Import AI

Free data!: Either someone has made off with the keys to Google’s data vault, or the organization has made a strategic decision to give away a teeny-tiny percentage of its competitive advantage. Now, along with sharing code and programing tools (TensorFlow), the company has begun to share large amounts of data as well. Case in point: this week it announced the Open Images dataset, which consists of 9 million URLs to images that have been annotated with labels spanning over 6000 categories. However, these labels were applied by Google’s own AI systems so there’s a chance they could not reflect reality as nicely as human-annotated labels. It also announced YouTube-8M, a mammoth video dataset. Salesforce MetaMind released WikiText, a large text dataset to help people produce better language models.

If data is the new coal, who can use it? Coal is useful. Neil Lawrence (University of Sheffield and now a member of Amazon’s ML team), has said data is as important to machine learning as coal is to the generation of power. So releases of data by organizations and companies is a great thing. But it’s worth bearing in mind that turning coal into energy is expensive and requires a significant amount of infrastructure. “£20k buys 3 machines for AI research. Typical experiments by companies use hundreds or thousands of machines. See the gap that worries me?,” writes Nando de Freitas of Google DeepMind. “Hardware is the least of our worries. Having high quality open datasets in areas that matter is vastly more important issue,” counters Google’s Francois Chollet. Either way, it’s worth bearing in mind that developing modern AI is tremendously expensive and requires vast hardware investments, so merely having access to the data isn’t enough, you need to be able to marshal the resources and technology to deploy on large pools of infrastructure.

The machine that builds the machine that builds the machine: Skip to the last half of this (excellent) lecture by Nando de Freitas to get a good overview of the latest AI research, which seeks to develop computers that can learn to design certain aspects of AI systems, which then solve tasks. Currently, good AI researchers have strong intuitions which they use to develop certain arrangements of neural network software. The next generation fo software will replicate some of this intuition and design aspects of the machinery itself.

Better cloud computers: After years of developers grumbling about its ageing fleet of GPUs, someone at Amazon has taken the wraps off of its new AI infrastructure. You can now rent computers that lash together up to 8 NVIDIA Tesla K80 accelerators from Amazon Web Services, and you can pair that with a nice software bundle developed by Amazon for running AI applications. Though companies are dabbling in other chips, ranging from FPGAs, to ASICs, to novel coprocessors, GPUs look like they’ll remain the standard workhorse for AI for years to come.In related news, Nvidia’s share price has more than doubled in the last year.

Relatively good robots: Deep learning techniques are washing into the field of robotics, speeding up progress there. That was most visible in this year’s Amazon Picking Challenge where entrants used robotic arms to pick up objects in a (simplified) warehouse. The results were significantly better than the year before and many of the teams had adopted deep learning techniques to make their robots better at seeing and acting in the world. Now, an MIT team which ranked highly in the competitions has published a research paper, Multi-view Self-supervised Deep Learning for 6D Pose Estimation in the Amazon Picking Challenge (PDF), outlining the system it used in the competition. One thing worth noting is that datasets like ImageNet don’t provide the right sort of data needed to train these industrial robots, so one of the inventions the team came up with was a method to train its robots to create their own large datasets of the objects they were trying to pick up.

The incredible, surprising, inevitable, metastasizing nature of artificial intelligence: Microsoft has gone through one of its periodic restructurings to create an AI research group consisting of more than 5,000 computer scientists and engineers. Why? Because the flexibility and utility of AI software has reached the point where researchers can come up with new techniques and engineers can (relatively) easily port these over to work on specific tasks within specific business divisions. Therefore, it makes sense to throw more resources into AI organizations because the inventions usually allow you to extract some kind of short-term information arbitrage advantage over your competitors, or let you reduce the cost of carrying out some part of your business. The same attitude underlies the ‘Brain’ group at Google, and Facebook’s Applied Machine Learning Group. Amazon is making similar moves to expand and build-up its AI group, as is IBM. “AI is shifting the computer science research supply chain and blurring lines between research and product,” writes MSR’s AI czar Harry Shum. “End-to-end innovation in AI will not come from isolated research labs alone, but from the combination of at-scale production workloads together with deep technology advancements in algorithms, systems and experiences.”

Please, build this for me: (And by me, I mean Google). “Better debugging tools will help researchers understand why their models aren’t learning, better experimentation management will make it easier for them to run and analyze more experiments,” writes Rajat Monga, the engineering director for TensorFlow, in a Quora session.

First the words, then the worlds: The recent surge of interest in artificial intelligence prompted many companies to build and release free AI programming frameworks; these include Google’s TensorFlow, Microsoft’s CNTK, Amazon’s DSSTNE, Baidu’s PADDLE, Skymind’s DL4J, Nervana’s (now, Intel’s) NEON, and others. Now we’re onto the next thing: environment interfaces. Just as an AI framework gives programmers a reasonably high-level language to use to get computers to perform the sorts of complex commands suited to AI development, environment interfaces will make it relatively simple for programmers to hook an AI system up to an environment(s) for it to learn and grow in. OpenAI released OpenAI Gym earlier this year to do just that. Now others are investing in the space, creating more free tools for the global AI community. Facebook this week announced CommAI-env (Environment fgor Communication-based AI), to make it easy for programmers hook agents up to a text-based communication layer, which gives them a way to talk to a text-based environment where they need to solve a variety of challenging tasks.

(These tasks range from being silent for a period of time determined by the CommaAI software-based teacher, to reading a list and manipulating it, to navigating through a virtual world described by text, to many others. “While the tasks might appear almost trivial (but try solving them in the scrambled mode we support, where your knowledge of English won’t be of help!), we believe that most of them are beyond the grasp of current learning-based algorithms, and that a Learner able to solve them all would have already made great strides towards the level of communicative intelligence required to interact with, and learn further from human teachers,“ the Facebook researchers write.)

After environment interfaces it’s likely they’ll be an arms race to release the environments themselves. Deepmind will probably release its 3D in-development ‘Labyrinth’ world. Microsoft and Facebook have already released their Minecraft-based Malmo, and maze-based Mazebase environments.

Free lectures! Next time you’re hankering for something to listen to while performing a tedious but necessary activity, like cleaning your living room while a neural net is trained, you might want to listen to the roughly 20 hours of lectures from September’s Deep Learning School. Replays available here.

Second edition of the bible (of reinforcement learning) from Richard Sutton and Andrew Barto is out now. Read it online here (PDF).

Smart summarization: being able to read something and remember only the salient components is a hallmark of intelligence. So it’s encouraging to see new AI papers report better results at compressing language against the dataset used in the Hutter Prize. Two recent papers seek to do this in two different ways. The ‘Multiplicative LSTM for sequence modelling’ paper from researchers at the University of Edinburgh and the Toyota Technological Institute pairs a recurrent neural network with an LSTM to achieve good results (pg 11, PDF), while the ‘hierarchical multiscale recurrent neural networks’ paper from the University of Montreal builds RNNs that develop a hierarchy of different representations of the underlying data (pg 8, PDF).

Special chips for AI: Microsoft has given more details on its ‘Catapult’ FPGA-based AI hardware, which will let developers load software onto its Windows Azure cloud and speed it up by running it on specially designed chips. The company first announced the Catapult system years ago and said in August of this year that it would make it available as a service other developers could rent. Meanwhile, a new company called Cornami has decloaked, with plans to build FPGA-liker chips with thousands of simple cores, which sound like they’ll be a good fit for common AI tasks (hello, matrix multiplication.)

Measure ten thousand ways, run once: As more chips are developed for AI systems and developers invest more hours in writing complicated code to parallelize and chain-together these various FrankenClusters, it’s going to be important to measure the performance of standard deep learning procedures on different hardware substrates. A new free tool from Baidu, called DeepBench, may help.

Monkeys for better robots: a reasonable article about robotics startup Kindred, which a bunch of ex-D-Wave (the quantum computer company) are involved in. Kindred appears to be betting that a person (or monkey!) can be an effective remote operator of a robot and could provide the raw data to train a computer to move a robot all by itself. Tantalising.

Better robots, no monkeys required: But there may be easier ways to train robots. DeepMind has published a paper on progressive nets which can help with transfer learning, and now a team from the University of California at Berkeley have come up with their own way of training multiple neural networks to work across multiple robots. This will make it easier to get robots to learn from each other in the same way Tesla’s vehicles are able to use ‘fleet learning’ to make sure cars don’t make the same mistake twice. It could also help robots tackle never-before-seen tasks by fusing together previously seen ideas. “In some cases, previously untrained combinations might generalize immediately to the new task, while in other cases, the composition of previously trained modules for a new previously unseen task can serve as a very good initialization for speeding up learning,” they write (PDF). Check out the video and paper here.

Freaky, nightmarish, convnet faces: Nice overview of how neural networks can be used to generate objects. Come for the accessible descriptions and stay for the hall-of-mirrors face transformations.

Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.

Deep learning goes to Africa, helps some kids: Last week I told you about Bay Labs and some collaborators taking technology to Africa to help identify symptoms of  Rheumatic Heart Disease (RHD) in Kenyan school children. The Bay Labs software uses deep learning to analyze data derived from an ultrasound to take a good educated guess as to whether it’s seeing something consistent with RHD. During the trip, medical professionals scanned 1200 children in four days and were able to spot 48 children with RHD or congenital heart disease. During this, they had a chance to test out the Bay Labs tech and see if it worked. “The feedback from our tests was overwhelmingly positive, particularly coming from Kenyans who never used an ultrasound scanning device before. John for instance, a clinical officer working for the Eldoret hospital, was able to acquire the right view after a few minutes of using the prototype, and to see the recommendations of the Bay Labs prototype (it was a non-pathological case here). I spent some time interviewing him after the fact and it was hard to contain his enthusiasm. He performed what usually takes a sonographer few years of training in a few minutes!,” Bay Labs’ Johan Mathe tells me. Check out some pictures from the trip here. If you or anyone you know is trying to deploy deep learning into (affordable) healthcare systems to help people, then please let me know.

Intelligent Utilities: Sci-Fi author Stephen Baxter has a pet theory that one of the test-beds for really sophisticated AI systems will be planet-spanning utility systems. The idea is that if you’re tasked with managing power for a sufficiently large system then you’ll need some degree of intelligence to match the inputs with the outputs, distribute load effectively, and even manipulate some of your edge hardware (fields of solar panels, dams, etc) to modify inputs. So it’s interesting to see this postdoc position at Oxford which is seeking a researcher to apply machine learning methods to the noisy, local measurements generated by large energy storage systems.

The (synthetic) players of games: Starcraft, a real-time strategy game released in 1998 that is played and watched by tens of thousands of people a month in South Korea could well be the next ‘grand challenge’ companies are likely to test their artificial intelligence systems on. The game pits players against one another in a battle containing numerous units that spans land and air, full of subterfuge, fast-paced play, imperfect information,and all dependent on an underlying resource extraction economy which each player must carefully build, tend, and defend. Google Deepmind has dropped numerous hints that Starcaft is a game it’s paying attention to, and last week Facebook AI Research published a paper where it used neural networks to learn some troop movement policies within a Starcraft game.

The self-modifying, endlessly mutating, e-commerce website: a new product from AI startup Sentient makes it possible for a website to ‘evolve’ over time to achieve higher sales. Sentient Ascend will convert a web page into numerous discrete components, then shuffle through various arrangements of them, and breed and evolve its way to a page that is deemed successful, eg one which generates more purchases. This relies on the company’s technology which pairs a specialism in evolutionary computation with a massive, million-CPU-plus computer farm spread out across the world. No surprise that University of Texas evolutionary algorithm professor Risto Miikkulainen has been working there since mid-2015.

Dealing with the deep learning research paper deluge: Because deep learning is currently suffused with money and interest and postgraduate students there’s been a corresponding rise in the number of research papers being published. Andrej Karpathy’s Arxiv Sanity has been a handy tool for navigating this. Now Stephen Merity has released another tool, called Trending Arxiv that makes it easier to spot papers that are being widely talked about.

Studying deep learning: The Deep Learning textbook, a general primer on deep learning, is now available to purchase in hardcover, if you’re into that sort of thing. Try before you buy by reading the online version for free. Another great online (free) resource is the ’neural networks and deep learning’ book from Michael Nielsen.

Imagination, generative models, and platonic forms: One of the truly weird things about young children is you can show them a stylized picture of something, like a wedge of cheese wearing a dinner jacket, tell them something about it (for instance: this cheese is named Frank and works in insurance), then show them a real version of the object and they’ll figure out what it is. (In this case, the child will examine the lump of cheddar replete with miniature knitted jacket and exclaim ‘that’s Frank, he works in insurance!). Why is this? Well, the child has developed an idea in their head of what the object is and can then generalize to other versions of it. You may know this from philosophy, where Plato is famous for talking about the ‘platonic forms’, which is a notion that we carry around ideas in our head of The Perfect Dog or The Perfect Steak, and then use these rich, perfect representations to help us categorize the imperfect steaks and dogs we find in the world. Clearly, it’d be helpful to develop software that can observe the world and develop similarly rich, internal representations of it. This would make it easier to build, for example, robots that possess a general idea of what a door handle is and therefore be able to manipulate never-before-seen handles. Generative adversarial networks (GANs) are one promising route to coding this kind of rich representation into computers. So keep that in mind when looking at this work from UC Berkeley and Adobe that lets you generate new shoes and landscapes from simple visual tweaks, or this GAN which is able to generate videos, or this new paper from the Twitter / Magic Pony team that uses GANs to scale-up low-resolution images. And there’s new research from NYU / FAIR that may make it easier to train the (notoriously unstable) GANs.

Neural nets grow up: As recently as a year ago companies would view neural networks and other currently in-vogue AI techniques as being little more than research projects. Now they’re actively trying to hire people with expertise in these areas for production projects around categorization and reinforcement learning. And the interest doesn’t show any signs of dimming, says Jake Klamka, CEO of Insight Data Science. To get an idea of just how many places people are finding neural nets useful for, take a look at this (heroic) round-up of recent research papers by The Next Platform… Weather forecasting! Detection of deep-sea animals! Fault diagnosis in satellites! And much, much more.

What can’t AI do? Lots! The best way to describe current AI is probably the Churchillian phrase ‘the end of the beginning’. We’ve deployed smart software into the world that is capable of doing a few useful things, like saving on power consumption of data centers, performing basic classification of perceptual inputs, and helping to infer some optimal arrangements of various things. But our AI systems can’t really act independently of us in interesting ways, and are frustratingly obtuse in many others. There’s a lot to work on, as replies to this tweet show.

Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.

Deep learning + heart doctors in Africa: Good healthcare is punishingly expensive. It relies on vast infrastructure and, in most countries, huge amounts of government support. If you’re unlucky enough to be born in a part of the world with poor healthcare infrastructure then your life will be shorter and your opportunities will be smaller. So it’s great to see examples of AIhelping to reduce the cost of healthcare. This week, deep learning startup BayLabs is doing work with the American society of echocardiography to help a Kenyan team scan hundreds of school children in the village of Eldoret for signs of Rheumatic Heart Disease – the most common acquired heart disease in children, particularly those in developing countries. The company is using a prototype device that looks like a ultrasound machine, though miniaturized. It’s got a GPU in it, naturally. The device uses artificial intelligence to spot RHD symptoms. It does this locally so it doesn’t need to phone home to a cloud system to work. “The probe acquires heart images and we run inference on a whole video clip of a given view or set of views of the heart (basically a sliced view of the moving heart),” says BayLabs ‘mad scientist’ Johan Mathe. (Note to concerned parents, I’ve met Johan and he appears to be reasonably sane.)

Play it again, HAL: New research from DeepMind shows how to teach computers to generate voices, music, and anything else. This brings us closer to a day where our phones can talk to us with intonation and, eventually, sarcasm, like Marvin the Paranoid Android. Check out the synthetic voices on the DeepMind blog and relax to some of the ghostly neural network piano tunes. This technology will also make it easier for people to create synthetic audio clips from known individuals, so propagandists could eventually conjure up an audio clip of Barack Obama calling for universal basic income, or another world leader issuing a declaration of war. The technique’s drawback is it involves processing 16,000 datapoints a second. This means it is – to use a technical term – bloody expensive. Optimization and hardware should change this over time.

Rise of the accelerators: Speaking of hardware… Intel is buying computer vision chip company Movidius, just weeks after snapping up the deep learning experts at Nervana. Intel’s view is that AI will require dedicated processors, probably paired with a traditional (Intel-made) CPU, and modifiable FPGAs (from recent Intel-acquisition Altera). Nvidia is continuing to design more deep learning-specific chips, adapting its graphical systems for AItasks. Meanwhile, companies like Google are designing their own systems from the ground up. It’s not clear yet if Intel can win this, but it’s certainly paying to get a seat at the table. The Next Platform has a nice analysis of these trends. Nuit Blanche points out need for radical new hardware – so, crazy IC geeks, please dive in! One reassuringly crazy idea is optical computing, see the website of startup LightOn.

Montréal Megabucks: the Université de Montréal, Polytechnique Montréal and HEC Montréal have been awarded $93,562,000 (Canadian) dollars to carry out research into deep learning, machine learning, and operations research. So I think this means UMontreal AI expert Yoshua Bengio can pick up the bill next time he goes out to dinner with his fellow researchers? It’s fantastic to see the Canadian government shovel money into a field that it helped start, long may the funding continue.

Is math mandatory?: How much math do you need to know to understand deep learning? There’s some debate. The proliferation of new software makes it relatively easy to get started with the software, but you’ll likely need to understand some of the technical components to diagnose complex bugs and to develop entirely new algorithms. That may require a greater understanding of the math involved. “ML has deep pitfalls, and mitigating them requires a foundational understanding of the mechanisms that make ML work,” writes Anton Troynikov. “Math is a tool, a language of sorts. Having a math background does not magically allow to “understand” anything, and in particular not ML,” writes Francois Chollet. “Math & CS can be used to model chess, but you don’t need to understand this formalism in order to play chess. Not even at the highest level. The same is true of the relationship between math & ML. Doing ML relies on intuitions which come from the practice of ML, not from math.” (Personally, I think learning more math can help you conceptualize aspects of deep learning.)

Neural network diagrams: Here’s a Google primer to some modern aspects of neural network development that pairs accurate, easy-to-grasp descriptions with some very powerful visualizations.

Too good to be true: Recently the AI research community was astir with the great and surprising results contained in a new paper, called Stacked Auto Regression Machines, that was published on Arxiv. The paper has now been withdrawn. One of the authors says they left out key evidence in the paper. “In the future, I will release a software package for public verification, along with a more detailed technical report,” they write. Good! The best way to attain trust in the AI community is to give people the code to replicate your results.

Oh dear. No, no, no, that’s not right at all is it? Deep learning perception systems do not work like human perception systems. University of Toronto AI researcher and ‘neural network technician’ Jamie Ryan Kiros has been exploring the faults inherent to one of these software systems and publishing the bloopers on Twitter. Check out Usain Bolt’s secret frisbee habit and the marvels of this edible boxing glove!

Thanks for reading. If you have suggestions, comments or other thoughts you can reach me at jack@jack-clark.net or tweet at me@jackclarksf

Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.

Baidu Paddles into AI Software: ‘It is a truth universally acknowledged, that an ambitious company in possession of a large technology team must be in want of its own artificial intelligence software’ – Jane Austin. No surprise then that Chinese technology giant Baidu announced its own free deep learning framework named ‘Paddle’, last week. Deep learning frameworks make it easier to design and build AI software, saving programmers time. Paddle will compete with similar software from Google (TensorFlow), Microsoft (CNTK), Amazon (DSSTNE), academic projects like Torch and Theano, and a smorgasbord of other add-ons, libraries, and tools from other companies. These tools have becoming marketing devices in their own right; Baidu announced Paddle at Baidu World in Beijing last week, ahead of the full release of the documentation, which is due for Sep 30th. Now taking bets on when Apple will release its own open source AI software…

There Can Only Be Three or Four: AI software will likely go the way of Linux distributions, with a few popular ones garnering the vast majority of users, and a long, long tail of less widely used ones extending out as far as the eye can see. Companies are likely going to compete heavily to gain the top spot(s), as that will make it easier for them to sell compatible cloud services to developers and to identify new talent to hire. This Hacker News poll shows Google’s TensorFlow and Francois Chollet’s Keras to be in the lead for now.

Amazon Builds UK AI Team: Amazon is building a new machine learning group in the UK and has persuaded Neil Lawrence of the University of Sheffield to join it. Multiple large companies have previously approached Mr Lawrence about joining, so Amazon has pulled off something of a coup here. The group will work alongside Ralf Herbrich’s group in Berlin. We hope Neil Lawrence will continue to write about AI for the research community and general public. He says he plans to. All the best!

Quantum AI Could Be Closer Than You Think: Experts don’t know if truly intelligent machines will require quantum computers, but a quantum computer certainly wouldn’t hurt, because the machines can solve problems that today’s computers simply can’t handle. Companies like Microsoft, Google, and IBM, are all looking into using quantum computers for applications in security and simulation, as well as to improve computer vision techniques and more. Now Google plans to use a quantum computer to simulate the behaviour of a random arrangement of quantum circuits. If it manages to do that, then the company will have made a significant step in tackling tasks that exceed the capabilities of traditional computers, say experts interviewed by New Scientist.

Industrial-scale Recommendations: Google has published a paper outlining the recommender system that underpins YouTube. It does some interesting things with regards to the age of videos to help it cotton-on to popular trends.

The (Real) Future of AI: Stanford has released the 2016 report of its One Hundred Year Study on Artificial Intelligence. Experts like Rodney Brooks (Rethink Robotics), Oren Etzioni (Allen Institute for AI), Astro Teller (Google X), Eric Horvitz (Microsoft), and many others have analyzed AI and tried to make some informed projections about where it is going. The answer includes autonomous cars, smarter home robots, greater health care systems, and the usual fuzzy projections about how it may influence employment. “AI will gradually invade almost all employment sectors, requiring a shift away from human labor that computers are able to take over,” they write. From a policy level it’d be great to have more funding for interdisciplinary study of the societal impacts ofAI, as well as clarifying how some rules (like the DMCA) could influence AIresearch, and employing more AI experts within government, they write. Download the report here.

Oooh, Fashion!: A website from Google and Zalando uses machine learning, massive amounts of data mining, and the input of over 600 fashion ‘professionals and influencers’, to create a tool that can procedurally generate new fashion items. Project Muze’s results are rather underwhelming, but it’s notable that both parties appeared to invest so many resources into the project. More to come, I expect.

You Have Questions, They Have Answers: Andrej Karpathy of OpenAI is doing a Quora Session on Thursday. Ask away.

Deep Learning Education: Jeremy Howard of Kaggle/Enlitic/Fast.AI has launched a Deep Learning course with the University of San Francisco. “We think this is the first ever in-person university-accredited deep learning certificate course. Only real prerequisite is decent coding expertise, although some memory of high-school level linear algebra (matrix-matrix products) and calculus (the chain rule) would be helpful,” he says. “There are scholarships available for anyone who isn’t able to pay for the course – although we’ve tried to keep the cost down too. (And I’m donating my teaching fees.)” Check out the course (and a good introductory deep learning video) at this link.

Thanks for reading. If you have suggestions, comments or other thoughts you can reach me at jack@jack-clark.net or tweet at me@jackclarksf

Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.

Cyborgs Are Closer Than You Think: Elon Musk says it would be a good idea for people to get some machinery wired into their brain to make them smarter and better able to compete with robots and AI. It turns out this is easier to do than you’d assume. A group of researchers published a paper yesterday that described “a lace-like electronic mesh that ‘you could literally inject’ into three-dimensional synthetic and biological structures like the brain.” This technology could eventually be used to deal with medical conditions and/or to enhance cognitive performance. “I think our goal is to do something, and I think it’s possible to, number one, correct deficiencies. And I wouldn’t mind adding a terabyte of memory,” said Professor Mark Hyman in an interview with Nautilus.

Please build this for me: Most modern AI techniques require a vast amount of labelled training data. If you talk to experts you’ll find that they each have intuitions about exactly how much data you’d need for a given task, whether that is a few thousand pictures for detailed classification, or a few hundred thousand words for text generation. Is it possible to build a small application that can, given a rough outline of the task (say, classify images at resolution Y with Z accuracy), estimate how much data you’ll need?

Neural phones: Samsung gave a talk at Hot Chips last week in which it said it was using neural networks for branch prediction in the M1 processor cores inside its S7 and S7 Edge smartphones. “If your CPU can predict accurately which instructions an app is going to execute next, you can continue priming the processing pipeline with instructions rather than dumping the pipeline every time you hit a jump. ‘The neural net gives us very good prediction rates,’ said Brad Burgess, who is Samsung’s chief CPU architect”, reports The Register. (As many have subsequently pointed out, people may have been using similar techniques for many years, but instead of calling it a neural network, they called it a perceptron. Marketing!)

Unleash the robots (with free data)! Right now, making truly smart robots is a challenge. That’s because the common way of developing modern AI software involves spending hours training your computers on test data. Modern computers are very fast so this means the computer can play a hundred games of Tetris in a few minutes. This approach doesn’t work very well for robotics. That’s because the simulators the robots are being trained on don’t fully reflect the fizzing complication of the real world. (New simulation environments are being developed, though, including Google DeepMind’s ‘Labyrinth’.)

So, if simulation is inefficient, what else can you do? The answer, if you have lots of money, time, and access to an office containing 14 robot arms, is to train your robots in the real-world. That’s what Google did earlier this year, when it created what Googlers called the ‘arm farm’. Over the course of several months its robots learned to pick up a variety of different objects through a process of (smart) trial and error. Data from one real-world robot was transferred to the others, letting the Mountain View, California company’s dexterous servants learn in the same networked way that Tesla’s self-driving cars do. So it was with a pleasant surprise that we saw Googlerelease the data from the experiments last week. This data gives researchers around 650,000 examples ofrobot grasping attempts and 59,000 examples ofpushing motions.

Transfer Learning: It’s likely that clever robots will be created through a combination of training in the virtual world and the real world. Being able to take insights gleaned from one environment, like a simulator, and apply them to another, like a real-world disaster zone, is one of the grand challenges in AI. Google DeepMind recently published a paper on ‘progressive networks‘, which lets it take a neural network that has learned to tackle one problem and daisy-chain it via gradient descent to a new network. This lets the new network tap into insights learned in other networks, reducing training time. This means you can train a bunch of networks in a simulator, then attach those to a neural network that is trying to tackle problems on a real-world robot, and learn to do things in less time than usual.

Computer, Enhance! Last week I asked for a service that could let me upscale my pictures using neural networks. It exists! Now I’ve found another nice example on Github with code. For an example of how upscaling can go wrong please look at the third-from-bottom picture on the Github repo.

Hands Across The Human-Machine Divide: Berkeley professor Stuart Russell will lead the Center for Human-Compatible Artificial Intelligence, which launched this week. “The center will work on ways to guarantee that the most sophisticated AI systems of the future, which may be entrusted with control of critical infrastructure and may provide essential services to billions of people, will act in a manner that is aligned with human values,” says Berkeley News. Funding for the center comes from the Open Philanthropy Project and the Leverhulme Trust and the Future of Life Institute.

Thanks for reading. If you have suggestions, comments or other thoughts you can reach me at jack@jack-clark.net or tweet at me@jackclarksf

Welcome to Import AI, a newsletter about artificial intelligence. Subscribe here.

Deep learning’s dirty data secret: modern deep learning approaches require a ridiculous quantity of data. How much data? A hacker house in SF used 3.8 terabytes of bandwidth in July. No coincidence that one of its residents is George Hotz, founder of self-driving car startup Comma.ai. The house has now upgraded to an ‘unlimited’ bandwidth plan, making Hotz foot the bill for the bit-flurry.

Free data! Comma.ai released seven and a quarter hours of highway driving data so world+dog can make their own AVs, and François Chollet, creator of the Keras deep learning framework, has made three image classification models available for free. That means you can now classify an image in as little as three lines of code.
( model = VGG16(weights=’imagenet’)
preds = model.predict(imgs)
print(decode_predictions(preds)) ).
How many lines to drive a car badly?

Show me the research: Spotting the difference between an AI startup with genuine technology and one that has re-implemented commodity systems is tricky. It’s easier to judge if the company participates in the wider AI research community. Therefore points should be awarded to robotics company Brain Corporation, text analysis startup Maluuba, and image recognition outfit Curious AI, which have all published papers recently. Keep them coming! (Extra kudos to Curious AI for publishing code.)

Spy VS Spy! Seven autonomous computer systems battled each other at the Darpa Cybersecurity Challenge in Las Vegas last week. ‘Mayhem’, the winner, was built by CMU startup For All Secure. The system performed automated program analysis (via symbolic execution) to find and exploit weaknesses in running programs. It appears to use a smart scheduler to let it run multiple checks on a piece of software in parallel, then selectively pause certain jobs to throw resources at promising vulnerabilities. This lets it efficiently explore the vast underbelly of the program and selectively focus on weak points, like a sophisticated, thieving octopus. (Unfortunately, some of the papers are paywalled.) Podcast with more information here.

Care for a little AI with your global government, sir? Google DeepMind would eventually like to donate its technologies to the United Nations, according to CEO Demis Hassabis.

Deep learning chips: Typical x86 processors are a bad fit for (most) modern AI tasks. GPUs are a bit better, but still not optimal. So expect new hardware substrates for AI. Microsoft will launch a new cloud service within a few months that lets people accelerate their neural network workloads with FPGA co-processors, said Qi Lu at the Scaled Machine Learning Conference. Google has already indicated it will offer its TPUs to people to accelerate TensorFlow workloads. Startup Nervana plans to produce its own chip to accelerate its ‘Neon’ software further. There are also numerous stealth chip and hardware startups that are re-thinking systems around deep learning (think: tens of thousands to millions of processors for low-precision matrix multiplication, etc).

GoodAI, a European AI research group founded by game developer Marek Rosa, has fleshed out its research and development roadmap and begun work on analyzing the overall AI landscape. It’s also developed some software for prototyping AI systems “with highly dynamic neural network topologies” named Arnold. It’s very pretty! The company will release it as open source software in a few months months, Rosa tells me.

Open VS Closed AI development? “Developing a joint set of openness guidelines on the short and long term would be a worthwhile endeavor for the leading AI companies today,” says Victoria Krakovna, co-founder of the Future of Life Institute.

Imagine that, computer. Generative models are gathering interest (as mentioned last week) and have a huge number of potential applications. For more information check out these two slide decks from Russ Salakhutdinov at CMU and Shakir Mohamed from DeepMind, from their sessions at the Deep Learning Summer School this week.

You have questions, they have answers: Google’s Brain team will be taking questions on Reddit and OpenAI’s Ian Goodfellow on Quora. Both on August 11th.

Thanks for reading. If you have suggestions, comments or other thoughts you can reach me at jack@jack-clark.net or tweet at me@jackclarksf

Artist’s depiction of the surprising popularity of deep learning techniques across a variety of disciplines.

The pace of development of artificial intelligence is going to get faster. And not for the typical reasons — More money, interest from megacompanies, faster computers, cheap&huge data, and so on. Now it’s about to accelerate because other fields are starting to mesh with it, letting insights from one feed into the other, and vice versa.

That’s the gist of a new book by David Beyer, which sees him interview 10 experts about artificial intelligence. It’s free. READ IT. The main takeaway is that neural networks are drawing sustained attention from researchers across the academic spectrum.  “Pretty much any researcher who has been to the NIPS Conference [a big AI conference] is beginning to evaluate neural networks for their application,” says Reza Zadeh, a consulting professor at Stanford. That’s going to have a number of weird effects.

(Background: neural networks come in a huge variety of flavors — RNNs! CNNS! LSTMs! GANs! Various other acronyms! — but people like them because they basically let you chop out a bunch of hand-written code in favor of feeding inputs and outputs into neural nets and getting computers to come up with the stuff in-between. In technical terms, they infer functions. In the late 00’s some clever academics rebranded a subset of neural network techniques to ‘Deep Learning’, which just means a stack of different nets on top of one another, forming a sort of computationally-brilliant lasagne. When I say ‘machine learning’ in this blogpost, I’m referring to some kind of neural network technique.)

Robotics has just started to get into neural networks. This has already sped up development. This year, Google demonstrated a system that teaches robotic arms to learn how to pick up objects of any size and shape. That work was driven by research conducted last year at Pieter Abbeel’s lab in Berkeley, which saw scientists combine two neural network-based techniques (reinforcement learning and deep learning) with robotics to create machines that could learn faster. Robots are also getting better eyes, thanks to deep learning as well. “Armed with the latest deep learning packages, we can begin to recognize objects in previously impossible ways,” says Daniela Rus, a professor in CSAIL at MIT who works on self-driving cars.

More distant communities have already adapted the technology to their own needs. Brendan Frey runs a company called Deep Genomics, which uses machine learning to analyze the genome. Part of the motivation for that is that humans are “very bad” at interpreting the genome, he says. That’s because we spent hundreds of thousands of years evolving finely-tuned pattern detectors for things we saw and heard, like tigers. Because we never had to hunt the genome, or listen for its fearsome sounds, we didn’t develop very good inbuilt senses for analyzing it. Modern machine learning approaches give us a way to get computers to analyze this type of mind-bending data for us. “We must turn to truly superhuman artificial intelligence to overcome our limitations,” he says.

Others are using their own expertise to improve machine learning. Risto Miikkulainen is an evolutionary computing expert who is trying to figure out how to evolve more efficient neural networks, and develop systems that can help transfer insights from one neural network into another, similar to how reading books lets us extract some data from a separate object (the text) and port into our own grey-matter. Benjamin Recht, a professor at UC Berkeley, has spent years studying control theory — technology that goes into autonomous capabilities in modern aircraft and machines. He thinks that fusing control theory and neural networks “might enable safe autonomous vehicles that can navigate complex terrains. Or could assist us in diagnostics and treatments in health care”.

One of the reasons why so many academics from so many different disciplines are getting involved is that deep learning, though complex, is surprisingly adaptable. “Everybody who tries something seems to get things to work beyond what they expected,” says Pieter Abbeel. “Usually it’s the other way around.” Oriol Vinyals, who came up with some of the technology that sits inside Google Inbox’s ‘Smart Reply‘ feature, developed a neural network-based algorithm to plot the shortest routes between various points on a map. “In a rather magical moment, we realized it worked,” he says. This generality not only encourages more experimentation but speeds up the development loop as well.

(One challenge: though neural networks generalize very well, we still lack a decent theory to describe them, so much of the field proceeds by intuition. This is both cool and extremely bad. “It’s amazing to me that these very vague, intuitive arguments turned out to correspond to what is actually happening,” says Ilya Sutskever, research director at OpenAI., of the move to create ever-deeper neural network architectures. Work needs to be done here. “Theory often follows experiment in machine learning,” says Yoshua Bengio, one of the founders of the field. Modern AI researchers are like people trying to invent flying machines without the formulas of aerodynamics, says Yann Lecun, Facebook’s head of AI.)

The trillion-dollar (yes, really) unknown in AI is how we get to unsupervised learning — computers that can learn about the world and carry out actions without getting explicit rewards or signals. “How do you even think about unsupervised learning?” wonders Sutskever. One potential area of research is generative models, he says. OpenAI just hired Alec Radford, who did some great work on GANs. Others are looking at this as well, including Ruslan Salakhutinov at the University of Toronto. Yoshua Bengio thinks it’s important to develop generative techniques, letting computers ‘dream’ and therefore reason about the world. “Our machines already dream, but in a blurry way,” he says. “They’re not yet crisp and content-rich like human dreams and imagination, a facility we use in daily life to imagine those things which we haven’t actually lived.”

The Deep Learning tsunami, tsunami-ing.

My personal intuition is that deep learning is going to make its way into an ever-expanding number of domains. Given sufficiently large datasets, powerful computers, and the interest of subject-area experts, the Deep Learning tsunami (see picture), looks set to wash over an ever-larger number of disciplines. Buy a swimsuit! (And read the book!)

KREUZBERG, BERLIN, 2017 — EXHIBITION LAUNCH OF “YES, COMPUTERS DO DREAM OF ELECTRIC SHEEP”:

Gallery-goers wander halls full of lifesize sculptures of sheep. Each sheep has a different, unique mutation, leading to one with six legs and another whose wool has been replaced with miniature spanners and hammers, the size of boardgame pieces, woven together in shining metal braids. Suddenly, the din of the house music is stalled and a trio of suited people stride in. One of them has a megaphone. “THIS IS AN ILLEGAL ART SHOW ,” they crackle. “THESE SHEEP WERE GENERATED USING PROPRIETARY DATA. WE’RE CONFISCATING THE SCULPTURES AND ALL GALLERY-GOERS WILL BE SUBJECTED TO A DATA AUDIT.” – fictional scenario, based on current AI research.

THE TROUBLE WITH MACHINE ART 

The increasing sophistication of Deep Learning artificial intelligence techniques are going to lead to a new type of generative art. That’s going to be exciting for our culture, but may draw the ire of rights holders.

Artists have been working with computers since they were invented and have used techniques like procedural programming, cellular automata, and more to explore the new creative territories that computers let them access. New technologies coming out of the current AI boom will accelerate this.

Don’t believe me? Take a look at this recent paper from a group of researchers at German and US institutions: “A Neural Algorithm of Artistic Style” [1]

Researchers have figured out how to take the artistic style from one painting and apply it to a new image.

In the days after this research was published third-parties figured out how to implement the system and started generating their own images.

Third-parties are able to rapidly re-implement new discoveries in free software, accelerating the speed with which new artistic techniques find their way into culture.

They were able to do this because there’s a wealth of free software packages available for running deep learning algorithms ranging from Theano to Caffe to Torch, and more, and the researchers published their paper as open-access, so people could access it for free. That speaks to the overall speed of invention within AI, which is accelerating as more people enter the field and publish research, or free code.

[Edit: One day after this post was published someone posted an animation to Reddit [2] showing a generative system drawing the Eiffel Tower in a style reminiscent of Van Gogh.]

This shows how a neural network can “imagine” a never-before-seen image in a particular aesthetic style.

Where this gets complicated is the issue of copyright as deep learning systems need to be fed with vast amounts of data. Typically, that’s done through open access datasets compiled by academic researchers, or private stores of information amassed by companies like Google, Facebook, and others.

Individual artists have other needs, and my suspicion is that they’ll do what they’ve always done – hunt through the available imagery, pick the ones they like, and make great art out of the images. And, just as in the past, this will raise valid and complex questions about the originality of the generated work, just as it has done with the free-for-all collage art we see coming out of social platforms like Tumblr and Vine. That’s going to create conversations about fair use as people share their Neural Network recipes with others.

I chatted about this issue recently with @Samim and @graphific on the Ethical Machines [3] podcast. [4]

This is not an isolated incident: it follows Google outlining an earlier system called “DeepDream” [5] in a blog post in July. That system let you use the feature detectors from a trained neural network to enhance new images, applying the proclivities of the AI system to never-before-seen entities. The internet was rapidly flooded with pictures made using this technique. Google even published the code on GitHub [6]. And, inevitably, it led to websites like DeepDreamGenerator [7] where anyone — no neural network expertise required – can make their own images, leading to ghastly moments like this on peoples’ Facebook feeds. Hold on tight, things are about to get WEIRD.

New AI art techniques get mainstream very, very quickly.

[1] http://arxiv.org/pdf/1508.06576v1.pdf
[2] https://www.reddit.com/r/MachineLearning/comments/3iygt2/neural_art_in_action/
[3] http://ethicalmachines.com/
[4] https://soundcloud.com/samim/ethical-machines-ep2-jack-clark
[5] http://googleresearch.blogspot.com/2015/06/inceptionism-going-deeper-into-neural.html
[6] https://github.com/google/deepdream
[7] http://deepdreamgenerator.com/