you will need this.   To get the default application credentials follow the instructions here.  We will say a few more words about this below.

## The Analysis Example in detail.

Now to describe  Celery and how to use Celery and Kubernetes in the many-task scenario described above.

To use Celery we start with our analysis program.   We have previously described the analysis algorithms in detail in another post, so we won’t duplicate that here.  Let’s start by assuming we have a function predict(doc) that takes a document as a string as an argument and returns a string containing the result from our trained machine learnging classifiers.  Our categories are “Physics”, “Math”, “Bio”, “Computer Science” and “Finance” and the result from each classifier is simply the category that that classifier determines to be the most likely correct answer.

Celery is a distributed remote procedure call system for Python programs.   The Celery view of the world is you have a set of worker processes running on remote machines and a client process that is invoking functions that are executed on the remote machines.   The workers and the clients all coordinate through a message broker running somewhere else on the network.

Here we use a RabbitMQ service that is running on a Linux VM on the NSF JetStream cloud as illustrated in Figure 2.

Figure 2.  Experimental Configuration with Celery workers running on Kubernetes in the Google Cloud Container Service,  the RabbitMQ broker running in a VM on the NSF Jetstream Cloud and a client program running as a notebook on a laptop.

The code block below illustrates the basic Celery worker template.    Celery is initialized with a constructor that takes the name of the project and a link to the broker service which can be something like a Redis cache or MongoDB.   The main Celery magic is invoked with a special Python “decorator” associated with the Celery object as shown in the predictor.py file below.

from celery import Celery
app = Celery('predictor', backend='amqp')

#Now initialize and load all the data structure that will be constant
#and recused for each analysis.  In our case this will include
#all the machine learning models that were trained on the data
#previously. And create a main worker function to invoke the models.
def invokeMLModels(statement):
....
return analysis

#define the functions we will call remotely here
def predict(statement):
prediction = invokeMLModels(statement)
return [prediction]


What this decorator accomplishes is to wrap the function in a manner that it can be invoked by a remote client.   To make this work we need create a Celery worker from our predictor.py file with the command below which registers a worker instance as a listener on the RabbitMQ queue.

>celery worker -A predictor -b 'amqp://guest@brokerIPaddr'


Creating a client program for our worker is very simple.   It is similar to the worker template except that our version of the predict( ) function does nothing because we are going to invoke it with the special Celery apply_async( ) method that will push the argument to the broker queue and return control immediately to the client.   The object that is returned from this call is similar to what is sometimes called a “future” or a “promise” in the programming language literature.  What it is a placeholder for the returned value.   Once we attempt to evaluate the get() method on this object our client will wait until a reply is returned from the remote worker that picked up the task.

from celery import Celery

def predict(statement):
return ["stub call"]

res = predict.apply_async(["this is a science document ..."])

print res.get()


Now if we have 10000 documents to analyze we can send them in sequence to the queue as follows.

#load all the science abstracts into a list

res = []
for doc in documents:
res.append(predict.apply_async([doc])

#now wait for them all to be done
predictions = [result.get() for result in res]
#now do an analysis of the predictions


Here we push each analysis task into the queue and save the async returned objects in a list.  Then we create a new list by waiting for each prediction value to be returned.   Our client can run anywhere there is Internet access.  For example this one was debugged on a Jupyter instances running on a laptop.  All you need to do is “pip install celery” and run Juypter.

There is much more to say about Celery and the interested reader should look at the Celery Project site for the definitive guide. Let us now turn to using this with Kubernetes.  We must first create a container to hold the analysis code and all the model data.   For that we will need a Docker file and a shell script to correctly launch celery one the container is deployed.   For those actually interested in trying this, all the files and data are in OneDrive here.   The Docker file shown below has more than we need for this experiment.

# Version 0.1.0
FROM ipython/scipystack
MAINTAINER yourdockername "youremail"
RUN easy_install celery
RUN pip install -U Sphinx
RUN pip install Gcloud
RUN easy_install pattern
RUN easy_install nltk
RUN easy_install gensim
COPY bookproject-key.json /
COPY models /
COPY config /
COPY sciml_data_arxiv.p /
COPY predictor.py /
COPY script.sh /
ENTRYPOINT ["bash", "/script.sh"]


To build the image we first put all the machine learning configuration files in a directory called config and all the learned model files in a directory called models.  At same level we have the predictor.py source.   For reasons we will explain later we will also include the full test data set: sciml_data_arxiv.p. The Docker build starts with the ipython/scipystack container.   We then use easy_install to install Celery as well as four packages used by the ML analyzers: pattern, nltk and gensim.   Though we are not going to use the Gcloud APIs here, we include them with a pip install.   But to make that work we need an updated copy of Sphinx.  To make the APIs work we would need our default client authentication keys.   They are stored in a json file called bookproject-key.json that was obtained from the Gcloud portal as described previously.   Finally we copy all of the files and directory to the root path ‘/’.  Note that the copy from a directory is a copy of the all contained files to the path ‘/’ and not to a new directory.   The ENTRYPOINT runs our script which is shown below.

cp /predictor.py .
export C_FORCE_ROOT='true'
echo $C_FORCE_ROOT celery worker -A predictor -b$1


Bash will run our script in a temp directory, so we need to copy our predictor.py file to that directory.  Because our bash is running as root, we need to convince Celery that it is o.k. to do that.  Hence we export C_FORCE_ROOT as true.  Next, if we were using the Gcloud APIs we need to export the application credentials.   Finally we invoke celery but this time we use the -b flag to indicate that we are going to provide the IP address of the RabbitMQ amqp broker as a parameter and we remove it from the explicit reference in the predictor.py file.  When run the predictor file will look for all the model and configuration data in ‘/’.   We can now build the docker image with the command

>docker build -t “yourdockername/predictor” .


And we can test the container on our laptop with

>docker run -i -t “yourdockername/predictor ‘amqp://guest@rabbitserverIP’


Using “-i -t” allows you to see any error output from the container.   Once it seems to be working we can now push the image to the docker hub.   (to use our version directly, just pull dbgannon/predictor)

We can now return to our Google cloud shell and pull a version of the container there.   If we want to launch the predictor container on the cluster, we can do it one at a time with the “kubectl run” command.  However Kubernetes has a better way to do this using a pod configuration file where we can specify the number of pod instances we want to create.  In the file below, which we will call predict-job.json we specify a job name, the container image in the docker hub,   and the  parameter to pass to the container to pass to the shell script.   We also specify the number of pods to create.   In this case that is 6 as identified in the “parallelism” parameter.

apiVersion: batch/v1
kind: Job
name:predict-job
spec:
parallelism: 6
template:
name: job-wq
spec:
containers:
- name: c
image: dbgannon/predictor
restartPolicy: OnFailure


One command in the cloud shell will now launch six pods each running our predictor container.

$kubectl create -f predict-job.json  ## Some Basic Performance Observations. When using many tasks system based on a distributed worker model there are always three primary questions about the performance of the system. 1. What is the impact of wide-area distribution on the performance? 2. How does performance scale with the number of worker containers that are deployed? More specifically, if we N workers, how does the system speed up as N increases? Is there a point of diminishing return? 3. Is there a significant per/task overhead that the system imposes? In other words, If the total workload is T and if it is possible to divide that workload into k tasks each of size T/k , then what is the best value of k that will maximize performance? Measuring the behavior of a Celery application as a function of the number workers is complicated by a number of factors. The first concern we had was the impact of widely distributing the computing resources on the overall performance. Our message broker (RabbitMQ) was running in a virtual machine in Indiana on the JetStream cloud. Our client was running a Jupyter notebook on a laptop and the workers were primarily on the Midwest Google datacenter and on a few on other machines in the lab. We compared this to a deployment where all the workers, the message broker and the client notebook were all running together on the Google datacenter. Much to our surprise there was little difference in performance between the two deployments. There are two ways to view this result. One way is to say that the overhead of wide area distribution was not significant. The other way to say this is that the overhead of wide area distribution was negligible compared to other performance problems. A second factor that has an impact on performance as a function of the number of workers is the fact that a single Celery worker may have multiple threads that are responding to asynchronous function calls. While we monitored the execution we noticed that the number of active threads in one worker could change over time. This made performance somewhat erratic. Celery’s policy is that it will never have more threads than the number of available cores, so to limit the thread variability we ran workers in container pods on VMs with only one core. Concerning the question of the granularity of the work partitioning we configured the program so that a number of documents could be processed in one invocation and this number could be set remotely. By taking a set of 1000 documents and a fixed set of workers, we divided the document set into blocks of size K where K ranged from 1 to 100. In general, larger blocks were better because the number of Celery invocations was smaller, but the difference was not great. Another factor involved Celery’s scheduling for deciding which worker get the next invocation. For large blocks this was not the most efficient because this left holes in the execution schedule when workers were occasionally idle while another was over scheduled. For very small blocks these holes tended to be small. We found that a value of K=2 gave reasonably consistent performance. Finally to test scalability we used three different programs. 1. The document topic predictor described above where each invocation classified two documents. 2. A simple worker program that does no computation but just sleeps for 10 seconds before returning a “hello world” string. 3. A worker that computes part of the Euler sequence sum(1/i**2, i=1..n) where n = 109 . Each worker computes a block of 107 terms of the sequence and the 100 partial results are added together to get the final result (which approximates pi2/6 to about 7 decimal places). The document predictor is very computational intensive and uses some rather large data matrices for the trained machine learning models. The size of these arrays are about 150 megabytes total. While this does fit in memory, the computation is going to involve a great deal of processor cache flushing and there may be memory paging effects. The example that computes the Euler sum requires no data other than the starting point index and the size of the block to sum. It is pure computation and it will have no cache flushing or memory paging effects, but it will keep the CPU very busy. The “sleep” example leaves the memory and the CPU completely idle. We ran all three with one to seven workers. (6 workers using 6 cores from the small Google demo account and one on another other remote machine). To compare the results, we computed the time for each program on one worker and plotted the speed-up ratio for 2, 3, 4, 5, 6 and 7 workers. The results are shown in the graphs below. Figure 3. Performance as speed-up for each of the three applications with up to 7 workers. As can be seen, the sleeper scales linearly in the number of workers. In fact, when executed on multi-core machines it is almost super-linear because of the extra threads that can be used. (It is very easy for a large number of threads to sleep in parallel.) On the other hand, the predictor and the Euler examples reached a maximum speed up with around five workers. Adding more worker pods to the servers did not show improvement because these applications are already very compute intensive. This was a surprise as we expect all three experiments to scale well beyond seven workers. Adding more worker pods to the servers did not show improvement because these applications are already very compute intensive. When looking for the cause of this limited performance, we considered the possibility that the RabbitMQ broker was a bottleneck, but our previous experience with it has allowed us to scale applications to dozens of concurrent reader and writers. We are also convinced that the Google Container Engine performed extremely well and it was not the source of any of these performance limitations. We suspect (but could not prove) that the Celery work distribution and result gathering mechanisms have overheads that limit scalability as the number of available workers grows. ## Conclusion Google has made it very easy to deploy containerized applications using Kubernetes on their cloud container service. Kubernetes has some excellent architectural features that allow multiple containers to be co-located on a single server within a pod. We did not have time here to demonstrate this, but their documentation gives some excellent examples. Celery is an extremely elegant way to do remote procedure calls in Python. One only needs to define the function and annotate it with a Celery object. It can then be remotely invoked with an asynchronous call that returns control to the caller. A future like object is returned. By calling a special method on the returned object the caller will pause until the remote call completes and the value is provided to the caller. Our experiments demonstrated that Celery has limited scalability if it is used without modification and with the RabbitMQ message broker. However, celery has many parameters and it may be possible that the right combinations will improve our results. We will report any improved results we discover in a later version of this document. # The State of the Cloud: Evolving to Support Deep Learning and Streaming Data Analytics and Some Research Challenges (Note: This is an updated version on 7/21/2016. The change relates to containers and HPC and it is discussed in the research topics at the end.) I was recently invited to serve on a panel for the 2016 IEEE Cloud Conference. As part of that panel I was asked to put together 15 minutes on the state of cloud technology and pose a few research challenges. Several people asked me if I had published any of what I said so I decided to post my annotated notes from that mini-talk here. The slide deck that goes along with this can be found here. There were three others on the panel who each made some excellent points and this document does not necessarily reflect their views. Cloud computing has been with us for fifteen years now and Amazon’s Web Services have been around for ten. The cloud was originally created to support on-line services such as email, search and e-commerce. Those activities generated vast amounts of data and the task of turning this data into value for the user has stimulated a revolution in data analytics and machine learning. The result of this revolution has been powerful and accurate spoken language recognition, near real-time natural language translation, image and scene recognition and the emergence of a first generation of cloud-based digital assistants and “smart” services. I want to touch on several aspects of cloud evolution related to these exciting changes. ## Cloud Architecture Cloud architectures have been rapidly evolving to support these computational and data intensive tasks. The cloud data centers of 2005 were built with racks of off-the-shelf server and standard networking gear, but the demands of the new workloads described above are pushing the cloud architects to consider some radically different approaches. The first changes were the introduction of software defined networks that greatly improved bisection bandwidth. This also allowed the data center to be rapidly reconfigured and repartitioned to support customer needs as well as higher throughput for parallel computing loads. Amazon was the first large public cloud vendor to introduce GPUs to better support high-end computation in the cloud and the other providers have followed suit. To accelerate the web search ranking process, Microsoft introduced FPGA accelerators and an overlay mesh-like network which adds an extra dimension of parallelism to large cloud applications. The advent of truly large scale data collections made it possible to train very deep neural networks and all of the architectural advances described above have been essential for making progress in this area. Training deep neural nets requires vast amounts of liner algebra and highly parallel clusters with multiple GPUs per node have become critical enablers. Azure now support on-demand clusters of nodes with multiple GPUs and dedicated InfiniBand networks. The FPGAs introduced for accelerating search in the Microsoft data centers have also proved to be great accelerators for training convolutional neural networks. GPUs are great for training deep networks but Nirvana has designed a custom ASIC that they claim to be a better accelerator. Even Cray is now testing the waters of deep learning. To me, all of these advances in the architecture of cloud data centers points to a convergence with the trends in supercomputer design. The future exascale machines that are being designed for scientific computing may have a lot in common with the future cloud data centers. Who knows? They may be the same. ## Cloud System Software The software architecture of the cloud has gone through a related evolution. Along with software defined networking we are seeing the emergence of software defined storage. We have seen dramatic diversification in the types of storage systems available for the application developer. Storage models have evolved from simple blob stores like Amazon’s S3 to sophisticated distributed, replicated NoSQL stores designed for big data analytics such as Google’s BigTable and Amazon’s DynamoDB. Processor virtualization has been synonymous with cloud computing. While this is largely still true, container technology like Docker has taken on a significant role because of its advantages in terms of management and speed of deployment. (It is worth noting that Google never used traditional virtualization in their data centers until their recent introduction of IaaS in GCloud.) Containers are used as a foundation for microservices; a style of building large distributed cloud applications from small, independently deployable components. Microservices provide a way to partition an application along deployment and language boundaries and they are well suited to Dev-Ops style application development. Many of the largest applications running on the cloud by Microsoft, Amazon and Google are composed of hundreds to thousands of microservices. The major challenges presented by these applications are management and scalability. Data center operating systems tools have evolved to coordinate, monitor and attend to the life-cycle management of many concurrently executing applications, each of which is composed of vast swarms of containerized microservice. One such systems is Mesos from Mesosphere. ## Cloud Machine Learning Tools The data analytics needed to create the smart services of the future depend upon a combination of statistical and machine learning tools. Bayesian methods, random forests and others have been growing in popularity and are widely available in open source tools. For a long time, neural networks were limited to three levels of depth because the training methods failed to show improvements for deeper networks. But very large data collections and some interesting advances in training algorithms have made it possible to build very accurate networks with hundreds of layers. However, the computation involved in training a deep network can be massive. The kernels of the computation involve the dense linear algebra that GPUs are ideally suited and the type of parallelism in the emerging cloud architecture is well suited to this task. We now have a growing list of open source machine learning toolkits that have been recently released from the cloud computing research community. These include Amazon’s Tensorflow, AzureML, Microsoft Research Computational Network Tool Kit (CNTK), Amazon’s Deep Scalable Sparse Tensor Network Engine (DSSTNE), and Nervana’s NEON. Of course the academic research community has also been extremely productive in this area. Theano is an important Python toolkit that has been built with contributions from over a dozen universities and institutes. Figure 1. cloud ML tools and services stack Not every customer of cloud-based data analytics wants to build and train ML models from scratch. Often the use cases for commercial customers are similar, hence another layer of services has emerged based on pre-trained models. The use cases include image and language recognition, specialized search, and voice-driven intelligent assistants. As illustrated in Figure 1, these new services include Cortana (and MSR project Oxford components), Google ML, Amazon Alexa Skills Kit, IBM Watson Services and (using a different style cloud stack) Sentient Aware. ## Streaming Data Analytics Services There are several “exponentials” that are driving the growth of cloud platforms and services. These include Big Data, mobile apps, and the Internet of things. The ability to analyze and act on data in motion is extremely important for application area including urban informatics, environmental and ecological monitoring and recovery, analysis of data from scientific experiments and web and data center log analysis. The Cloud providers and open source research community has developed a host of new infrastructure tools that can be used to manage massive streams of data from remote sources. These tools can be used to filter data streams, do on-line analysis and use the backend cloud machine learning services. The tools include Spark Streaming, Amazon Kinesis, Twitter Heron, Apache Flink, Google Dataflow/Apache Beam and the Azure Event hub and data lake. A more detailed analysis of these tools can be found here. ## A Few Research Challenges As was evident at the IEEE cloud conference, there is no shortage of excellent research going on, but as promised here are a few topics I find interesting. 1. Cloud Data Center Architecture. If you are interested in architecture research the Open Compute Project has a number of challenging projects that are being undertaken by groups of researchers. They were founded by people from companies including Facebook, Intel, Google, Apple, Microsoft, Rackspace, Ericsson, Cisco, Juniper Networks and more and they have contributed open data center designs. And it is open, so anybody can participate. 2. Cloud & Supercomputer convergence. As the sophistication of the cloud data centers approach that of the new and proposed supercomputers it is interesting to look at what architectural convergence might look like. For example, which modes of cloud application design will translate to supercomputers? Is it possible that the current microservice based approach to interactive cloud services could be of value to supercomputer centers? Can we engineer nanosecond inter-container messaging? Can we do a decent job of massive batch scheduling on the cloud with the same parallel efficiency as current supercomputers? Update: It seems that there is already some great progress on this topic. The San Diego Supercomputer Center has just announced deployment of Singularity on two of their big machines. Singularity is a special container platform from Gregory M. Kurtzer of LBNL. There is a great article by Jeff Layton that gives a nice overview of Singularity. 3. Porting Deep Learning to Supercomputers. There is currently serious interest in doing large scale data analytics on large supercomputers such as those at the national centers. Some believe that the better algorithms will be available with these advance parallel machines. Can we compile tensorflow/CNTK/ DSSTNE using MPI for exascale class machines? In general, are there better ways to parallelize NN training algorithms for HPC platforms? 4. The current open source stream analytics platforms describe above are designed to handle massive streams of events that are each relative small. However, many scientific event streams are more narrow and have event object that may be massive blobs. What would it take to modify the open source streaming tools to be broadly applicable to these “big science” use cases. I welcome feedback on any of the items discussed here. Many of you know more about these topics than I, so let me know where you think I have incorrectly or overstated any point. # Fun with Recurrent Neural Nets: One More Dive into CNTK and TensorFlow In a previous article I set about comparing Microsoft’s Computational Network Took Kit for deep neural nets to Google’s TensorFlow. I concluded that piece with a deep dive into how recurrent neural nets (RNNs) were represented in each system. I specifically went after the type of RNNs known by the strange name of Long Short-Term Memory (LSTM) networks. I wanted to learn a bit more about how these systems worked. I decided to treat them like laboratory specimens so that I could poke and prod them to see what I could learn and what I could get them to do. This article is essentially my lab notebook. Warning: With the exception of a bit toward the end, this is not technically very deep. In fact, I did not discover anything that has not been extensively reported on elsewhere. But I learned a lot and had some fun. Perhaps it will be of interest to students just starting to learn about this subject. Before I get to far into this, I would like to mention that I recently discovered an excellent series of tutorials on RNNs by Denny Britz that are definitely worth reading. # CNTK’s LSTM and Hallucinating Bloomberg Financial News One of the many good examples in CNTK is language modeling exercise in Examples/Text/PennTreebank. The documentation for this one is a bit sparse and the example is really just of a demo for how easy it is to use their “Simple Network Builder” to define a LSTM network and train it with stochastic gradient decent on data from the Penn Treebank Project. One command starts the learning: cntk configFile=../Config/rnn.cntk Doing so trains the network, tests it and saves the model. However, to see the model data in an easily readable form you need a trivial addition to the configfile: you need to add the following dumpnode command to put a dump file a directory of your choosing. dumpnode=[ action = "dumpnode" modelPath = "$ModelDir$/rnn.dnn" outputFile = "$OutputDir$/modeltext/dump" ]  This creates a big text file with all the trained data. To experiment with the trained model, I decided to load it into a python notebook and rebuild the LSTM network from the defining equations. From the CNTK book those equations are I was pleased to see that the dumped model text had the same W and b tensors names as in the equations, so my job was relatively easy. I extracted each of the tensors and saved them into a file (I will make these available in Github). The python code for the LSTM based on the equations above is below. def rnn(word, old_h, old_c): Xvec = getvec(word, E) i = Sigmoid(np.matmul(WXI, Xvec) + np.matmul(WHI, old_h) + WCI*old_c + bI) f = Sigmoid(np.matmul(WXF, Xvec) + np.matmul(WHF, old_h) + WCF*old_c + bF) c = f*old_c + i*(np.tanh(np.matmul(WXC, Xvec) + np.matmul(WHC, old_h) + bC)) o = Sigmoid(np.matmul(WXO, Xvec)+ np.matmul(WHO, old_h)+ (WCO * c)+ bO) h = o * np.tanh(c) #extract ordered list of five best possible next words q = h.copy() q.shape = (1, 200) output = np.matmul(q, W2) outlist = getwordsfromoutput(output) return h, c, outlist  As you can see, this is almost a literal translation of the equations. The only different is that this has as input a text string for the input word. However the input to the equations is a vector encoding of the word. The model generates the encoding matrix E which has the nice property that the ith column of matrix corresponds to the word in the ith position in the vocabulary list. The function getvec(word, E) takes the embedding tensor E, and looks up the position of the word in the vocabulary list and returns the column vector of E that corresponds to that word. The output of one pass through the LSTM cell is the vector h. This is a compact representation of the words likely to follow the input text to this point. To convert this back into “vocabulary” space we multiply it by another trained vector W2. The size of our vocabulary is 10000 and the vector output is that length. The ith element of output represents the relative likelihood that that ith word is next word to follow the input so far. Getwordsfromoutput simply returns the top 5 candidate words in order of likelihood. Before going further, it is worth looking closer at the properties of the word embedding matrices E and W2. There is a fascinating paper by Mikolov, Yih and Zweig entitled “Linguistic Regularities in Continuous Space Word Representations” where they suggest that the embedding space for word has several interesting properties. I decided to investigate that. Their point is that words that are similar in a linguistic sense will be nearby in the embedding space. For example, present tense verbs should be near other present tense verbs and singular nouns should be near each other, etc. I decided to try that. However, there are two embedding mappings. One is based on the tensor E and the other based on the W2 tensor. E has dimension 150 by 10000 and W2 is 200 by 10000. The difference in dimensionality are because of arbitrary decisions made in defining the hidden layers in the network. But both represent word imbeddings. I experimented with both. I wrote a function getnear(word, M) which takes a word and looks for the 5 most nearby words in the space where M is transpose of either E or W2. (I used cosine distance as the metric.) Verb tense locality and noun plurals worked best in the W2 space as illustrated below. These are only illustrations. For a deeper statistical analysis look at the Mikolov paper. A more interesting conjecture from their study was that there may be some linearity in these embedding that might allow one to try simple analogies of the form “A is to B, as C is to __”. Their idea is that if a, b and c are the vector embeddings of the words A, B and C, then the embedding of “__” may be computed as d = c + (b-a). So I wrote a little function AistoBasCisto(A, B, C) that does this computation. In the results I had to delete A, B and C from the candidate answers because they came up often as nearby. In this case my results were less encouraging. It worked better with the E space than with W2. For example, for E we have And for the W2 space the results looked like As you can see the “run running walk __” example failed with E but was close, but still incorrect, with W2. You may wonder why these particular words came up. The data we used to train the system came from a small subset of the Penn TreeBank collection as provided in the CNTK package. It is heavily dominated by financial news items. This explains why the plural of person could be managers or customers. A larger vocabulary and data collection would be needed to truly test the analogy by linearity conjecture. #### Now on to hallucinating the financial news. Now to test the LSTM as a truly recurrent network. We provide the network with a starting word and let it suggest the next word. And then we repeat this process constructing a “sentence”. In the code below we randomly pick one of the top three suggest by the network as the next word. c = np.zeros(shape = (200, 1)) h = np.zeros(shape = (200, 1)) output = np.zeros(shape = (10000, 1)) word = 'my' sentence= word for _ in range(40): h, c, outlist = rnn(word, h, c) word = outlist[randint(0,3)] sentence = sentence + " " +word print sentence+"."  In this case we start with the word “my” and let it generate a 40 word sentence. The output is my new rules which would create an interest position here unless there should prove signs of such things too quickly although the market could be done better toward paying further volatility where it would pay cash around again if everybody can.  This is a great example of hallucinating financial news. Let’s try it again starting with the word “president”. president michael de brown wrote himself against democratic union law which represents an emergency relief agreement during a new york state district or early tuesday before july after a federal government agency created early losses without mr. krenz or perhaps.  Now with the word “the”. the company reported third-quarter results reflecting a number compared between N barrels including pretax operating loss from a month following fiscal month ending july earlier compared slightly higher while six-month cds increased sharply tuesday after an after-tax loss reflecting a strong.  The “sentences” end rather abruptly because of the 40 word limit I set. If you let it go it will run on until the state vector for the sentence seems to break down. Try this yourself. To make it easy to play with this example, I have put the code in GitHub. The trained model text files are in OneDrive and is a zipped file of about 50MB. There are many more excellent and fun examples. Andrej Karpathy has a great blog article showing how RNNs can mimic Shakespeare, or Latex science articles and many more. ## TensorFlow’s seq2seq French Lesson. One of the most interesting examples in the TensorFlow tutorials is an English to French translator. As with the CNTK example it was trivial to start the translator learning following the instructions in the tutorial. After letting this run for about a week, I wanted to see how well it would do. As with the CNTK example, I created a Jupyter IPython notebook and loaded the trained model. I will explain how that was done in more detail below but, for now, I will show how we can invoke it to test its translation ability. This particular trained model was not very big and with a relatively small data set, so I didn’t expect much. In fact, as you will see, to a French speaker it is a disaster. On the other hand, it learned more French in a week of training that I did in three semesters of French in college. (For full disclosure, this was my weakest subject in college and my grade was a hard-fought “C” each semester.) The code below demonstrates how the model is invoked. First you have to tokenize the input sentence. The algorithm uses a system of buckets of fixed sizes to make the training more efficient. You next find the smallest bucket that can contain your sentence and convert this to the input vector list needed by the model. The step function takes a Tensorflow session, the input vector list and a null list of decoder inputs (to be explained later) and generates a list of vectors as outputs. Each vector represents the likelihood that individual vocabulary words are the correct word at that point in the translated sentence. We pick the most likely and print the sentence. sentence = " I am not the president of France. " token_ids = data_utils.sentence_to_token_ids(sentence, en_vocab) # Which bucket does it belong to? bucket_id = min([b for b in xrange(len(_buckets)) if _buckets[b][0] > len(token_ids)]) # Get a 1-element batch to feed the sentence to the model. encoder_inputs, decoder_inputs, target_weights = model.get_batch({bucket_id: [(token_ids, [])]}, bucket_id) _, _, output_logits = model.step(sess, encoder_inputs, decoder_inputs, target_weights, bucket_id, True) outputs = [int(np.argmax(logit, axis=1)) for logit in output_logits] print(" ".join([rev_fr_vocab[output] for output in outputs])) Je ne suis pas le président de la France .  This example is not too bad. However, if I ask “In which city does the president of France live?” I get “Dans quelle ville le président de la France ?”. This is not exactly correct. If I feed this into Google translate and ask what this means in English I get “In which city the President of France?”. If I give it this one, What is the name of a good restaurant? The system responds with Quel est le nom d’une bonne bonne bonne ?” Which translates back to “What is the name of a good good good?”. Probably not very helpful on the streets of Paris. It turns out restaurant is not in the tiny training vocabulary used here. Finally, given this sentence ” The article stated that the President of the United States is here today. “ The translator returned Le paragraphe a indiqué que le président des États-Unis est aujourd ‘ hui aujourd ‘ hui .” The end of this reply is “is today today”. As I said, this is still much better than I could do with my college French. However, as you can see from the previous two examples, our little translator runs out of gas at the end of sentences and tends to repeat itself. You should try this yourself. I have put the notebook file in github or you can execute these directly from the Tensorflow python code. All you need to do is train the model from TensorFlow and run the notebook with the path to the model output directory. While loading and using the trained model was easy and fun, understanding the seq2seq model used in this example takes a bit of work. So this part of this article will get a bit more technical. The TensorFlow translate program is based on a sequence-to-sequence model constructed from more primitive recurrent neural nets. By sequence-to-sequence we mean a network that takes a sequence as input and produces a sequence as output. It consists of two parts: an “encoder RNN” and a “decoder RNN” as shown in Figure 1 below. Figure 1. A sequence-to-sequence RNN English to French translator with the encoder and decoder unrolled to show the flow of messages. In this figure the RNNs are “unrolled” to show the flow of messages. The state vector at the end of the encoder is a vector embedding of the input sentence. This state vector is used to start the decoder along with a “GO” token. The diagram shows the network after it has been trained. During training the inputs to the decoder are the French version of the English sentences. I won’t talk about the training here because is enough to try to understand how this works. Before I go any further I want to point you to some important papers. Sutskever, Vinyals and Le published an early important paper on sequence to sequence models that is worth reading. To understand how it is built the network we need to dig into the code a bit. The building blocks are a set of classes of base type RNNCell with specializations 1. BasicRNNCell 2. GRUCell 3. BasicLSTMCell 4. LSTMCell 5. OutputProjectionWrapper 6. InputProjectionWrapper 7. EmbeddingWrapper 8. MultiRNNCell The ones we will see used here are GRUCell, MultiRNNCell and EmbeddingWrapper. We discussed LSTMCell in our previous article but we need to look at GRUCell here because that is the one used in the example. The GRUCell is a “Gated Recurrent Unit” invented by Cho et. al. in “Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation”. The “gated” phrase comes from the way the output is defined as coming mostly from the previous state or from a combination with the new input. The diagram below tries to explain this a bit better. Figure 2. GRU wiring diagram It also helps to see it in terms of the defining equations. The quantity ut is a gate vector. Recall the sigmoid function switches sharply between one and zero. So when ut is one then h is just a copy of the old h and we are ignoring the input x it is based on the value ct. The gate rt is determines how much of the old state goes into defining the value of ct. To understand how this is encoded in TensorFlow you need to understand the function. linear(args, output_size, bias, bias_start=0.0, scope=None)  where args is a list of tensors each of size batch x n . Linear computes sum_i(args[i] * W[i]) + bias where W is a list of matrix variables of size n x outputsize and bias is a variable of size outputsize. In the equations above we have represented linear algebra as a matrix times a column vector. Tensorflow uses the transpose notation: row vector on the left times the transpose of the matrix. So in linear the args are a list of row vectors. Where is the matrix W and offset b? This is fetched from memory based on the variable current scope, because W and b are variable tensors that are learned values. If you look at the first two equations above, you will see they are almost identical. In fact, we can write them as If you transpose the last one from column form into row form you can now compute both with one invocation of the linear function. The code for the GRUCell is below. As you can see they have encoded one pass through the GRU cell with only two matrix vector multiplies. You can also see that the way the variable scope is used to pick out the W’s for the gates and the W for the state/output. Another point to remember that an invocation of the “__call__ function operator does not cause the tensor to execute the operation, rather it builds the graph. class GRUCell(RNNCell): def __init__(self, num_units): self._num_units = num_units ... stuff deleted .... def __call__(self, inputs, state, scope=None): with vs.variable_scope(scope or type(self).__name__): with vs.variable_scope("Gates"): # Reset gate and update gate. # We start with bias of 1.0 to not reset and not udpate. r, u = array_ops.split(1, 2, linear([inputs, state], 2 * self._num_units, True, 1.0)) r, u = sigmoid(r), sigmoid(u) with vs.variable_scope("Candidate"): c = tanh(linear([inputs, r * state], self._num_units, True)) new_h = u * state + (1 - u) * c return new_h, new_h  The top level class we invoke for building our model is seq2seqModel. When we create an instance of this class it sets in motion a set of flowgraph building steps. I am going to skip over a lot of stuff and try to give you the big picture. The first graph building step in the initialization of an instance of this object is # Create the internal multi-layer cell for our RNN. single_cell = rnn_cell.GRUCell(size) … if num_layers > 1: cell = rnn_cell.MultiRNNCell([single_cell] * num_layers)  As you can see we are creating a GRU cell graph generator instance and making a list of num_layers of this object and passing that to the constructor for MultiRNNCell. In our case, num_layers has been set to 2. MultiRNNCell is pretty easy to understand. It builds a graph consisting of a stack of (in this case) GRU cells where the output state vector of each level is fed to the input of the level above it. This new compound cell has an output that is the state of the top sub-cell and whose output state is the concatenation of the output states of all the sub-cells. The next part is not so easy to follow. We will take our MultiRNNCell graph builder and use it to create and encoder and a special decoder. But first we must make a short digression. ## Paying Attention There is a problem that is encountered in the sequence-to-sequence model. The encoder encodes the entire sentence into a state vector which is used by the decoder as its input. That state vector is an abstract representation of our entire sentence as a single point in a very high dimensional space. The decoder has been trained to use that point as a starting point to unroll a translated version of the sentence. I find the fact that it works at all to be rather remarkable. It is as if the decoder takes the English state vector and transforms it into a similar point in “French” space. Unfortunately, the longer the input sentence, the more difficult it is to decode it. How much information can we pack into one point? The problem is that at each decoding step we need a little bit more information than is provided by the state vector as it passes through the decoder loop. The idea used here is to help the decoder by providing it a bit of focus derived from the input sequence at each stage of the decoder loop. This is generally referred to as “attention”, as in “at this step of decoding please pay attention to what the encoder was doing here”. Bahdanau, Cho and Bengio had an early paper about this that used a bidirectional pass over the input sequence. As they put it, they wanted to “automatically (soft-)search for parts of a source sentence that are relevant to predicting a target word”. (Denny Britz has a lovely blog article about attention and describes several fascinating applications. It is well worth reading.) The mechanism for attention used in the TensorFlow example is based on a paper by Vinyals et. al. and we will follow that one here. The key idea is rather than take the single final state vector from the encoder, let’s collect the state vectors at each stage of the encoder. Following Vinyals, let the encoder state vectors for each input word be And let the decoder state vectors be Then for each decoder time step t compute Where the Ws are learned matrices and v is a learned vector. Then as the input to the t+1 state vector of the decoder we use the concatenation The idea is this new state vector at time t+1 puts much more focus on the corresponding words in in the encoder string. This all happens in a function called seq2seq.attention_decoder that is called in another constructor function seq2seq.embedding_attention_seq2seq that wraps and an embedding around a graph generated by our MultiCellRNN graph builder to generate the final decoder graph. These graphs are all stitched together in the Seq2SeqModel constructor. It is fair to say that there are many levels of abstraction here that are used to build the decoder and link it to the encoder. I am leaving out many details that are critical for the training such as the part that implements the bucket handler. The final graph, in its most abstract form is pictured below in figure 3. Figure 3. The Translate.py sequence to sequence translator is based on a two level GRU cell encoder and an attention-augmented two level GRU cell decoder. The input English is entered in reverse order as an optimization ## Final Thoughts As I have said above, I have not included all the details of how the seq2seq translator is put together, but I tried to include the highlights that I found most interesting. I encourage you to dive into the code and discover the rest. You will likely find some errors in what I described above. If so, please let me know. There is really a lot of exciting results that have come out in the last few years relating to RNNs. For example, Lei Ba, Mnih and Kavukcuoglu demonstrated that RNNs with attention can be applied to interesting image analysis challenges, such as reading the house number from a street scene. In “Teaching Machines to Read and Comprehend” Hermann et. al. excellent paper demonstrate the use of an attentive RNN build to answer simple questions about text. I personally don’t think any RNN can pass a Turing test yet, so it ain’t A.I. But these little statistical machines are certainly wonderful mimics and they can speak better French than I. # TensorFlow Meets Microsoft’s CNTK Updated April 4, 1017. Much of this material has been updated and improved and now appears as Chapter 10, Cloud Computing for Science and Engineering. It can be accessed at the book’s website. Update Nov 10, 2016. Microsoft now has a new release of CNTK. We have a post now that provides a quick look at this new version. Go read that one instead of this one. Update Oct 25, 2016. this post describes the early version of CNTK. Microsoft just released a very nice new version called the cognitive toolkit. I would not base your impression of CNTK on the following post. I’ll update this as soon as I have time. ———————– CNTK is Microsoft’s Computational Network Toolkit for building deep neural networks and it is now available as open source on Github. Because I recently wrote about TensorFlow I thought it would be interesting to study the similarities and differences between these two systems. After all, CNTK seems to be the reigning champ of many of the image recognition challenges. To be complete I should also look at Theano, Torch and Caffe. These three are also extremely impressive frameworks. While this study will focus on CNTK and TensorFlow I will try to return to the others in the future. Kenneth Tran has a very nice top level (but admittedly subjective) analysis of all five deep learning tool kits here. This will not be a tutorial about CNTK or Tensorflow. Rather my goal is to give a high level feel for how they compare from the programmer’s perspective. This is not a performance analysis, but rather a programming model analysis. There is a lot of code here, so if you don’t like reading code, skip to the conclusions. CNTK has a highly optimized runtime system for training and testing neural networks that are constructed as abstract computational graphs. In that sense, CNTK is very much like TensorFlow. However, there are some fundamental differences. To illustrate these features and differences I will take two standard examples that are included with both systems and work through the approach taken by each system. The first example is a not-too-deep convolutional neural net solution to the standard MNIST handwritten digit recognition example. I will conclude with some comments about how they differ in their approach in the case of recurrent neural networks. Both TensorFlow and CNTK are basically script-driven. By this I mean that the construction of the neural network flow graph is described in a script and the training is done using some very clever automated processes. In the case of TensorFlow the script is embedded in the Python language and Python operators can be used to control the flow of execution of the computational graph. CNTK does not currently have a Python or C++ binding (though one is promised) so currently the control flow of the execution of the training and testing is highly choreographed. As I will show, this is not as much of a limitation as it sounds. There are actually two scripts associated with a CNTK network: a configuration file that controls the training and test parameters and a network definition language file for constructing the network. I’ll start with the description of the neural network flow graph because that is where the similarity to TensorFlow is the greatest. There are two ways to define the network in CNTK. One approach is to use the “Simple Network Builder” that will allow you to create some simple standard networks by specifying only a few parameter settings. The other is to use their Network Definition Language (NDL). The example here (taken directly from their download package in Github) uses NDL. Below is a slightly abbreviated version of the Convolution.ndl file. (I have used commas to put multiple lines on one line to fit the page better.) CNTK network graphs have a set of special nodes. These are FeatureNodes and LabelNodes that describe the inputs and training labels, CriterionNodes and EvalNodes that that are used for training and result evaluation, and OutputNodes that represent the outputs of the network. I will describe these below as we encounter them. At the top of the file we have a set of macros that are used to load the data (features) and labels. As can be seen below we read images of the MNIST digits as features which are now arrays of floating point numbers that we have scaled by a small scalar constant. The resulting array “featScaled” will be used as input to the network. load = ndlMnistMacros # the actual NDL that defines the network run = DNN ndlMnistMacros = [ imageW = 28, imageH = 28 labelDim = 10 features = ImageInput(imageW, imageH, 1) featScale = Const(0.00390625) featScaled = Scale(featScale, features) labels = Input(labelDim) ] DNN=[ # conv1 kW1 = 5, kH1 = 5 cMap1 = 16 hStride1 = 1, vStride1 = 1 conv1_act = ConvReLULayer(featScaled,cMap1,25,kW1,kH1,hStride1,vStride1,10, 1) # pool1 pool1W = 2, pool1H = 2 pool1hStride = 2, pool1vStride = 2 pool1 = MaxPooling(conv1_act, pool1W, pool1H, pool1hStride, pool1vStride) # conv2 kW2 = 5, kH2 = 5 cMap2 = 32 hStride2 = 1, vStride2 = 1 conv2_act = ConvReLULayer(pool1,cMap2,400,kW2, kH2, hStride2, vStride2,10, 1) # pool2 pool2W = 2, pool2H = 2 pool2hStride = 2, pool2vStride = 2 pool2 = MaxPooling(conv2_act, pool2W, pool2H, pool2hStride, pool2vStride) h1Dim = 128 h1 = DNNSigmoidLayer(512, h1Dim, pool2, 1) ol = DNNLayer(h1Dim, labelDim, h1, 1) ce = CrossEntropyWithSoftmax(labels, ol) err = ErrorPrediction(labels, ol) # Special Nodes FeatureNodes = (features) LabelNodes = (labels) CriterionNodes = (ce) EvalNodes = (err) OutputNodes = (ol) ]  The network is defined in the block DNN. The network consists of two convolutional-maxpooling layers followed by an all-to-all standard network with one hidden later of 128 nodes. In convolutional layer one we have 5×5 convolutional kernels and we specify 16 of these (cMap1) for the parameter space. The operator ConvReLULayer is actually a shorthand for another subnetwork defined in a macro file. Algebraically we would like to represent the parameters of the convolution as a matrix W and a scale vector B so that if the input is X, the output of our network layer is of the form output = f(op(W, X) + B). In this case the operator op is convolution and f is the standard relu function relu(x)=max(x,0). The NDL code for the ConvReLULayer is given by ConvReLULayer(inp, outMap, inWCount, kW, kH, hStride, vStride, wScale, bValue) = [ convW = Parameter(outMap, inWCount, init="uniform", initValueScale=wScale) convB = Parameter(outMap, 1, init="fixedValue", value=bValue) conv = Convolution(convW, inp, kW, kH, outMap, hStride,vStride, zeroPadding=false) convPlusB = Plus(conv, convB); act = RectifiedLinear(convPlusB); ]  The W matrix and B vector are defined as Parameters and they will be the entities that are given an initial value and then modified during training to define the final model. In this case convW is a matrix with 16 rows of 25 columns B is a scale vector of length 16. Convolution is a built-in function that has been set to not use zero padding. This means that convolution over the 28×28 image will be centered on the 24 by 24 interior region and the result will be 16 variations of a 24×24 output sudo-image. We next apply Maxpooling based on 2×2 regions and the result is now 12×12 by 16. For the second convolutional layer we up the number of convolutional filters from 16 to 32. This time we have 16 channels of input so the size of the W matrix is 32 rows of 25×16 = 400 and the B vector for this layer is 32 long. The convolution is now over the interior of the 12×12 frames so it is size 8×8 and we have 32 copies. The second maxpooling step takes us to 32 frames of 4×4 or a result of size 32*16 = 512. The final layers have the 512 maxpooling output and a hidden layer of 128 nodes to a final 10 node output defined by the two operators DNNSigmoidLayer(inDim, outDim, x, parmScale) = [ W = Parameter(outDim, inDim, init="uniform", initValueScale=parmScale) b = Parameter(outDim, 1, init="uniform", initValueScale=parmScale) t = Times(W, x) z = Plus(t, b) y = Sigmoid(z) ] DNNLayer(inDim, outDim, x, parmScale) = [ W = Parameter(outDim, inDim, init="uniform", initValueScale=parmScale) b = Parameter(outDim, 1, init="uniform", initValueScale=parmScale) t = Times(W, x) z = Plus(t, b) ]  As you can see these are defined by the standard linear algebra operators as W*x+b. The final part of the graph definition is the cross entropy and error nodes followed by a binding of these to the special node names. We will define the training process soon, but first it is fun to compare this to the construction of a very similar network in TensorFlow. We described this in a previous post but here it is again. Notice that we have the same set of variables as we did with CNTK except they are called variables here and parameters in CNTK. The dimensions are also slightly different. The convolutional filters 5×5 in both cases but we have 16 copies in the first stage and 32 in the second in CNTK and 32 in stage one and 64 in stage two in the TensorFlow example. def weight_variable(shape, names): initial = tf.truncated_normal(shape, stddev=0.1) return tf.Variable(initial, name=names) def bias_variable(shape, names): initial = tf.constant(0.1, shape=shape) return tf.Variable(initial, name=names) x = tf.placeholder(tf.float32, [None, 784], name="x") sess = tf.InteractiveSession() W_conv1 = weight_variable([5, 5, 1, 32], "wconv") b_conv1 = bias_variable([32], "bconv") W_conv2 = weight_variable([5, 5, 32, 64], "wconv2") b_conv2 = bias_variable([64], "bconv2") W_fc1 = weight_variable([7 * 7 * 64, 1024], "wfc1") b_fc1 = bias_variable([1024], "bfcl") W_fc2 = weight_variable([1024, 10], "wfc2") b_fc2 = bias_variable([10], "bfc2")  The network construction is also almost identical. def conv2d(x, W): return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') def max_pool_2x2(x): return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') #first convolutional layer x_image = tf.reshape(x, [-1,28,28,1]) h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1) h_pool1 = max_pool_2x2(h_conv1) #second convolutional layer h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2) h_pool2 = max_pool_2x2(h_conv2) #final layer h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64]) h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1) h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob) y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)  The only differences are that the convolutional operators here are defined with padding so the output of the first convolutional operator had dimensions of 28 by 28 flowed by a pooling reduction to 14 by 14. The second convolutional operator and max pooling reduces this to 7×7, so the input to the final layer is 7x7x64 = 3136 with 1024 hidden nodes (with a relu instead of a sigmoid function). (For training purposes the last stage uses a probabilistic dropout function that randomly set values to zero. If keep_prob = 1, this is a no-op. ) ## Network Training The way network training is specified in CNTK differs substantially from the TensorFlow approach. The training and testing is specified in a file called convolution.config. Both CNTK and TensorFlow use a symbolic analysis of the flow graph to compute the gradient of the network for use in gradient decent training algorithms. The CNTK team has a very nice “book” that describes a great deal about how the gradients are computed. Currently CNTK only supports one learning method: Mini-batch Stochastic Gradient Decent, but they promise to add more in the future. He, Zhang, Ren and Sun have a lovely paper that describes how they train extremely deep (up to 1000 layers) networks using a nested residual reduction method reminiscent of algebraic multi-grid, so it will be interesting to see if that method makes its way into CNTK. An abbreviated version of the config file is shown below. command = train:test modelPath = "$ModelDir$/02_Convolution" ndlMacros = "$ConfigDir$/Macros.ndl" train = [ action = "train" NDLNetworkBuilder = [ networkDescription = "$ConfigDir$/02_Convolution.ndl" ] SGD = [ epochSize = 60000 minibatchSize = 32 learningRatesPerMB = 0.5 momentumPerMB = 0*10:0.7 maxEpochs = 15 ] reader = [ readerType = "UCIFastReader" file = "$DataDir\$/Train-28x28.txt"

features = [
dim = 784
start = 1
]

labels = [
# details deleted
]
]
]
test = [
….
]


The command line indicates the sequence to follow:  train then test.   Various file paths are resolved and then the train block specifies the network to be trained and the parameters for the Stochastic Gradient Decent (SGD).  A reader block specifies the way the “features” and “labels” from the network NDL file are read.   A test block is also included to define the parameters of the test.

Running this on a 16-core (non-GPU) linux VM took 62.95 real-time minutes to do the train and test and 999.01 minutes of user time and 4 minutes of system time.    The user time indicated that all 16 cores were all very busy (999/63 = 15.85).   Of course this means little as CNTK is designed for parallelism and massive GPU support is the true design point idea.

The training used by TensorFlow is specified much more explicitly in the Python control flow.   However, the algorithm is also a gradient based method called Adam introduced by Kingma and Ba.    Tensorflow has a number of gradient based optimizers in the library, but I did not try any of the others.

As can be seen below, the cross_entropy is defined in the standard way and fed to the optimizer to produce a “train_step” object.

y_ = tf.placeholder(tf.float32, [None, 10])
cross_entropy = -tf.reduce_sum(y_*tf.log(y_conv))
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
sess.run(tf.initialize_all_variables())
for i in range(20000):
batch = mnist.train.next_batch(50)
train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

print("test accuracy %g"%accuracy.eval(feed_dict={
x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))


Then for 20000 iterations the python program grabs a batch of 50 and runs the train_step with 50% random dropout.  The test step is to evaluate the accuracy subgraph on the entire test set.

Aside from the magic of the automatic differentiation and the construction of the Adam optimizer trainer, this is all very straightforward.   I also ran this on the same 16 core server with the same data as was used for the CNTK case.   Much to my surprise the real time was almost exactly the same as CNTK.   The real time was 62.02 minutes, user time 160.45 min, so much less parallelism was exploited.    I don’t believe these numbers mean much.   Both CNTK and Tensor flow are designed for large scale GPU execution and they are not running exactly the same training algorithm.

## Recurrent Neural Nets with CNTK and TensorFlow

Recurrent Neural Networks (RNNs) are widely used in language modeling such as predicting the next word you are going to type when texting or in automatics translation systems.   (see Andrej Karpathy’s blog for some great examples.) It is really a lovely idea. The input to the system is a word (or set of words) along with the state of the system based on words seen so far and the output is a predicted word list and a new state of the system as shown in Figure 1.

Figure 1.

There are, of course, many variations of the basic RNN.   One of the most popular is the Long-Short Term Memory (LSTM) version that is defined by the equations

Figure 2. LSTM Equations (taken from the CNTK book)

where    is the sigmoid function.

If you want to read a great blog article about LSTMs and how they work, I recommend this one by Christopher Olah.   In fact, he has a diagram that makes it a bit easier to see the flow of the equations above.  I had to modify it a tiny bit to fit the CNTK version of the equations and the result is shown in Figure 3.

Figure 3.  Adapted from Christopher Olah’s excellent article.

The notation in the picture uses sigmoid and tanh boxes and concatenated variables to represent this expression.

As can be seen, this is the form of the equations in figure 2 where the Ws and the bs are the learned weights.

## CNTK version

Below is the network definition language specification for the LSTM graph.   There are two things to notice here.    The first is the way the recurrence is handled directly in the network using a delay operator called “PastValue” that takes variable, its dimension and a time delay value and returns a buffered copy of that value.   The second thing to see is the way the W matrix is handled and how it differs from our concatenated operator describe above and in Figure 3.   Here they “stack” all the Ws that belong to x and all the Ws that belong to h and a stack of b values.   They then compute one W*x and one W*h and add them and then add b.   They then use a row slice operator to pull them apart to be used in the separate sigmoid functions.   Also note that they use the fact the Ws for c are all diagonal matrices.

LSTMPComponent(inputDim, outputDim, cellDim, inputx, cellDimX2, cellDimX3, cellDimX4) = [
wx = Parameter(cellDimX4, inputDim,  init="uniform", initValueScale=1);
b = Parameter(cellDimX4,  1,         init="fixedValue", value=0.0);
Wh = Parameter(cellDimX4, outputDim, init="uniform", initValueScale=1);

Wci = Parameter(cellDim, init="uniform", initValueScale=1);
Wcf = Parameter(cellDim, init="uniform", initValueScale=1);
Wco = Parameter(cellDim, init="uniform", initValueScale=1);

dh = PastValue(outputDim, output, timeStep=1);
dc = PastValue(cellDim, ct, timeStep=1);

wxx = Times(wx, inputx);
wxxpb = Plus(wxx, b);

whh = Times(wh, dh);

wxxpbpwhh = Plus(wxxpb,whh)

G1 = RowSlice(0, cellDim, wxxpbpwhh)
G2 = RowSlice(cellDim, cellDim, wxxpbpwhh)
G3 = RowSlice(cellDimX2, cellDim, wxxpbpwhh);
G4 = RowSlice(cellDimX3, cellDim, wxxpbpwhh);

Wcidc = DiagTimes(Wci, dc);
it = Sigmoid (Plus ( G1, Wcidc));

bit = ElementTimes(it, Tanh( G2 ));

Wcfdc = DiagTimes(Wcf, dc);
ft = Sigmoid( Plus (G3, Wcfdc));

bft = ElementTimes(ft, dc);

ct = Plus(bft, bit);

Wcoct = DiagTimes(Wco, ct);
ot = Sigmoid( Plus( G4, Wcoct));

mt = ElementTimes(ot, Tanh(ct));

Wmr = Parameter(outputDim, cellDim, init="uniform", initValueScale=1);
output = Times(Wmr, mt);
]


## The TensorFlow version

The TensorFlow version of the LSTM recurrent neural network is very different from the CNTK version.   While they both execute the same underling set of equations the way it is represented in TensorFlow make strong use of the Python control flow.    The conceptual model is simple.  We create a LSTM cell and define a “state” which is input to the cell and also an output.   In pseudo code:

cell = rnn_cell.BasicLSTMCell(lstm_size)
# Initial state of the LSTM memory.
state = tf.zeros([batch_size, lstm.state_size])

for current_batch_of_words in words_in_dataset:
# The value of state is updated after processing each batch of words.
output, state = cell(current_batch_of_words, state)


This is a nice pseudo code version of figure 1 taken from the tutorial.   The devil is in the very subtle details.   Remember that most of the time python code in Tensor flow is about building the flow graph, so we have to work a bit harder to build the graph with the cycle that we need to train and execute.

It turns out that the greatest challenge is defining how we can create and reuse the weight matrices and bias vectors inside a graph with a cycle.   CNTK uses the operator “PastValue” to create the needed cycle in the graph.  TensorFlow uses the literal recurrence above and a very clever variable save and recall mechanism to accomplish the same thing.   The moral equivalent of “PastValue” in Tensorflow is a function called tf.get_variable( “name”, size, initializer = None) whose behavior depends  upon a flag called “reuse” associated with the current variable scope.  If reuse==False and no variable already exists by that name in this scope then get_variable returns a new variable with that name and uses the initializer to initialize it.  Otherwise it returns an error.  If reuse == True then get_variable returns the previously existing variable by that name.  If no such variable exists, it returns an error.

To illustrate how this is used below is a simplified version of one of the functions in TensorFlow used to create the sigmoid function from eq. 1 above.  It is just a version of W*x+b where x is a list [a, b, c, …].

def linear(args, output_size, scope=None):
#Linear map: sum_i(args[i] * W[i]), where W[i] is a variable.
with vs.variable_scope(scope):
matrix = vs.get_variable("Matrix", [total_arg_size, output_size])
res = math_ops.matmul(array_ops.concat(1, args), matrix)
bias_term = vs.get_variable(
"Bias", [output_size],
initializer=init_ops.constant_initializer(1.))
return res + bias_term


Now to define the BasicLSTMCell we can write it roughly as follows.   (To see the complete versions of these functions look at rnn_cell.py in the TensorFlow Github repository.)

class BasicLSTMCell(RNNCell):
def __call__(self, inputs, state, scope=None):
with vs.variable_scope(scope):
c, h = array_ops.split(1, 2, state)
concat = linear([inputs, h], 4 * self._num_units)
i, j, f, o = array_ops.split(1, 4, concat)
new_c = c * sigmoid(f) + sigmoid(i) * tanh(j)
new_h = tanh(new_c) * sigmoid(o)
return new_h, array_ops.concat(1, [new_c, new_h])


As you can see, this is a fairly accurate rendition of the diagram in Figure 3.  You will notice the operator split above is the counterpart to the rowslice operation in the CNTK version.

We can now create instances of a recurrent neural network that can be used for training and using the same variable scope we can create another one to use for testing that share the same W and b variables.  The way this is done is shown in ptb_word_lm.py in the TensorFlow tutorials for recurrent neural nets.  There are two additional points worth observing.  (I should say they were critical for me to understand this example.)   They create a class lstmModel that can be used to build the networks for training and test.

class lstmModel:
def __init__(self, is_training, num_steps):
self._input_data = tf.placeholder(tf.int32, [batch_size, num_steps])
self._targets = tf.placeholder(tf.int32, [batch_size, num_steps])
cell = rnn_cell.BasicLSTMCell(size, forget_bias=0.0)
outputs = []
states = []
state = self._initial_state
with tf.variable_scope("RNN"):
for time_step in range(num_steps):
if time_step > 0:
tf.get_variable_scope().reuse_variables()
(cell_output, state) = cell(inputs[:, time_step, :], state)
outputs.append(cell_output)
states.append(state)
… many details omitted …


Where this is used is in the main program were we create a training instance and a test instance (actually there is a third instance which I am skipping to keep this as simple as possible).

with tf.variable_scope("model", reuse=None, initializer=initializer):
m = PTBModel(is_training=True, 20)
with tf.variable_scope("model", reuse=True, initializer=initializer):
mtest = PTBModel(is_training=False, 1)


What is happening here is that the instance m is created with 20  steps with no reuse initially.  As you can see from the initializer above that will cause the loop to unroll 20 copies of the cell in the graph and after the first iteration the reuse flag is set to True, so all instances will share the same W and b.   The training works on this unrolled version.   The second version mtest has reuse = True and it only has one instance of the cell in the graph.   But  the variable scope is the same as m, so it shares the same trained variables as m.

Once trained, we can invoke the network with a kernel like the following.

cost, state = sess.run([mtest.cost, mtest.final_state],
{mtest.input_data: x,
mtest.targets: y,
mtest.initial_state: state})


Where x and y are the inputs. This is far from the complete picture of the tutorial example. For example, I have not gone into the training at all and the full example uses a stacked LSTM cell and a dropout wrapper. My hope is that the detail I have focused on here will help the reader understand the basic structure of the code.

## Final Observations

I promised a programming model comparison of the two systems.    Here are some top level thoughts.

1. TensorFlow and CNTK are very similar for the simple convolutional neural network example.  However, I found the TensorFlow version easier to experiment with because it is driven by python. I was able to load it as a IPython notebook and try different things.  With CNTK one needed to completely understand how to express things with the configuration file.   I found that difficult.   With TensorFlow I was able to write a simple k-means clustering algorithm (see my previous post on Tensorflow).   I was unable to do this with CNTK and that may be due to my cluelessness rather than a limit of CNTK.  (If somebody knows how to do it, I would appreciate a tip.)
2. In the case of the LSTM recurrent neural network, I found the CNTK version to be completely transparent.   In the case of Tensorflow I found the top level idea very elegant, but I also found it very difficult to understand all the details because of the clever use of the variable scoping and variable sharing.   I had to dig very deep to understand how it worked.   And it is not clear that I have it all yet!   I did find one trivial bug in the Tensorflow version that was easy to fix and I am not convinced that the variable scoping and reuse flags, which are there to solve an encapsulation problem, are the best solutions.  But the good think about TensorFlow is that I can easily experiment with alternatives.
3. I must also say that the CNTK book and the TensorFlow tutorials are both excellent introductions to the high level concepts.  I am sure more detailed, deep-dive books will come out soon.

I am also convinced that as both systems mature they will improve and become easier to program.   I did not discuss performance, but CNTK is the current champ in terms of speed on some difficult challenges.  But with the rapid evolution of these systems I expect to see the competition to heat up.