Updated with Google’s TensorFlow: Artificial Intelligence, Neural Networks, and Deep Learning

Deep Learning (image courtesy of Google Semantics)

Deep Learning (Courtesy of Google Semantics)

**Note: there are links all over this post. Every link is unique, even if it has the same words as the link. So, click away for some awesome additional information! Some are amazingly informative!**

**Note #2: This is a work in progress. The amount of time this article took led me to simply provide links near the end rather than explaining what the links basically “said”. I’m going to fine-tune this over time. This could actually become a book; I’d love to do it!**

Artificial Intelligence (AI), Neural Networks and Deep Learning are fascinating topics to be sure. More interesting to me, however, answers the question “How did we get where we are today, and, where ‘are we’ today?

This post is about all 3 technologies, the pioneers that led to where we are today with these technologies, where we “are” today with these technologies; along with some correlations to one of my other favorite technologies, Augmented Reality.

The most inspiring and persistent AI revolutionaries is a gentleman by the name of Yann LeCun. Initially ridiculed for researching topics that mimic certain features of the brain with the thought that by doing so, he could discover ways that led to intelligent machines, you’ll see by reading this entire post he’s still a main thought leader in this discipline today. He ignored obstacles and stayed steadfast in his theories. You can read about his research here.

Yann LeCun, Google Scholar, computer scientist with contributions in machine learning, computer vision, mobile robotics and computational neuroscience

Yann LeCun, Google Scholar, Computer Scientist with Contributions in Machine Learning, Computer Vision, Mobile Robotics and Computational Neuroscience

Despite the ridicule, in the mid-90’s while at Bell Labs (then owned by AT&T) LeCun created software that simulated neurons and learned to read handwritten text by looking at many different examples.  AT&T used it to sell the first machines capable of reading the handwriting on checks and written forms. The bank machines could read thousands of checks per hour! How he accomplished this is documented by the IEEE here; a fascinating read!

Handwriting Recognition via Convolutional Neural Networks

Handwriting Recognition via Yann LeCun’s Convolutional Neural Networks


The Convolutional Networks LeCun Diagram

The Convolutional Networks LeCun Diagram

Quite unbelievably, notwithstanding this huge success – literally the same day this huge breakthrough was announced – AT&T split into 3 companies dedicated to different markets, and LeCun was directed to work on other things. Unfortunately his brain-inspired approach to AI fizzled out.

Almost 2 decades later, LeCun’s ideas have come full circle. The brain-inspired approach to AI, now called “Deep Learning“, is back and being used by top technology companies making great strides in facial and speech recognition today.

What Is Deep Learning

What Is Deep Learning?

There are other pioneers who led innovations of the deep learning technique.

One such pioneer was psychologist Frank Rosenblatt back in the 1950’s. Biologists then were developing mathematical theories of how intelligence and learning emerge from signals passing between neurons in the brain. The idea – still current today – was that the links between neurons are stronger if those cells communicated frequently. The bombardment of neural activity triggered by a new experience adjusts the brain’s connections so it can better understand the experience the second time around.

Frank Rosenblatt and Charles W Wightman

Frank Rosenblatt (right) and Charles W. Wightman (left)

In 1956, Rosenblatt used those theories to invent a way of making simple simulations of neurons in software and hardware, announced by the New York Times article ‘”Electronic ‘Brain’ Teaches Itself.” Rosenblatt’s design, which he called “Perceptron“, could learn how to sort simple images into categories such as triangles and squares. This was implemented on giant machines thickly tangled with wires, but they established the basic principles at work in artificial neural networks today.

A 1957 Diagram of Perceptron

A 1957 Diagram of Perceptron

The computer he built had 8 simulated neurons made from motors and dials connected to 400 light detectors. Each of the neurons received a share of the light signals, combined them, and depending on what they added up to, produced either a “1” or a “0”; together those digits was the perceptron’s description of what it saw. Initially this failed, but with using a method called supervised learning, he was able to train a perceptron to correctly distinguish different shapes. He’d show the perceptron an image along with the correct answer. Then the machine would tweak how much attention each neuron paid to the incoming signals, shifting “weights” toward settings that would produce the right answer. Eventually the perceptron could correctly categorize images it had never seen before.

Today’s deep learning networks use very sophisticated algorithms and have millions of simulated neurons, with billions of connections between them. However, they are trained the same way Rosenblatt trained his perceptron.

In 2004, the IEEE Frank Rosenblatt Award was established. Rosenblatt is regarded as one of the founders of neural networks. The award is presented to people or groups for outstanding contribution(s) to the advancement of design, techniques, or theory in biology and linguistically motivated computational paradigms, including but not limited to neural networks, connectionist systems, evolutionary computation, fuzzy systems, and hybrid intelligent systems in which these paradigms are contained.

The Rosenblatt Award

The Rosenblatt Award

Rosenblatt predicted that perceptrons would one day be capable of doing things such as greeting people by name; this idea became a linchpin of the nascent field of artificial intelligence. Research was done on perceptrons with more complex networks arranged in a hierarchy of multiple learning layers. The idea was, by passing images or other data successively through layers would allow a perceptron to solve more complex problems. Unfortunately, his learning algorithm couldn’t work on multiple layers.

Artificial Intelligence Visualization

Artificial Intelligence Visualization

In 1969, AI pioneer Marvin Minsky published a critique of perceptrons, claiming getting more layers working wouldn’t make perceptrons powerful enough to be useful. With that one publication, interest in neural networks was again halted.

AI specialist Marvin Minsky

Marvin Minsky, Cognitive Scientist in the Field of AI, Co-Founder of the MIT AI Lab and Author of Several Texts on AI and Philosophy

When AI researchers abandoned software that learned, they turned to using logic to make working facets of intelligence – such as an aptitude for winning at chess or Jeopardy, for example. The first video below shows IBM Watson beating undefeated Jeopardy champions, and the subsequent video shows IBM Deep Blue beating world chess champion Garry Kasparov:



(There are some amazing articles to read on IBM Watson today, both here and here, and on IBM Watson vs. IBM Deep Blue here, and more to come later.)

In the early 1980’s, LeCun was astounded that Rosenblatt’s perceptron theory was abandoned. This led him to a group of underground researchers working on Rosenblatt’s idea of working with neural networks with multiple layers. It’s central figure was Geoff Hinton (who now works at Google and the University of Toronto.) LeCun and Hinton became mutual admirers and LeCun joined the underground movement in 1985.

Google Scholar Geoff Hinton, cognitive psychologist and computer scientist, most noted for his work on artificial neural networks.

Google Scholar Geoff Hinton, Cognitive Psychologist and Computer Scientist, Most Noted for His Work on Artificial Neural Networks

Prior to LeCun’s success at Bell Labs, Hinton and others perfected a learning algorithm for neural networks with multiple layers known as backpropagation:



After LeCun’s check-reading project ended, backpropagation proved tricky to adapt to other problems, and a new way to train software to sort data was invented by a Bell Labs researcher which didn’t involve simulated neurons and thus was seen as mathematically more elegant. It quickly became a cornerstone of  Internet companies such as Google, Amazon, and LinkedIn that used it to train systems to block spam or suggest to you things to buy.

In 2003 LeCun, Hinton and a new collaborator University of Montreal professor Yoshua Bengio formed what LeCun calls “the deep-learning conspiracy“.

Yoshua Bengio, Google Scholar and most noted for his work on artificial neural networks. He is noted for his work in deep learning, along with Yann LeCun, Geoffrey Hinton, Andrew Ng et al.

Yoshua Bengio, Google Scholar, Most Noted for His Work on Artificial Neural Networks. He is Noted for His Work in Deep Learning, Along with Yann LeCun, Geoffrey Hinton, Andrew Ng et al.

In order to prove that neural networks are useful, they quietly developed ways to make them bigger, train them with larger data sets, and run them on more powerful computers. In contrast to LeCun’s 5 layers of neurons in his handwriting recognition system, they now could have 10 or many more.

Approximately in 2010, “deep learning” started to beat established techniques on real-world tasks like sorting images. Microsoft, Google, and IBM added it to their speech recognition systems.

Speech Recognition

Speech Recognition Visualization Image

Below is a short video entitled “Speech Recognition Breakthrough for the Spoken, Translated Word“:


Regardless, neural networks were still not considered widely useful. In 2012 LeCun’s rejection from a major conference on setting a new record on a standard vision task incensed LeCun. Read below to discover who was accepted into that competition and won.

Six months later, Hinton and 2 graduate students used a network similar to the one LeCun made for reading checks to rout the field in the leading contest for image recognition, the “ImageNet Large Scale Visual Recognition Challenge.” The challenge asks software to identify 1,000 types of objects as diverse as a mosquito to a cathedral. Hinton’s teams entry correctly identified the object in an image within 5 guesses about 85% of the time – more than 10 percentage points better than the second-best system. The deep learning software’s initial layers of neurons optimized themselves for finding simple things like edges and corners, with the layers after that looking for successively more complex features like shapes and eventually dogs or people. Click here to read the paper on the winning entry. Below you’ll find the award diagram:

Deep Learning with GPUs ImageNet Award Diagram

Deep Learning with GPUs ImageNet Award Diagram

And below here is an image that gives you a more visual description:

Learning of Object Parts

Learning of Object Parts

The slideshow below describes “Deep Learning” and “AI” using multiple GPUs, just like in Hinton’s ImageNet award-winning entry.  Click on the caption to watch the slideshow:

For the first time, LeCun saw the very people who ignored neural networks be absolutely amazed. They finally “got it.” Academics working on computer vision abandoned their old methods and deep learning suddenly became one of the main strands in AI.

Below is Hinton’s Google Tech Talk about Deep Learning, which won geeks over and virally spread the word that people should be watching neural networks, finally giving it the credibility it deserved:

Google bought a company founded by Hinton and the 2 students behind the 2012 result, and they started working on a research team known as “Google Brain“. Here’s a very in-depth article on what Google Brain is and what “brains” are behind that brain 🙂 http://bit.ly/GoogleBrainInDepth.

Google Brain Image

Google Brain Image

LeCun harbors mixed feelings about the 2012 research that brought the world around to his point of view. He should have been credited with the breakthrough system, and even Hinton (credited for the breakthrough) agrees that LeCun’s group had done more work than anyone else to prove out the techniques used to win the ImageNet challenge. The only reason he didn’t enter the challenge was a conflict with graduation schedules and other commitments.

Microsoft and other companies started to create new projects to investigate deep learning. In 2013, Facebook CEO Mark Zuckerberg appeared at the largest neural-network research conference and announced that LeCun was starting FAIR (Facial Artificial Intelligence Research), a new research program at Facebook studying Facial Recognition and Artificial Intelligence. Below is a video of LeCun talking about this research program:

Apart from Hinton working on Google Brain are many other “big brains”. One is his student from the 2012 competition, Google Fellow (one link) Ilya Sutskever (another link), shown below:

Google Scholar Ilya Sutskever, key member of Google Brain research team, co-founder of DNNresearch, postdoc in Stanford, student in the Machine Learning group of Toronto, working with Geoffrey Hinton

Google Fellow Ilya Sutskever, Key Member of Google Brain research team, Co-Founder of DNNresearch, Postdoc in Stanford, Student in the Machine Learning Group of Toronto, Working with Geoffrey Hinton

Another amazing person is Ray Kurzweil, the Panglossian (definition: “characterized by or given to extreme optimism, especially in the face of unrelieved hardship or adversity“) philosopher of AI.

Ray Kurzweil, Author, Entrepreneur, Futurist and Inventor

I actually love this guy’s innovative, futuristic (and IMHO probably true) philosophies. Watch this short TED Talk he did on how by 2030, we will be a hybrid of biological and non-biological thinking, and the non-biological portion is subject to his laws of accelerating returns, growing exponentially:

Adding to this list is Google Scholar (one link) Peter Norvig (another link), who wrote the standard textbook for AI courses and many others:

Google Scholar Peter Norvig

Google Scholar Peter Norvig, Computer Scientist, Director of Research (Formerly Director of Search Quality) at Google

Next is Sebastian Thurn (one link), Google [x] (another link) Founder and a key inventor of the self-driving car:

Google VP and Fellow Sebastian Thrun, educator, programmer, robotics developer and computer scientist, CEO and cofounder of Udacity

Google VP and Fellow Sebastian Thrun, Educator, Programmer, Robotics Developer and Computer Scientist, CEO and Co-Founder of Udacity

Then there’s the man they call “Google’s Baddest Engineer“, Jeff Dean, a Google Senior Fellow. As you’d have read in the long article I suggested a while back, he’s a legend among legends. He was the leader in creating Google’s software infrastructure! “Dean Fans” have collected “Jeff Dean Facts“, some of which are listed below:

  • Jeff Dean can beat you at connect four. In three moves.
  • One day Jeff Dean grabbed his Etch-a-Sketch instead of his laptop on his way out the door. On his way back home to get his real laptop, he programmed the Etch-a-Sketch to play Tetris.
  • Jeff Dean is still waiting for mathematicians to discover the joke he hid in the digits of Pi.
  • When Jeff gives a seminar at Stanford, it’s so crowded Don Knuth has to sit on the floor. (TRUE)
Don Knuth, Computer Scientist, Mathematician, Professor Emeritus at Stanford University, Author of the Multi-Volume Work "The Art of Computer Programming"

Don Knuth, Computer Scientist, Mathematician, Professor Emeritus at Stanford University, Author of the Multi-Volume Work “The Art of Computer Programming”

  • Jeff Dean once bit a spider, the spider got super powers and C readability
  • Jeff Dean has Perl Readability. (TRUE)
  • Jeff Dean got promoted to level 11 in a system where max level is 10. (TRUE)
  • Compilers don’t warn Jeff Dean. Jeff Dean warns compilers.
  • When Graham Bell invented the telephone, he saw a missed call from Jeff Dean. (my favorite!)

Here’s a plain photo of him:

Google Senior Fellow Jeff Dean, Google's Baddest Engineer

Google Senior Fellow Jeff Dean, “Google’s Baddest Engineer

And here he is in a panel of experts with Jeff Dean is wearing the “Google Brain TShirt” – Jeff Dean is the only man on Earth that is allowed to wear that TShirt!

Jeff Dean Wearing His Distinctive Google Brain TShirt

Jeff Dean Wearing His Distinctive Google Brain TShirt

Building the Google Brain, courtesy of

Building the Google Brain (Image Courtesy of “Backchannel”)

In 2011, Dean ran into Stanford AI professor Andrew Ng, who was visiting GooglePlex. Dean asked Ng what he was doing currently, and Ng answered “We’re trying to train neural nets.” Ng continued to state that after the deep learning breakthrough, they worked well, but if Google could figure out how to train really big nets, amazing things would happen.”

Andrew Ng, associate professor in the Department of Computer Science and the Department of Electrical Engineering by courtesy at Stanford University. He is chairman of the board of Coursera, an online education platform that he co-founded , Chief Scientist at Baidu Research in Silicon Valley

Andrew Ng, Associate Professor in the Department of Computer Science and the Department of Electrical Engineering by Courtesy at Stanford University. He is Chairman of the Board of Coursera, an Online Education Platform that He Co-Founded , Chief Scientist at Baidu Research in Silicon Valley

Eventually what happened is Dean and Ng made pursuing building a massive neural net system, informally known as “The Google Brain” and based within Google X. The team started experimenting with unsupervised learning because there’s more unsupervised data in the world than supervised. This resulted in a massive milestone in machine learning and AI.

This milestone was that we’re talking about is this “biggest neural network possible” (16,000 computer processors with 1 billion connections) discovered the concept of a cat by itself. It was exposed to 10 million randomly selected YouTube videos over the course of 3 days and, after being presented with a list of 20,000 different items, it began to recognize pictures of cats using a deep learning algorithm – despite being fed no information on distinguishing features that might help identify one!!! Quoting Jeff Dean, “It basically invented the concept of a cat.

Picking up on the most commonly occurring images featured on YouTube, the system achieved 81.7% accuracy in detecting human faces, 76.7% accuracy when identifying human body parts and 74.8% accuracy when identifying cats. These results reveal that it’s possible to train a face detector without having to label images as containing a face or not.

Deep Learning Computational Power

Deep Learning Computational Power Image

These findings are useful in the development of speech and image recognition software, including translation services. It’s the process of learning through repetition. So instead of having teams of researchers trying to find out how to find edges, you instead throw a ton of data at the algorithm and you let the data speak and have software learn from the data.

Google considers it such an advance that the research has made the leap from the X lab into it’s main labs.

This remarkable discovery is explained in the video below:


Rather than learning from tagged or labeled data in speech recognition for example, by soaking up massive amounts of transcripted speech data using tens of thousands of hours of audio together with the transcript of what was actually said to train neural networks to make predictions, it’s now massively less expensive to use unsupervised learning (learning from unlabeled data.) This is what’s driving most of the economic value of deep learning today.

Ng has since left Google to oversee a new artificial-intelligence research lab in Silicon Valley for “China’s Google“, Baidu.

Four of Google’s deep learning scientists published a paper entitled “Show and Tell.” Here was another scientific breakthrough. But not only that, it introduced a “neural image caption generator” (NIC) designed to provide captions for images without any human intervention! (I just love this stuff!) This enormous experiment involved vision and language. What made this system unusual is that it layered a learning system for visual images onto a neural net capable of generating sentences in natural language (which by the way, LeCun predicted in 2013.)

Show, Attend, and Tell: Neural Image Caption Generation with Visual Attention

Show, Attend, and Tell: Neural Image Caption Generation with Visual Attention (Image Courtesy of BackChannel)

The scientists said to find “a group of young people playing a game of Frisbee“, “a person riding a motorcycle on a dirt road“, and “a herd of elephants walking across a dry grass field“. The system “learned” on its own the concepts of Frisbee, road, and herd of elephants; very impressive!

Games, Science, and Turing's Machine, Image Courtesy of BackChannel

Games, Science, and Turing’s Machine (Image Courtesy of BackChannel)

Given that Hinton’s hiring and Dean’s brain were major steps in pushing the company towards deep learning, perhaps the biggest move occurred in 2013, when Google spent $400 million to acquire DeepMind, an AI company. DeepMind has its own take on deep learning, based on a closer study of the brain itself.

Image Courtesy of Center for Data Science to Advance Deep Learning Research

Image Courtesy of Center for Data Science to Advance Deep Learning Research

DeepMind’s CEO is another one of deep learning’s leaders, Demis Hassabis. At 14, Hassabis was an avid computer game programmer as well as a chess prodigy. He had key roles in landmark game titles such as “Black and White” and “Theme Park“. Then he started his own game company, eventually employing 60 people, while still in his twenties. But gaming, he says, was a means to an end, the end being the development of an intelligent general purpose artificial intelligence machine. By 2004, he felt that he had taken gaming AI as far as he could in that field. But it was too soon to start an AI company — the computer power he needed wasn’t cheap and plentiful enough. So he studied for a doctorate in cognitive neuroscience at the University College London.

Google Scholar Demis Hassabis

Google Scholar Demis Hassabis, AI Researcher, Neuroscientist, Computer Game Designer, World-Class Gamer, Chess Prodigy

DeepMind’s neural net system was left to its own deep learning devices to learn game rules, trying at millions of sessions at Pong, Space Invaders, Beam Rider and other classics, and taught itself to do equal or surpass an accomplished adolescent. Even more intriguing, some of its more successful strategies were ones that no humans had ever envisioned! To quote Hassabis, “This is a particular potential of this type of technology. We’re imbuing it with the ability to learn for itself from experience just like a human would do and therefore it can master things that maybe we don’t know how to program. It’s exciting to see that when it comes up with a new strategy in an Atari game that the programmers didn’t know about.

Hassabis has some ideas how DeepMind technology might enhance people’s lives. A more proactive version of search – not only finding things for people but making decisions for them – would be a valuable provider to save time. For example, there’s more books in the world than anyone could read in their lifetime. If you’re on a flight or relaxing,  you shouldn’t have to think about what book to read; it should be automated. He also envisions using DeepMind in Google’s self-driving car or Google’s new company Calico, focused on health and well-being.

Below is a new video of Hassabis describing Google’s DeepMind – this is an amazing video I suggest everyone watch:


Hassabi’s big goal is to create a general artificial intelligence machine that will process information anywhere it can get it, then do practically anything with it. Again to quote Hassabis, “The general AI that we work on here is a process that automatically converts unstructured information into useful, actionable knowledge. The human brain does this, so we have a prototype already.” (This intrigues me so much!)

To many, this sounds scary, even to Hassabis. He acknowledges that the advanced techniques his own group is pioneering may lead to a problem where AI gets out of human control, or at least becomes so powerful that its uses might be best constrained (DeepMind investor Elon Musk – of SpaceX, PayPal, Tesla Motors, Hyperloop and SolarCity –   just invested $10 million to study AI dangers). That’s why, as a condition of the DeepMind purchase, Hassabis and his co-founders demanded Google set up an outside board of advisers to monitor the progress of the company’s AI efforts.

ex machina movie cover

With the “uncontrollable fear-factor” part removed from AI mentioned above, what was just described (and also will be described more below), is illustrated in a sensational movie called “Ex Machina.” In the movie (not to ruin it – I encourage you to watch it! I can’t count how many time I’ve done so!) an uber-genius, eluding to but not mentioning Google, created a ground-breaking experiment in AI by creating a breath-taking female AI. A young programmer is brought to a super secret place where the creator of this AI challenges him to evaluate the human qualities of his design; basically passing the “Turing test“. The Turing test is based on Alan Turing‘s book “The Turing Machine” that legendary author Charles Petzold transcripted into a “more” readable version in “The Annotated Turing” book. The Turing test is a test of a machine’s ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human. Turing proposed that a human evaluator would judge natural language conversations between a human and a machine that is designed to generate human-like responses. Below is the movie’s trailer:


DeepMind recently published another major paper, describing a project that synthesizes some ideas from neuroscience memory techniques to create a neural network with the properties of a Turing Machine, which is synonymous for a universal computing device. This means that this system, given enough time and memory, can in theory compute anything. The paper focused on the practical: with the ability to “record” information and draw on it later – a type of artificial version of a person’s working memory – the “Neural Turing Machine” was not only able to learn faster and to perform more complex tasks that previous neural nets, but “to generalize well outside its training regime“.

Since Jeff Dean’s deep learning project has moved from Google X to the Knowledge division (which includes search), his team has been working closely with many search-related teams, including language, image recognition, ads, street view and self-driving cars.

Google's Lexus RX 450h

One of Google’s Self-Driving Car, a Lexus RX 450H

Dean’s neural model for end-to-end language translation is superior in picking up nuances in diction that are key to conveying meaning. And their speech recognition technology has only an 8% word error rate.

Google's Speech Recognition

Google’s Speech Recognition Diagram of Deep Neural Networks

Back to the topic of IBM Watson, it now has upgraded speech and vision analysis and also it’s computing system is helping doctors at the University of Texas’s MD Anderson Cancer Center to spot patterns in the medical charts and histories of more than 100,000 patients. Watch the video below:


Microsoft has launched its InnerEye computing program, which is aimed at analyzing medical images and identifying disease progression. There’s a lot of good being done with the advancement of AI.

LeCun’s group at Facebook has created software that matches faces almost as well as you do. For example, they’re envisioning AI that keeps you from uploading embarrassing pictures of yourself! And they’ve also created a virtual assistant named Moneypenny. Click here to see how it differs from Siri and Cortana.

Facebook Moneypenny virtual assistant (Image courtesy of eWeek)

Facebook’s “Moneypenny” Virtual Assistant (Image Courtesy of eWeek)

Now to tie this all in with augmented reality, Microsoft’s Cortana combined with Hololens appears in your Windows car as a virtual assistant, projected right onto your windshield!

Cortana and Hololens in Your Windows Car (Image courtesy of GeekSnack)

Cortana and Hololens in Your Windows Car (Image courtesy of GeekSnack)

Compare that to InfiniyAR‘s Virtual Assistant that you can touch:

Infinity AR's Virual Assisstant

Infinity AR’s Virual Assisstant

And just for those of you who are not familiar with augmented reality, I’ll give you a “teeny, teeny” taste of what it is, sticking with the company InfinityAR for consistency (although its’ title contains the words “concept video“, these technologies literally exist today and are used mainly in the enterprise):


To finish off this article, I’ll leave your with some extra links to videos by the pioneers mentioned in this article on AI, Neural Networks and Deep Learning for you to view at your leisure:

Deep Learning – World Changing, Disruptive, Artificial Intelligence“:


Large-Scale Deep Learning for Building Intelligent Computer Systems“:


IBM Watson: Intelligence as a Service“:


Dark Knowledge: TTIC Distinguished Lecture Series – Geoffrey Hinton“:


Google's Always Up To Something New!

Google’s Always Up To Something New!

Google’s Admirable Announcement of Their Open-Source Machine Learning Library, “TensorFlow(added 11/10/2015)

On November 9, 2015, Google made one of their most astonishing announcements ever: giving the world free access to their latest machine learning algorithm called “TensorFlow“:

You might be wondering, “What’s the big deal about ‘this’ machine learning vs. all of the other machine learning algorithms available by many other companies?

Other companies charge to use their machine learning tools and the code behind it is hidden; therefore, it can’t be expanded upon by others unless you’re on the team in that company’s machine learning division. Google, on the other hand, is being true to science: they benevolently made their advanced machine learning FREE for everyone. Their noble philosophy behind doing this goes back to the root of scientific discovery and the invention of the internet itself: research communication and collaboration. That is very impressive to me, indeed.

To paraphrase Greg Corrado, Google Scholar and Sr. Research Scientist, “It doesn’t make sense for researchers in machine learning to have different tools than the people developing the products. They all should have access to the same set of tools, so if ideas the researchers have done could be moved directly into products without having to rewrite code, it brings better products to market faster.

Greg Corrado, Google Research Scientist Interested in Biological Neuroscience, AL, and Scalable Machine Learning, One of the Founding Members of the Co-Technical Lead of Google's Large Scale Deep Neural Networks Project

Greg Corrado, Google Scholar and Research Scientist Interested in Biological Neuroscience, AL, and Scalable Machine Learning, One of the Founding Members of the Co-Technical Lead of Google’s Large Scale Deep Neural Networks Project

Imagine that? All researchers, developers, scientists using the same code base, sharing their new code and ideas back and forth, working together. Utopia? I’d call it “Googliness.”

Here is a YouTube video that was embedded in the original announcement on Google’s blog page yesterday:

TensorFlow is a successor to “DistBelief“, the machine intelligence engine that Google has been using since 2011. TensorFlow has significant improvements over its predecessor. According to Google Fellow Jeff Dean (from earlier in the article) and Rajat  (one link) Monga (another link), Google Technical Lead of TensorFlow, “TensorFlow is general, flexible, portable, easy-to-use, and completely open source. We’ve added all this while improving upon DistBelief’s speed, scalability and production-readiness – in fact, on some benchmarks, TensorFlow is twice (to five times) as fast as DistBelief.”

Ragat Monga, Google Scholar, Technical Lead of TensorFlow

Ragat Monga, Google Scholar, Technical Lead of TensorFlow

TensorFlow is loosely built on how the brain works, using “large scale unsupervised learning“. It can run on a single smartphone or across thousands of computers in data centers. Jeff Dean and his researchers have used TensorFlow with over 50 different teams at Google, deploying this system in real products across a wide spectrum of areas. It’s used in Google’s speech recognition, their new photo app “Google Photos“, and in “Smart Reply” for Gmail’s inbox. It excels at perceptual and language understanding tasks, so the computer can actually see what’s in an image or a short video clip when you’re looking at it.

You can download the library on the official site.  To quote the front page of the site, “The library is for computation using data flow graphs. Nodes in the graph represent mathematical operations, while the graph edges represent the multidimensional data arrays (tensors) communicated between them. The flexible architecture allows you to deploy computation to one or more CPUs or GPUs in a desktop, server, or mobile device with a single API. TensorFlow was originally developed by researchers and engineers working on the Google Brain Team within Google’s Machine Intelligence research organization for the purposes of conducting machine learning and deep neural networks research, but the system is general enough to be applicable in a wide variety of other domains as well.

Here is first an image of TenorFlow’s flow, followed by a very short video showing the flow:

Diagram of TensorFlow's "Flow"

Diagram of TensorFlow’s “Flow”



Again from TensorFlow’s home page, “Data flow graphs describe mathematical computation with a directed graph of nodes & edges. Nodes typically implement mathematical operations, but can also represent endpoints to feed in data, push out results, or read/write persistent variables. Edges describe the input/output relationships between nodes. These data edges carry dynamically-sized multidimensional data arrays, or tensors. The flow of tensors through the graph is where TensorFlow gets its name. Nodes are assigned to computational devices and execute asynchronously and in parallel once all the tensors on their incoming edges becomes available.

There’s great documentation on TensorFlow’s site. They have documentation for both Python and C++:

TensorFlow's Python Documentation Menu

TensorFlow’s Python Documentation Menu


TensorFlow's C++ Documentation

TensorFlow’s C++ Documentation

It has TensorFlow Mechanics information:

TensorFlow Mechanics Menu

TensorFlow Mechanics Menu

And also some machine learning examples:

TensorFlow Machine Learning Examples

TensorFlow Machine Learning Examples


You can download the whitepaper here, read about the technology on Stack Overflow here, and even fork your own repository from Google’s original source code on GitHub here!

Google keeps feeding our brains! Thank you, Google!

TensorFlow "Feeding Our Brains" (courtesy of Wired Magazine)

TensorFlow “Feeding Our Brains” (Courtesy of Wired Magazine)

(Continuing with the original article) And to leave you with some final thoughts:
Exponential advances will allow deep learning to handle significant intellectual activities within 5 years. Health, the law and education are just a few of the vast array of human enterprises that will benefit.

Some recent applications of deep learning include:

– The “Google Brain” identifying a cat simply by watched 10 million YouTube videos
– Microsoft Research’s “Project Adam” integration into Cortana for vastly improved image interpretation
– Facebook’s 97.25% accurate face matching deep learning software
– Google’s “neural image caption generator” (NIC) designed to produce captions for images without human input
– Netflix’s application of deep learning to improve video recommendations
– DeepMind’s Neural Turing Machine learns to play video games by playing each game many times.
– Skype’s real time translator
– “Enlitic“‘s use of deep learning to process medical images, test results and patient histories and deliver more accurate and comprehensive diagnostic assessments

Some 80% of employment opportunities in developed economies will be affected by software systems based on deep learning principals leading to social disruption if governments don’t plan the re-shaping of economies as change occurs.

Eventually these developments will lead to significant productivity benefits delivering reduced working hours and an improved quality of life. Find out why deep learning leader like Jeff Dean Google senior fellow, John Platt of Microsoft Research, Google Scholar Adam Berenzweig from Clarifai, Elliot Turner Founder of AlchemyAPI and Ilya Sutskever working on Google Brain are impassioned about this breakthrough field.

See other In5Years” Future Technology videos:

Qualcom Tricoder Prize“:


5 Amazing Future Defense Technologies“:


Subscribe to “In5Years” at  http://goo.gl/4EK7QQ

Like “In5Years” on Facebook: https://www.facebook.com/infiveyears
Follow “In5Years” on Google+: http://google.com/+In5yearsInfo
Visit the Website: http://in5years.info


  1. WOW!

  2. Did you post my article? A Japanese online technical media company translated this & published it at my permission. I’d love to see it!

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: