AI Monthly Digest #11 – From OpenAI partnering with Microsoft to battle.net Blade Runners

/in , /by and

AI models are skilled in Chess, Go, StarCraft and, since July, six-player Texas Hold’em Poker. But the hunt for inhuman players has begun.

And when AI models get bored with beating humans in games, it’s apparently time for a bike ride. Read on to find out just why.

OpenAI partners with Microsoft

In March, OpenAI shifted to a for-profit paradigm, forging OpenAI LP. The company has now entered into a strategic partnership with Microsoft to build a new computational platform in Azure Cloud. OpenAI will port its codebase into Azure and develop new solutions with tools available there.

The main benefit for Microsoft will be the access it gains to the fruits of OpenAI’s work as a preferred partner for commercializing new AI technologies. OpenAI was formed with the goal of improving people’s lives with technology and has delivered multiple AI breakthrough models including MuseNet, which produces music in various styles, and GPT-2, a natural language processing model and gold standard in text generation.

For more on the joint venture, click on over to the official press release.

Why does it matter?

As the company recently announced, taking a non-profit approach to develop AI proved too daunting even for the likes of Elon Musk and Sam Altman, the company’s CEO. Unlike with most software, developing artificial intelligence requires not only skilled and talented people but also an astonishing amount of computing power. Training a single GPT-2 model is estimated to cost up to $50,000 – and that’s only for one of many experiments run in a given year. So, securing access to computing power is a must. By providing that, Microsoft will be getting access to the base of talented AI developers, significantly increasing its AI development potential.

AI beats skilled human players in six-player poker

Pluribus bot is the first AI-controlled agent capable of beating human pro players in six-player no-limit Hold’em poker, the most popular format of this game in the world. Unlike chess and Go, poker is a game with hidden information – the player cannot see the hands of other players. The game itself involves bluff and a great deal of psychological factors. AI bots were good at beating one opponent, but going up against more than one was a major milestone.

More details are available in the ai.facebook blogpost.

Why does it matter

Dealing with a one-on-one situation, though common in recreational games, is rare when solving real-life problems. Moreover, it is somewhere between difficult and impossible to get all the information one needs, as in chess or Go. Delivering models that work in a limited-information, multiple-agent environment is pushing models that support demand forecasting or managing cybersecurity threats.

When an autonomous car is not enough

Cars, be they traditional or autonomous, come with various disadvantages, especially in cities. They get stuck in traffic jams, and require parking, which can be hard to come by.

To provide more sustainable autonomous transport, Chinese scientists have brought out an autonomous bicycle. The machine responds to voice commands, avoids obstacles and maintains balance. It uses a new Tianjic chip, which supports processing neuroscience-inspired algorithms. Check out the video below to see how it works.

Why does it matter

The research itself sounds like a lot of fun, but it also constitutes an excellent foundation for further work on autonomous transportation. A bike can be used for delivering fast food, or modified to work as a motorized Rickshaw in our ever more crowded metropolises. Couriers can used them to deliver the mail and other documents or to transport individuals incapable of riding a bike.

Autonomous bikes need not be fully autonomous, by the way: AI can support the driving process or provide alerts to riders.

Alan Turing will be featured on 50 pound banknote

Often hailed as the godfather of modern computing, Alan Turing is widely known for his work on cracking the famed German Enigma code and as the leader of the team that enhanced the cracking methods delivered by Polish mathematicians.

Turing is also considered a pioneer of artificial intelligence. He came up with the Turing Test as a first way to determine if a machine mimicking a human in conversation is truly intelligent.

So great was Turing’s contribution to humanity that he will now be featured on Britain’s 50-pound banknote. Chosen from 227,299 nominations covering 989 eligible characters, Turing was ultimately picked by Mark Carney, Bank of England governor.

Why does it matter

The announcement is a sign that computer science is no longer considered a novelty and prominent AI researchers earn the same respect chemists, physicists or life sciences experts do, as representation on a banknote well attests.

BattleNet Blade Runners

Deepmind, in conjunction with Blizzard, has deployed AlphaStar model on Battle.net, to allow players to test their skills and mettle against artificial intelligence. Battle.net is an official platform connecting players from all around the world, enabling them to quickly find opponents for a multiplayer match.

There is just one twist: the famed, reinforcement learning-trained AlphaStar will play anonymously, thus allowing players to compete with the model as they would do in any match with a normal opponent.

AlphaStar has been developed significantly beyond the abilities it commanded in defeating human professional players MaNa and TLO. Deepmind capped the actions-per-minute and actions-per-second rate to make it more accurately appropriate human abilities limited by muscles and the need to operate a mouse and keyboard.  The model’s perception has also been narrowed to a single frame to come in line with what human players see on the screen.

Finally, the model is able to control and compete in any race given, be it Terran, Protoss or Zerg, representing all the factions available in the game. This represents serious progress: during matches in January, the model could only control Protoss units fighting against other Protoss.

Why does it matter

At the moment, it doesn’t. But let the experiment run its course and tune in later for an update. We anticipate more impressive progress.

Interestingly, this most recent news thrilled the players’ community, which all too clearly remembers the wounds AlphaStar inflicted in dominating renowned pros. Given that any match with the model is counted as a normal encounter and a ranking match, on-guard players try to spot and avoid AlphaStar lurking in the muddy waters of the battle.net rankings.

Players have reported “odd” behavior of some opponents and have been uploading videos on YouTube, where they discuss if the other player is actually AlphaStar incognito. They also advise each other to check if the partner responds to messages. Being called a noob by an opponent even once can be strong evidence that the opponent on the other side of the battlefield isn’t human.

So players are on the hunt for a replicant.

AI Monthly Digest #10 - AI tackles climate change and deciphers long-forgotten languages

AI Monthly Digest #10 – AI tackles climate change and deciphers long-forgotten languages

/in , /by and

June brought record-breaking temperatures, perfectly highlighting the global challenge of climate change. Is that AI-related news? Check and see in the latest AI Monthly Digest.

A common misconception about machine learning projects is that they are by definition big.  However, any number of AI-powered micro-tweaks and improvements are applied in everyday work. A good example of both micro and macro tweaks that can fix a major problem can be found in the paper described below.

AI tackling climate change

The world witnessed an extraordinarily hot June, with average temperatures 2 degrees celsius above normal in Europe. According to the World Meteorological Organization, the heatwave is consistent with predictions based on greenhouse gas concentrations and human-induced climate change.

Tackling this challenge will not be easy: according to World Bank Data, fossil fuel energy consumption still stacks to 79% of total. Furthermore, greenhouse gasses, particularly methane, are emitted by cattle, with livestock being responsible of 14.5% of total human-induced greenhouse emissions.

The most prominent figures in AI today, including DeepMind CEO Demis Hassabis, Turing award winner Yoshua Bengio, and Google Brain co-founder Andrew Ng, have authored a comprehensive paper on ways that AI can tackle the changing climate.

Their call for collaboration is meant to inspire practitioners, engineers and investors to deliver short- and long-term solutions for measures within our reach. Those include producing low-carbon electricity through better forecasting, scheduling, and control for variable sources of energy, mitigating the damage produced by high-carbon economies through, for example, better predictive maintenance as well as help minimize energy use in transportation, smart buildings and cities. The applications can vary from designing grid-wide control systems or optimizing scheduling with more accurate demand forecasting.

Why does it matter

Climate change is one of the greatest challenges mankind faces today, with truly cataclysmic scenarios approaching. Further temperature increases may lead to a variety of disasters, from flooding coastal regions due to melting ice caps, agricultural crises and conflicts over access to water.

Green energy promises solutions, yet these are not without their challenges, many of which could be solved with machine learning, deep learning or reinforcement learning. Responsibility is among deepsense.ai’s most important AI trends, and being responsible for the planet would be an excellent example of just why we chose to focus on that trend.

We will provide more in-depth content on climate change and AI-powered ways of tackling it. So stay tuned!

Giants racing to produce the best image recognition

If machine learning is today’s equivalent of the steam engine revolution, data and hardware are the coal and engine that power the machines. Facebook and Google are like the coal mines of yesteryear, having access to large amounts of fuel and power to build new models and experiment.

It should come as no surprise that breakthroughs are usually powered by the tech giants. Google’s state of the art in image recognition, EfficientNet, has been a recent giant step forward. The model was delivered by automated searching procedure uniformly scaling each dimension of the network in order to find the best combination.

EfficientNet stands for something.

The result is state-of-the-art in Image recognition. At least when it comes to combining efficiency and accuracy. But not when it comes to accuracy alone.

Not even a month later Facebook delivered a model that outperformed Google’s. The key lay in scaling the enormous dataset it was trained on. The social media mogul has access to Instagram’s database, which holds no less than billions of user-tagged images, a dataset ready to be chewed over by a hungry deep learning model.

The neural network was released to the public using a recently launched Pytorch Hub platform for sharing cutting edge models.

Why does it matter

Both advances show how important machine learning is for the tech giants and how much effort they invest in pushing their research forward. Every advancement in image recognition brings new breakthroughs closer. For example, models are becoming more accurate in detecting diabetic retinopathy using images of the eye. Every further development delivers new ways to solve problems that would be unsolvable without ML (Machine learning) – manufacturing for visual quality control is among the best examples.

XLNet outperforms BERT

As we noted in a past AI Monthly Digest, Google has released Bidirectional Encoder Representation from Transformations (BERT). BERT was, until recently, the state-of-the-art when it comes to Natural Language Processing benchmarks. The newly announced XLNet is an autoregressive pretraining method (as opposed to an autoencoder-like BERT) which learns a language model by predicting the next word in a sequence using the permutation of all the surrounding words. An intuitive explanation can be found (here).

The XLNet model proved more effective than BERT in beating all 20 benchmark tasks.

Why does it matter

Understanding a natural language was considered a benchmark for intelligence, with Alan Turing’s test being among the best examples. Every push forward delivers new possibilities in building new products and solving problems, be they business ones or something more uncommon, like the example below.

AI-powered archeology? Bring it on!

Deep learning-based models are getting even better at understanding natural language. But what about language that is natural, but has never been deciphered due to lack of knowledge or a frustratingly small amount of extant text?

Recent research from MIT and Google shows that a machine learning approach can deliver major improvements in deciphering ancient texts. In the basics of modern natural language processing techniques, all of the words in a given text are assumed to be related to each other. The machine itself doesn’t “understand” text it in a human way, but rather forms its own assumptions based on the relations and connotations of each word in a sentence.

Disc of Phaistos, one of the most famous mysteries of archaeology

In this approach, the translation process is not built on understanding the world, but rather finding similarly connotated words that transfer the same message. This is entirely different than humans’ approach to language.

By making the algorithm less data-hungry, the researchers deliver a model that translates texts from rare and long-lost languages. The approach is described in this paper.

Why does it matter

While there are countless examples of machine learning in business, there are also new horizons to discover in the humanities. Deciphering the secrets of the past is every bit as exciting as building defenses against the challenges of the future.

The more sophisticated approach to and possible brute-force breaking of unknown languages provides a way to uncover more language-related secrets.

A Disc of Phaistos? Or a Voynich manuscript maybe?

AI Monthly Digest #9 – the double-edged sword of modern technology

AI Monthly Digest #9 – the double-edged sword of modern technology

/in , /by and

This edition is all about AI morality-related themes, with a slight tinge of Talking Heads and Modern Talking.

Earlier this year, deepsense.ai highlighted AI morality and transparency as one of 2019’s dominant AI trends. May bore out our thesis, especially as it relates to potential misuse and malicious intent.  At the same time, though, AI provides unique chances to support entertainment and education, as well as deliver new business cases.

A bigger version of GPT-2 released to the public

Open-AI has recently shown the GPT-2 model has set a new gold standard for natural language processing. Following the acclaimed success of the model, OpenAI opted not to make it public due to the risk of malicious usage, particularly to produce spam and fake news at no cost.

This sparked an uproar. The industry good practice is to release AI research work as open-source software, so other researchers can push the boundaries further without having to repeat all the work done earlier from scratch. In other words – OpenAI threw up a major hurdle to NLP-model development by keeping GPT-2 under wraps.

To support the scientific side of the equation while reducing the malicious threat, OpenAI releases some smaller-scale models to the public. The model it recently released operates on 345M parameters, while the best original model consists of 1.5B parameters. Every parameter can be seen as a virtual neuron inside a neural network, so OpenAI is basically reducing the brain it designed.

The original network was released to OpenAI partners currently working on malice-proofing the system. The first independent applications of the downscaled network are already available at talktotransformer.com and onionbot headline generator.

Why does it matter?

OpenAI is currently facing a difficult choice between supporting the global development of AI and the fear of losing control over dangerous technology. In a world facing a potential avalanche of fake news and social media being used to perpetuate propaganda, building a system that writes coherent and convincing texts is undoubtedly dangerous.

This case allows one to see all the AI-related issues in a nutshell, including the technology’s amazing potential, the real threat of misuse or malicious intent. So the case may serve as a precedent for future cases.

Talking heads unleashed

A group of scientists working for Samsung’s AI Center in Moscow and Skolkovo Institute of Science and Technology designed a model that can produce a convincing video of a talking head from a single image, such as a passport photo or even a painting.

The model renders with consistency both the background and the head’s behavior. Most impressively, the model builds a convincing video of a talking head from even a single image of the frame.

The solution is searching for a similar face that was analyzed and extracts facial features including a nose, chin, mouth and eyes. The movement of those features is then applied on the image, as shown in the video.

The results are undoubtedly impressive.

Why does it matter?

Yet another AI ethics-related issue, the talking-head technology poses the threat of deepfakes, images that show a person making statements that he or she would never make. This raises obvious questions about the malicious ways such technology could be used.

On the other hand, when deepfakes are used for special effects in popular movies, no one seems to complain and critics even weigh in with their acclaim. Some of the better-known examples come from the Star Wars franchise, particularly Rogue One, which features Leia Organa wearing the face of a young Carrie Fisher.

AI has also proved itself useful in promoting art. By leveraging this technology it is possible to deliver the talking head of Girl with a Pearl Earring or the Mona Lisa telling visitors from screens about a painting’s historical context – a great way to put more fun in art lessons for kids. Or just to have some fun seeing what a Stallone-faced Terminator would look like.

Again, AI can be used for both good and evil ends. The ethics are up to the wielder of this double-edged sword.

Modern Talking – recreating the voice of Joe Rogan

Another example of deepfake-related technology is using AI to convincingly recreate Joe Rogan’s voice. The text-to-speech technology is not a new kid on the block, yet it is easy to spot due to the robotic and inhumanely calm style of speaking. Listening to automated text-to-speech was usually boring at best while delivering the unintentional comic effects of robotic speech, all in the absence of emotion or inflection.

Dessa engineers have delivered a model that is not only transforming text to speech, but also recreating Joe Rogan’s style of speaking. Joe is a former MMA commentator who went on to become arguably the most popular podcaster in the world. Speaking with great emotion, heavily accenting and delivering power with every word, Rogan is hard to mistake.

Or is he? The team released a quiz that challenges the listener to distinguish if a given sample comes from a real podcast or was AI-generated. The details can be found on Dessa’s blog.

Why does it matter?

Hearing a convincing imitation of a public personality’s voice is nearly as unsettling as watching a talking head talk. But the technology can be used for entertainment and educational purposes. For example, delivering a new Frank Sinatra single or presenting Winston Churchill’s comprehensive and detailed speech on reasons behind World War II.

Again, the ethics are in the user’s hands, not in the tool. Despite that, and as we saw with OpenAI’s GPT-2 Natural Language Processing model, researchers have decided NOT to let the model go public.

Machine learning-powered translations increase trade by 10,9%

Researchers at Olin Business School at Washington University in St.Louis have found a direct connection between machine learning-powered translations and business efficiency. The study was conducted on e-Bay and shows that moderate improvement in the quality of language translation increased trade between countries on eBay by 10.9%.

The study examined the trade between English speakers from the United States and their trade relations with countries speaking other languages in Europe, America and Asia. More on the research can be found on the Washington University of St.Louis website.

Why does it matter?

While there is no doubt that AI provides vital support for business, the evidence, while voluminous, remains largely anecdotal (sometimes called anec-data) with little quantitative research to back up the claim. Until the Olin study, which does provide hard and reliable data. Is justified true belief knowledge? That’s an entirely different question…

A practical approach to AI in Finland

AI Monthly Digest #5 presented a bit about a Finnish way of spreading the word about AI. Long story short: contrary to many approaches of building AI strategy in a top-down model, Finns have apparently decided to build AI-awareness as a grassroots movement.

To support the strategy, the University of Helsinki has released a digital AI course on the foundations and basic principles of AI. It is available for free to everyone interested.

Why does it matter?

AI is gaining attention and the reactions are usually polarised – from fear of losing jobs and machine rebellion to arcadian visions of an automated future with no hunger or pain. The truth is no doubt far from either of those poles. Machine learning, deep learning and reinforcement learning are all built on certain technological foundations that are relatively easy to understand, including their strengths and limitations. The course provides good basic knowledge on these issues, which can do nothing but help our modern world.

AI Monthly Digest #8 – new AI applications for music and gaming

AI Monthly Digest #8 – new AI applications for music and gaming

/in , /by and

The April edition of AI Monthly Digest looks at how AI is used in entertainment, for both research and commercial purposes.

After its recent shift from non-profit to for-profit, OpenAI continues to build a significant presence in the world of AI research. It is involved in two of five stories chosen as April’s most significant.

AI Music – spot the discord…

While machine learning algorithms are getting increasingly better at delivering convincing text or gaining superior accuracy in image recognition, machines struggle to understand the complicated patterns behind the music. In its most basic form, the music is built upon repetitive motifs that return based on sections of various length – it may be a recurrent part of one song or a leading theme of an entire movie, opera or computer game.

Machine learning-driven composing is comparable to natural language processing – the short parts are done well but the computer gets lost when it comes to keeping the integrity of the longer ones. April brought us two interesting stories regarding different approaches to ML-driven composition.

OpenAI developed MuseNet, a neural network that produces music in a few different styles. Machine learning algorithms were used to analyze the style of various classical composers, including Chopin, Bach, Beethoven and Rachmaninoff. The model was further fed rock songs by Queen, Green Day and Nine Inch Nails and pop music by Madonna, Adele and Ricky Martin, to name a few. The model learned to mimic the style of a particular artist and infuse it with twists. If the user wants to spice up the Moonlight Sonata with a drum, the road is open.

OpenAI has rolled out an early version of the model and it performs better when the user is trying to produce a consistent piece of music, rather than pair up a disparate coupling of Chopin and Nine Inch Nails-style synthesizers.

OpenAI claims that music is a great tool with which to evaluate a model’s ability to maintain long-term consistency, mainly thanks to how easy it is to spot discord.

…or embrace it

While OpenAI embraces harmony in music, Dadabots has taken the opposite tack. Developed by Cj Carr and Zack Zukowski, Databots model imitates rock, particularly metal bands. The team has put their model on YouTube to deliver technical death metal as an endless live stream – the Relentless Doppelganger.

While it is increasingly common to find AI-generated music on Bandcamp, putting a 24/7 death metal stream on YouTube is undoubtedly something new.

Fans of the AI-composed death metal have given the music rave reviews. As The Verge notes, the creation is “Perfectly imperfect” thanks to its blending of various death metal styles, transforming vocals into a choir and delivering sudden style-switching.

It appears that bare-metal has ushered in a new era in technical death metal.

Why does it matter?

Researchers behind the Relentless Doppelganger remark that developing music-making AI has mainly been based on classical music, which is heavily reliant on harmony, while death metal, among others, embraces the power of chaos. It stands to reason, then, that the  music generated is not perfect when it comes to delivering harmony. The effect is actually more consistent with the genre’s overall sound. What’s more, Databots’ model delivers not only instrumentals, but also vocals, which would be unthinkable with classical music. Of course, the special style of metal singing called growl makes most of the lyrics incomprehensible, so little to no sense is actually required here.

Related:  AI Monthly Digest #7 - machine mistakes, and the hard way to profit from non-profit

From a scientific point of view, OpenAI delivers much more significant work. But AI is working its way into all human activity, including politics, social problems and policy and art. From an artistic point of view, AI-produced technical death metal is interesting.

It appears that when it comes to music, AI likes it brutal.

AI in gaming goes mainstream

Game development has a long and uneasy tradition of delivering computer players to allow users to play in single-player mode. There are many forms of non-ML-based AI present in video games. They are usually based on a set of triggers that initiate a particular action the computer player takes. What’s more, modern, story-driven games rely heavily on scripted events like ambushes or sudden plot twists.

This type of AI delivers an enjoyable level of challenge but lacks the versatility and viciousness of human players coming up with surprising strategies to deal with. Also, the goal of AI in single-player mode is not to dominate the human player in every way possible.

Related:  AI Monthly Digest #6 - AI ethics and artificial imagination

The real challenge in all of this comes from developing bots, or the computer-controlled players, to deliver a multiplayer experience in single-player mode. Usually, the computer players significantly differ from their human counterparts and any transfer from single to multiplayer ends with shock and an instant knock-out from experienced players.

To deliver bots that behave in a more human way yet provide a bigger challenge, Milestone, the company behind MotoGP 19, turned to reinforcement learning to build computer players to race against human counterparts. The artificial intelligence controlling opponents is codenamed A.N.N.A. (Artificial Neural Network Agent).

A.N.N.A. is a neural network-based AI that is not scripted directly but created through reinforcement learning. This means developers describe an agent’s desired behaviour and then train a neural network to achieve it. Agents created in this way show more skilled and realistic behaviors, which are high on the wish list of Moto GP gamers.

Why does it matter?

Applying ML-based artificial intelligence in a mainstream game is the first step in delivering a more realistic and immersive game experience. Making computer players more human in their playing style makes them less exploitable and more flexible.

The game itself is an interesting example. It is common in RL-related research to apply this paradigm in strategic games, be it chess, GO or Starcraft II for research purposes. In this case, the neural network controls a digital motorcycle. Racing provides a closed game environment with a limited amount of variables to control. Thus, racing in a virtual world is a perfect environment to deploy ML-based solutions.

In the end, it isn’t the technology but rather gamers’ experience that is key. Will reinforcement learning bring a new paradigm of embedding AI in games? We’ll see once gamers react.

Bittersweet lessons from OpenAI Five

Defense of The Ancients 2 (DOTA 2) is a highly popular multiplayer online battle arena game with two teams, each consisting of five players fighting for control over a map. The game blends tactical, strategic and action elements and is one of the most popular online sports games.

OpenAI Five is the neural network that plays DOTA 2, developed by OpenAI.

The AI agent beat world champions from Team OG during the OpenAI Five Finals on April 13th. It was the first time an AI-controlled player has beaten a pro-player team during a live-stream.

Why does it matter?

Although the project seems similar to Deepmind’s AlphaStar, there are several significant differences:

  • The model was trained continuously for almost a year instead of starting from zero knowledge for each new experiment – the common way of developing machine learning models is to design the entire training procedure upfront, launch it and observe the result. Every time a novel idea is proposed, the learning algorithm is modified accordingly and a new experiment is launched starting from scratch to get a fair comparison between various concepts. In this case, researchers decided not to run training from scratch, but to integrate ideas and changes into the already trained model, sometimes doing elaborate surgery on their artificial neural network. Moreover, the game received a number of updates during the training process. Thus, the model was forced at some points not to learn a new fact, but to update its knowledge. And it managed to do so. The approach enabled the team to massively reduce the computing power over the amount it had invested in training previous iterations of the model.
  • The model effectively cooperated with human players – The model was available publicly as a player, so users could play with it, both as ally and foe. Despite being trained without human interaction, the model was effective both as an ally and foe, clearly showing that AI is a potent tool to support humans in performing their tasks — even when that task is slaying an enemy champion.
  • The research done was somewhat of a failure – The model performs well, even if building it was not the actual goal. The project was launched to break a previously unbroken game by testing and looking for new approaches. The best results were achieved by providing more computing power and upscaling the neural network. Despite delivering impressive results for OpenAI, the project did not lead to the expected breakthroughs and the company has hinted that it could be discontinued in its present format. A bitter lesson indeed.

Blurred computer vision

Computer vision techniques deliver astonishing results. They have sped up the diagnosing of diabetic retinopathy, built maps from satellite images and recognized particular whales from aerial photography. Well-trained models often outperform human experts. Given that they don’t get tired and never lose their focus, why shouldn’t they?

But there remains room for improvement for machine vision, as researchers from KU Leuven University in Belgium report. They delivered an image that fooled an algorithm, rendering the person holding a card with an image virtually invisible to a machine learning-based solution.

Why does it matter?

As readers of William Gibson’s novel Zero Hour will attest, images devised to fool AI are nothing new. Delivering a printable image to confound algorithm highlights a serious interest among malicious players interfering with AI.

Examples may include images produced to fool AI-powered medical diagnostic devices for fraudulent reasons or sabotaging road infrastructure to render it useless for autonomous vehicles.

AI should not be considered a black box and algorithms are not unbreakable. As always, reminders of that are welcome, especially as responsibility and transparency are among the most significant AI trends for 2019.

AI-monthly_digest_7

AI Monthly Digest #7 – machine mistakes, and the hard way to profit from non-profit

/in , /by and

March saw some major events concerning top figures of the ML world, including OpenAI, Yann LeCun, Geoffrey Hinton and Yoshua Bengio.

The past month was also the backdrop for inspiring research on how machines think and how different they can be from humans when provided with the same conditions and problem to solve.

OpenAI goes from non-profit to for-profit

OpenAI was initially the non-profit organization focused on pushing the boundaries of artificial intelligence in the same manner the open-source organizations are able to deliver the highest-class software. The Mozilla Foundation and the Linux Foundation, the non-profit powerhouses behind the popular products, are the best examples.

Yet unlike the popular software with development powered by human talent, AI requires not only brilliant minds but also a gargantuan amount of computing power. The cost of reproducing the GPT-2 model is estimated to be around $50,000 – and that’s only one experiment to conduct. Getting the job done requires a small battalion of research scientists tuning hyperparameters, debugging and testing the approaches and ideas.

Related:  AI Monthly Digest #6 - AI ethics and artificial imagination

Staying on technology’s cutting edge pushed the organization toward the for-profit model to fund the computing power and attracting top talent, as the company notes on its website.

Why does it matter?

First of all, OpenAI was a significant player on the global AI map despite being a non-profit organization. Establishing a for-profit arm creates a new strong player that can develop commercial AI projects.

Moreover, the problem lies in the need for computing power, marking the new age of development challenges. In a traditional software development world, a team of talented coders is everything one would need. When it comes to delivering the AI-models, that is apparently not enough.

The bitter lesson of ML’s development

OpenAI’s transition could be seen as a single event concerning only one organization. It could be, that is, if it wasn’t discussed by the godfathers of modern machine learning.

Related:  With your head in the clouds - how to harness the power of artificial intelligence

Richard Sutton is one of the most renowned and influential researchers of reinforcement learning. In a recent essay, he remarked that most advances in AI development are powered by access to computing power, while the importance of expert knowledge and creative input from human researchers is losing significance.

Moreover, numerous attempts have been made to enrich machine learning with expert knowledge. Usually, the efforts were short-term gains with no bigger significance when seen in the broader context of AI’s development.

Why does it matter?

The opinion would seem to support the general observation that computing power is the gateway to pushing the boundaries of machine learning and artificial intelligence. That power combined with relatively simple machine learning techniques frequently challenges the established ways of solving problems. The RL-based agents playing GO, chess or Starcraft are only top-of-mind examples.

Yann LeCun, Geoffrey Hinton and Yoshua Bengio awarded the Turing Award

The Association for Computing Machinery, the world’s largest organization of computing professionals, announced that this year’s Turing Award went to three researchers for their work on advancing and popularizing neural networks. Currently, the researchers split their time between academia and the private sector, with Yann LeCun being employed by Facebook and New York University, Geoffrey Hinton working for Google and the University of Toronto, and Yoshua Bengio splitting his time between the University of Montreal and his company Element AI.

Why does it matter?

Named after Alan Turing, a giants of mathematics and the godfather of modern computer science, the Turing Award has been called IT’s Nobel Prize. While the lack of such a prize in IT is obvious — IT specialists get the Turing Award.

Nvidia creates a wonder brush – AI that turns a doodle into a landscape

Nvidia has shown how an AI-powered editor swiftly transforms simple, childlike images into near-photorealistic landscapes. While the technology isn’t exactly new, this time the form is interesting. It uses Generative Adversarial Networks and amazes with the details it can muster – if the person drawing adds a lake near a tree, the water will reflect it.

Why does it matter?

Nvidia does a great job in spreading the knowledge about machine learning. Further applications in image editing will no doubt be forthcoming, automating the work of illustrators and graphic designers. But for now, it is amazing to behold.

So do you think like a computer?

While machine learning models are superhumanly effective in image recognition, if they fail, their predictions are usually at least surprising. Until recently, it was believed that people are unable to predict how a computer will interpret an image when not in the right way. Moreover, the totally inhuman way of recognizing the image is prone to mistakes – it is possible to prepare an artificial image that can effectively fool the AI behind the image recognition and, for example, convince the model that a car is in fact a bush.

Related:  AI Monthly digest #5 - AlphaStar beats human champions, robots learn to grasp and a Finnish way to make AI a commodity

The confusion about the machines identifying objects usually comes from the fact that most AI models are narrow AI. The systems are designed to work in a closed environment and solve a narrow problem, like identifying cars or animals. Consequently, the machine has a narrow catalog of entities to name.

To check if humans are able to understand how the machine is making its mistakes, the researchers provided volunteers with images that had already fooled AI models together with the names the machines were able to choose from for those images. In those conditions, people provided the same answers as the machines 75% of the time.

Why does it matter?

A recent study from John Hopkins University shows that computers become increasingly human even in their mistakes and that surprising outcomes are the consequence of extreme narrowness of the artificial mind. A typical preschooler has an incomparably larger vocabulary and amount of experience collected than even the most powerful neural network, so the likelihood of a human finding a more accurate association for the image are many times larger.

Again, the versatility and flexibility of the human mind is the key to its superiority.

Ai-ethics-artificial-imagination

AI Monthly Digest #6 – AI ethics and artificial imagination

/in , /by and

February came with the groundbreaking event of building a new state-of-the-art natural language processing neural network. And then the unthinkable happened.

Machine learning has already shown that it is a world-transforming technology used in various appliances, including drug discovery, saving endangered species and designing software for sophisticated prosthetic legs. Therefore it is natural that this should spark debates, both ethical and business-related. What reaction has there been to this in February?

1. Open-AI designed a new gold standard for natural language processing…

While in image processing artificial intelligence models outperform humans in most tasks, natural language processing (NLP) is still a challenge for machines to master. Although AI-based services are already usable with Google Translate being the pinnacle of today’s achievements, the texts produced by machines is easy to recognize by humans. But this could change soon with the latest state-of-the-art from OpenAI.

They designed a model called GPT-2 and prepared a dataset containing 8 million web pages for training. The network objective is to predict the next word given all the words to that point, which is the simplest way of doing unsupervised learning in NLP. The main improvements, in this case, are scaling up the model size (1.5B parameters) and training it on the gargantuan database of Internet text (40GB) at an unprecedented scale (32 x TPUv3).

The resulting network analyzes the starting word of a sentence and then it adds the next words to create a text based on “the most probable output”. The effect is surprisingly accurate. The model is able to recognize the type of the starting text. If the first sentence is a press title it will produce a legitimate-sounding short news article. It is also good at adopting various styles – something unimaginable for previously sturdy models which could produce problematic sentences.

Example:

Human written:

In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.

Model completion (machine-written, 10 TRIES)

The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science.

Now, after almost two centuries, the mystery of what sparked this odd phenomenon is finally solved.

Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans. Pérez noticed that the valley had what appeared to be a natural fountain, surrounded by two peaks of rock and silver snow.

Pérez and the others then ventured further into the valley. “By the time we reached the top of one peak, the water looked blue, with some crystals on top,” said Pérez.

Pérez and his friends were astonished to see the unicorn herd. These creatures could be seen from the air without having to move too much to see them – they were so close they could touch their horns.

While examining these bizarre creatures the scientists discovered that the creatures also spoke some fairly regular English. Pérez stated, “We can see, for example, that they have a common ‘language,’ something like a dialect or dialectic.”

More examples of AI-generated texts may be found at Open-AI blog.

Moreover – the network is capable of transfer learning – it is easier to perform a slight additional training to enable the model to perform more sophisticated tasks. Transfer learning is a gateway technology for modern image recognition appliances, including art recognition and visual quality control.

Why does it matter

Alan Turing, the godfather of modern information technology, was heavily convinced that the ability to understand language is a key indicator of intelligence. Natural language, the emotions hidden beneath the words and all the cultural and societal contexts behind sentences are truly unique for human beings.

Related:  AI Monthly digest #5 - AlphaStar beats human champions, robots learn to grasp and a Finnish way to make AI a commodity

Making this part of the world approachable for machines is undoubtedly the great breakthrough and passes countless possibilities, from battling hate crime to various business appliances.

2. … and they didn’t release it to the public

With great power comes great responsibility and the events from February 2019 show it clearly. Despite best practices and established custom, Open-AI decided NOT TO release the state-of-the-art neural network to the public as an open source software.

The organization did this to prevent the neural network from being used for malicious purposes, with perpetuating fake news as a top concern. The decision sparked a debate on Reddit, Twitter and other various media, where participants argued on the safety of releasing this “potentially dangerous technology” and tackling the established model of open-sourcing the effect of research. Researchers are afraid of building the doorway for “deepfakes for texts”.

Why does it matter

AI ethics and responsibility was highlighted as one of the major AI trends 2019. The situation where researchers choose safety over progress is unusual. On the other hand, fake news is considered one of the major threats of the future and was recognized one of the most dangerous online activities in 2018 by the World Economic Forum.

With the rise of autonomous cars and IoT revolution building a frame for sophisticated AI-powered solutions to work, such dilemmas may get increasingly common.

3. China’s AI strategy at a glance (or even more)

In the last AI monthly digest, we covered the AI strategy forged in Finland, where the national pursuit for the AI appliances grew from a grassroots movement. In contrast, China has a centralized strategy to make the country AI-giant.

The Center for a New American Security (CNAS) has published a report in which they provide deeper insights into understanding China’s AI strategy. The report is a long read written by Gregory C. Allen who had a chance to meet a few times with high-ranking Chinese officials on conferences focusing on Artificial Intelligence.

Topics covered in the text:

  • Chinese views on importance and security of AI
  • Strengths and weaknesses of China’s AI Ecosystem
  • China’s short-term goals in AI
  • The role of semiconductors

The author concludes that gaining expertise in and understanding of AI developments in China should help U.S. policymakers sort out their priorities. Instead of influencing China’s competitiveness, they should focus on boosting the technological and economic competitiveness of the United States.

Why does it matter

Early adopters are usually those who earn knowledge with first failures. Thus, for now, it is wise to observe and learn.

4. Introducing Paperswithcode

Staying updated on the latest developments in machine learning is not an easy task, especially considering the fact that not all the researchers decide to make their code available. Releasing the code is a good practice, but providing only the paper that “describes the matter enough” to reproduce the effect is also correct.

Related:  AI Monthly digest #4 - artificial intelligence and music, a new GAN standard and fighting depression

Using the paperswithcode makes the process of paper-and-code matching much easier, effectively saving a lot of time for people willing to broaden their knowledge.

Why does it matter

The portal itself is a handy tool to gain knowledge and, a bit unwillingly promotes the constituted approach of making the research code public. If it gains notoriety, it may be a tool of social pressure to publish code with the research.

5. Popularity of thispersondoesnotexist.com

The neural networks’ ability to create convincing human faces is one of the benchmarks for modern AI appliances. To make the images available for the public, the StyleGan-based model publishes random face on thispersondoesnotexist.com website.

As in every aspect of life, the devil is in the details. The website is a great tool to spot the weakest parts of generated faces.

  • Hair – it is common to place some hair in the air, with no attachment to other or floating in an unnatural way. Sometimes the skin around hair is “sticking” to them, creating some kind of scars that are disturbing when seen. The model sometimes messes up hairstyles, mixing dreadlocks with straight and curly hair, but it is more challenging to spot
  • Teeth – algorithms tend to mess up with teeth, placing them unnaturally, sometimes merged or blurred.
  • Glasses – it is common that the eyes inside the glasses do not fit the rest of the depicted person. Sometimes it means placing eyes of an old woman, surrounded with wrinkles, on a child’s face.

Why does it matter

Making these images publicly available with commentary, which all have been made by neural networks, is an interesting way to spread knowledge about AI-related topics. The concerns about deep fakes and other illicit content created with the support of artificial intelligence are mostly based on lack of knowledge – the AI-based techniques are seen as omnipotent. But they aren’t, as can be easily seen by the example of teeth and hair.

Related:  AI Monthly digest #3 - artificial intelligence in science getting big

Moreover, the initiative is another way of encouraging people to verify news sources and distrusting information and tackling the challenge of fake news and deep fakes, mentioned above.

And a bonus one:

deepsense.ai in cooperation with Google Brain has finished research on enabling reinforcement learning (RL) agents to make predictions about their upcoming actions to reduce the number of actions needed to train skills.

RL agents typically need incomparably more actions to train skill than humans and the ability to build the skill “in a head” may be a part of the answer. Starting from childhood, people repeat their actions in their minds when they seek perfection or just recall an enjoyable or pleasant time.

Following this idea, our researchers have designed neural networks that possess the imagination that enables the model to simulate the action before executing it. A more detailed story about these projects and their outcome may be found in our blogpost about artificial imagination and Arxiv paper.

Why does it matter?

Artificial imagination is basically the idea of building the world simulated only by the mind’s power. Our researchers were able not only to reduce training time, but also to replace the simulated environment with another neural network’s imagination.

Maintaining and providing the simulated environment is one of the highest costs in building reinforcement learning agents, so tackling this challenge may be a great step to making this technique more popular.

AI Monthly digest #5 – AlphaStar beats human champions, robots learn to grasp and a Finnish way to make AI a commodity

AI Monthly digest #5 – AlphaStar beats human champions, robots learn to grasp and a Finnish way to make AI a commodity

/in , , /by and

With a ground-breaking AlphaStar performance, January kickstarted 2019 AI-related research and activities.

The January edition of AI Monthly digest brings a marvel to behold–AlphaStar beating human champions at the real-time strategy game StarCraft. To find out why that’s important, read the full story.

AlphaStar beats top world StarCraft II pro players

AlphaStar, a DeepMind-created agent, beat two world-famous professional players in StarCraft II. The agent was using a Protoss race and playing against another Protoss. The game itself can also play Zerg and Terrans, but AlphaStar is trained only to play Protoss vs. Protoss matches.
The machine defeated Dario “TLO” Wunsch, a Zerg specialist playing Protoss for the occasion, 5-0 in the first five-match round. It then made quick work of professional Protoss player Grzegorz “MaNa” Komnicz beating the champion 5:0.
A noticeable advantage AlphaStar had against both players was its access to the entire StarCraft II map at once. It is still obscured by the fog of war, but the agent doesn’t have to mimic the human’s camera moves. DeepMind prepared one other agent to address this issue and plays using a camera interface, but it lost to MaNa 0:1.
To make the matches fair, the DeepMind team reduced the Actions Per Minute (APM) ratio to a human level and ensured the machine had no advantage in terms of reaction time. Nonetheless, it was clear at crucial moments that AlphaStar had bursts of APM far above human abilities. DeepMind is aware of this and will probably do something about it in the future. For now, however, we will be content to focus on what we have seen.

How the matches went

Unlike human players, AlphaStar had employed some unorthodox yet not necessarily wrong strategies – avoiding walling the entrance to the base with buildings was the most conspicuous one. What’s more, the model used significantly more harvesting drones than pro players normally use.

Beyond its superiority in micromanagement (the art of managing a single unit and using its abilities on the battlefield), the agent didn’t display any clearly non-human strategies or abilities. However, AlphaStar was seen at its finest when it managed to win the match by managing a large number of Stalkers, the unit that is normally countered by Immortals in a rock-paper-scissors manner. As MaNa, the human player confronting the agent, noted, he had never encountered a player with such abilities. As such, the gameplay was clearly on a superhuman level, especially considering the fact that MaNa executed the counter-tactic, which failed due to AlphaStar’s superior micromanagement.

How the Deepmind team did it

The initial process of training the agent took ten days – three of supervised learning built on the basis of replays of top StarCraft II players. The team then infused the agent with reinforcement learning abilities (an approach similar to our team’s cracking Montezuma’s Revenge) and created the “AlphaStar” league to build multiple agents competing against each other. league witnessed a similar cycle with some strategies emerging and being later countered.

Related:  Five top artificial intelligence (AI) trends for 2019

After that, the team selected five agents for a match with TLO. To further polish their skills, the agents were trained for another week before the match with the MaNa. As a Protoss specialist, MaNa posed a greater challenge than TLO, a Zerg-oriented player who was learning Protoss tactics only to square off against AlphaStar.
Courtesy of Blizzard, the developer of StarCraft II, Deepmind was delivered a significantly faster version of StarCraft II. This version enabled each agent in AlphaStar league to experience up to 200 years of real-time gameplay in just two weeks.

Why it matters

The AI community has grown accustomed to witnessing agents cracking Atari Classics and popular board games like chess or Go. Both environments provide a diverse set of challenges, with chess being a long-term fully observable strategy game and Atari delivering real-time experience with limited data.
StarCraft combines all manner of challenge by forcing players to follow the long-term strategy without knowledge of an opponent’s strategy and movement until it is in the line of sight of its own units (normally the battlefield is covered by the “fog of war”). Each encounter may show that a strategy needs to be fixed or adapted, as many units and strategies tend to work in a rock-paper-scissors manner, enabling players to play in a tactic-counter-tactic circle. Problem-solving in real time while sticking to a long-term strategy, constantly adapting to a changing environment and optimizing one’s efforts are all skills that can be later extrapolated to solve more challenging real-world problems.

Related:  AI Monthly digest #3 - artificial intelligence in science getting big

Thus, while playing computer games is fun, the research they enable is very serious. It also lays bare the contrast between human and machine abilities. The computer was able to beat a human player after about four hundred years of constant playing. The human expert, meanwhile, was twenty-five years old, had started playing StarCraft at the age of six and had to sleep or go to school while not playing StarCraft.
Nevertheless, impressive and inspiring.

Understanding the biological brain using a digital one

According to Brain Injury Alliance Wisconsin, approximately 10% of individuals are touched by brain injuries and 5.3 million Americans (a little more than 2% of the US population) live with the effects of a brain injury. Every 23 seconds someone suffers a brain injury in the US.
Such injuries add up to $76.5 billion in annual costs once treatment, transportation and the range of indirect costs like lost productivity are considered.
While brain trauma may sometimes be responsible for the loss of speech, strokes and motor neurone disease are also to blame. Although patients lose the ability to communicate, they often remain conscious. Stephen Hawking is perhaps the most famous such person. Hawking used a speech generator, which he controlled with the muscles in his cheek. The generators can also be controlled with the eyes.
Applying neural networks to interpret the signals within the brain enabled the scientists to reconstruct speech. Summarizing the efforts, Science magazine points out that the effects are more than promising.
Alzheimer’s disease is another challenge that may be tackled with neural networks. There are no medications that heal the disease, but applying the treatment early enough makes it manageable. With Alzheimer’s, the earlier the diagnosis is made, the more effective the treatment will be. The challenge is in the diagnosis, which often comes too late for the disease to be reversible.

Related:  Powering up demand forecasting with machine learning

By feeding the neural networks with glucose PET scans, researchers from the University of California delivered a system that can diagnose the early symptoms of Alzheimer’s disease up to six years earlier than doctors do.

Why it matters

The human brain is one of the most complex devices in the universe, so understanding how it works is obviously a great challenge. Applying neural networks to treat brain-related diseases may come with a bit of irony – we need an outer, artificial brain to outthink the way our own is working.

Democratizing the AI – the Finnish way

Machine learning and artificial intelligence, in general, tend to be depicted as a black box, with no way to get to know “what the machine is thinking”. At the same time, it is often shown as a miraculous problem-solver, pulling working solutions out of seemingly nothing like a magician procuring a rabbit from a hat. But this too is a misconception.
Like every tool before it, neural networks need to be understood if they are to yield the most valuable outcomes. That’s one reason Finland aims to train its population in AI techniques and machine learning. Starting with 1% of its population (or roughly 55,000 people), the country aims to boost its society and economy by being a leader in the practical application of AI.
Initially a grassroots movement, the initiative gained the support of the government and Finland’s largest employers.

Why it matters

The biggest barrier in using AI and machine learning-powered techniques is uncertainty and doubt. Considering that people are afraid of things they don’t understand, spreading the knowledge about machine learning will support the adoption and reduce societal reluctance to adapting these tools. Moreover, understanding the mechanisms powering ML-based tools will give users a greater understanding of just what the tools are and are not capable of.

New state-of-the-art in robotic grasping

The issues Artificial Intelligence prompts frequently ignite philosophical debate and add interesting insight and inspiration. This recent paper on robot grasping is short of neither insights nor inspiration.
[bctt tweet=”The idea behind the use of reinforcement learning to control robotic arms is simple – hard-coding all the possible situations the robot may encounter is virtually impossible, but building a policy to follow is much easier” via=”no”]
What’s more, building the controller for the robotic arm requires the mountains of data coming from the sensors to be cross-combined. Every change – be it lighting, color or position of an object — can confuse the controller and result in failure.
Thus, the research team built a neural network that processes the input into the “canonical” version, stripped of the insignificant details like shades or graphical patterns – so that grasping is the only thing that matters. Ushering in a new state of the art in robotic grasping, the results are impressive.

Why do the results matter?

There are two reasons these results are important. First, building the controllers of robotic arms is now simpler. Robots that can move in non-supervised, non-hardcoded ways and grasp objects will be used in astonishing ways to improve human lives–for example, as assistants for the disabled or by augmenting the human workforce in manufacturing.
The second breakthrough is how researchers achieved their improvements. Instead of building more powerful neural networks to improve the input processing, the researchers downgraded the data into a homogenous, simplified “canonical” version of the reality. It seems that when it comes to robotic perception, Immanuel Kant was right. There are “things that exist independently of the senses or perception”, but they are unknowable–at least for a robotic observer. Only operating within a simplified reality enables the robot to perform the task.

Keep informed on state-of-the-art machine learning

With the rapidly changing ML landscape, it is easy to lose track of the latest developments. A lecture given by MIT researcher Lex Fridman is a good way to start. The video can be seen here:

Read previous editions of AI Monthly digest:

AI Monthly digest #4 - artificial intelligence and music, a new GAN standard and fighting depression

AI Monthly digest #4 – artificial intelligence and music, a new GAN standard and fighting depression

/in , , /by and

December brought both year-end summaries and impressive research work pushing the boundaries of AI.

Contrary to many hi-tech trends, artificial intelligence is a popular term in society and there are many (not always accurate) views on the topic. The December edition of AI Monthly digest both shares the news about the latest developments and addresses doubts about this area of technology.
Previous editions of AI Monthly digest:

Tuning up artificial intelligence and music

Machines are getting better at image recognition and natural language processing. But several fields remain unpolished, leaving considerable room for improvement. One of these fields is music theory and analytics.
Music has different timescales, as some parts repeat at scale of seconds while others extend throughout an entire composition and sometimes beyond. Moreover, music composition employs a great deal of repetition.

Google’s Magenta-designed model leverages the relative attention to spot how far two tokens (motifs) are, and produced convincing and quite relaxing pieces of piano music. The music it generated generally evokes Bach more than Bowie, though.
Researchers provided samples of both great and flawed performances. Although AI-composed samples still refer to classical music, there are more jazz-styled improvisations.

Related:  Don’t waste the power. AI supply chain management

Further information about studies on artificial music can be found on Magenta’s Music Transformer website.
Why does it matter?
Music and sound is another type of data machine learning can analyze. Pushing research on music further will deepen our knowledge of music and styles in a way that AI made Kurt Vonnegut’s dream of analyzing literature a reality.
Furthermore, the music industry may be the next to leverage data and the knowledge and talents of computer scientists. Apart from tuning the recommendation engines for streaming services, they may contribute more to the creation of music. A keyboard is after all a musical instrument.

2. The machine style manipulator

Generating fake images from real ones is a thing. Generative Adversarial Networks enhance their training abilities by analyzing real images, generating fake ones and then training to be as good as possible in determining which is real and what is shown in the images.

The challenge neural networks and those who design them fact is in producing convincing images of people, cars, or anything else that is to be recognized by the networks. In a recent research paper, the group behind the “one hour of fake celebrity faces” project introduced a new neural network architecture that separates high-level attributes and stochastic variations. In the case of human images, the high-level attribute may be a pose while freckles or the hairdo are stochastic variations.
In a recent video, researchers show the results of applying a style-based generator and manipulating styles later to produce different types of images.

The results are impressive – researchers were able to produce convincing images of people from various ethnic backgrounds. By controlling different levels of styles, researchers were able to tune up everything on image – from gender and ethnicity to the shape of glasses worn.
Why does it matter?
That’s basically the next state-of-the-art in GAN networks, the best-performing image recognition technology used. Producing fancy-looking fake images of faces and houses can significantly improve the performance of image recognition models. Ultimately, however, this technology may be a life saver, especially when applied in medical diagnosis, for example in diabetic retinopathy.

3. AI failures – errare (not only) humanum est

2018 saw significant improvements in machine learning techniques and artificial intelligence proved even further how useful it is. However, using it in day-to-day human life will not be without its challenges.

In his 1896 novel “An Outcast of the Islands”, Joseph Conrad wrote “It’s only those who do nothing that make no mistakes”. This idea can also be applied to theoretically mistake-proof machines. Apart from inspiring successes, 2018 also witnessed some significant machine learning failures:

  • Amazon’s gender-biased AI recruiter – the machine learning model designed to pre-process the resumes sent to the tech giant overlooked female engineers due to bias in the dataset. The reason was obvious – the tech industry is male-dominated. As algorithms have neither common sense nor social skills, it assumed that women are just not a good match for the tech positions the company was trying to fill. Amazon ditched the flawed recruiting tool, yet the questions about hidden bias in datasets remain.
  • Uber’s fatal autonomous car crash – the story of fatal crash is a bitter lesson for all autonomous car manufacturers. Uber’s system not only detected the pedestrian it hit while driving, but also autonomously decided to proceed and ignore warnings, killing 49-year old Elaine Herzberg.
  • World Cup predictions gone wrong – The World Cup gave us another bitter lesson, this time for predictive analytics. While the model built to predict brackets may have been sophisticated, it failed entirely. According to its predictions, Germany should have met Brazil in the finals. Instead, the German team didn’t manage to get out of its group while Brazil bent the knee before South Korea. The final came down to France versus Croatia, an unthinkable combination, both for machine learning and football enthusiasts around the world. The case was further described in our blogpost about failure in predictive analytics.

More examples of AI failures can be found in the Synced Review Medium blogpost.
Why does it matter?
Nobody’s perfect. Including machines. That’s why users and designers need to be conscious of the need to make machine learning models transparent. What’s more, it is the next voice to ensure that machine learning model results are validated – a step that is tempting to overlook for early adopters.

4. Smartphone-embedded AI may detect the first signs of depression

A group of researchers from Stanford University has trained a model with pictures and videos of people who are depressed and people who are not. The model analyzed all the signals the subjects sent, including tone of voice, facial expressions and general behaviour. These were observed during interviews conducted by an avatar controlled by a real physician. The model proved effective in detecting depression more than 80% of the time. The machine was able to recognize slight differences between people suffering from depression and people who were not.
Why does it matter?
According to WHO, depression is the leading cause of disability worldwide. If not cured, it can lead to suicide, the second most common cause of death among 15-29-year-olds.

Related:  Powering up demand forecasting with machine learning

One barrier to helping people suffering from depression is inaccurate assessment. There are regions in the world where less than 10% of people have access to proper treatment. What’s more, mental illness is often stigmatized, and treatment is both costly and hard to access. These factors, and the fact that early symptoms are easily overlooked, lead many patients to avoid looking for care and medical support.
The experiment is a step toward building an automated and affordable system for spotting signs of depression early on, when the chance for a cure is highest.

5. So just how can AI hurt us?

Machine learning is one of the most exciting technologies of the twenty-first century. But science fiction and common belief have provided no lack of doomsday scenarios of AI harming people or even taking over the world. Dispelling the myths and disinformation and providing knowledge should be a mission all AI-developing companies. If you’re new to the discussion, here’s an essay addressing the threat of AI.
Why does it matter?
Leaving the doubts unaddressed may result in bias and prejudice when making decisions, both business and private ones. The key to making the right decisions is to be informed on all aspects of the issue. Pretending that Stephen Hawking’s and Elon Musk’s warnings about the cataclysmic risks AI poses were pointless would indeed be unwise.
On the other hand, the essay addresses less radical fears about AI, like hidden bias in datasets leading to machine-powered discrimination or allowing AI to go unregulated.
That’s why the focus on machine morality and the transparency of machine learning models is so important and comes up so frequently in AI Monthly digest.

Summary

December is the time to go over the successes and failures of the past year, a fact that applies equally to the machine learning community. Facing both the failures and challenges provides an opportunity to address common issues and make the upcoming work more future-proof.

AI Monthly digest #3 - artificial intelligence in science getting big

AI Monthly digest #3 – artificial intelligence in science getting big

/in , /by and

November brought a lot of significant AI-related breaking news. Machine learning and deep learning models were folding proteins, deriving the laws of physics from fictional universes and mimicking the human brain in a way never before seen.

This edition of AI Monthly digest looks at scientific improvements made by AI with significant support from tech giants. Stories from November show that AI is clearly a great tool to improve countless lives – and the protein-folding model is the best example of solving problems thought to be unsolvable. Or at least since our universe came into existence.

1. Deepmind’s AlphaFold wins Protein-Folding Contest

It’s a bit ironic that the protein-based human brain designed silicon-based tools to better predict protein folding. That’s precisely the case with AlphaFold, the neural network designed by Deepmind to predict the shape of a protein from the DNA code.
The basic building blocks of all known life forms, proteins come in many forms, shapes and levels of complexity. Their shape and function are encoded within DNA. Yet the DNA itself contains only a sequence of amino acids that form the protein, and the way they will form a structure is not encoded in any way. Thus it is challenging to establish what the protein will look like. As Levintha’s Paradox states, it would take longer than the age of the universe to enumerate the possible configurations of a typical protein before reaching the right structure. Yet, the shape of protein is crucial in the treatment and diagnosis of numerous diseases, including Alzheimer’s, Parkinson’s and Hutchington’s.
Thanks to the relatively high amount of gene-related data, Deepmind was able to build a neural network that could predict the shape of proteins only from a DNA sequence. The algorithm analyzes the distance between particular amino acids and compares that with existing models, preparing estimations about possible shapes and folds.
The effects are seen as one of the most significant breakthroughs and an unprecedented progress on protein folding.
To read more about the model and how it works, see recent Deepmind’s blogpost.

2. Towards building an AI Physicist

A Physicist’s job is to create models of the universe that give us a better understanding of the reality surrounding us. Newton, Galileo, Archimedes and many others had to convert measurements and observations into the fundamental laws of the universe. The ability to discover important answers from data is the best proof of their genius.
Last month, MIT’s Tailin Wu and Max Tegmark presented work reporting on an “AI Physicist” deriving the laws of physics in artificial worlds. To crack the mysterious environments, the AI agent uses four strategies embedded in its architecture: divide-and-conquer, Occam’s razor, Unification and Lifelong Learning, which together allow it to discover and manipulate theories without supervision.
The system was able to produce correct theories for various simulated worlds created only to test its accuracy. The success is significant from a scientific point of view, as it proves that neural networks may be the tools to speed up physical science in more ways than we ever expected.
The Arxiv paper can be read here.

3. SpiNNaker – Human brain supercomputer runs for the first time

Building neural networks is about mimicking the human brain’s activity and processing abilities. Most common neural networks work by communicating in a constant way and exchanging information in a stream. The spiking neural network is more “biological” insofar as it works by forcing artificial neurons to communicate in spikes and exchange information in a “flash”. Instead of sending the information from point A to point B, it supports sending multiple bits of parallel information.

AI Monthly digest #3 - Artificial Intelligence in science getting big - Brain

The spiking neural network is more “biological” insofar as it works by forcing artificial neurons to communicate in spikes and exchange information in a “flash”

This system is mimicked in a SpiNNaker (Spiking Neural Network Machine) built at the University of Manchester’s School of Computer Science, backed by EPSRC and supported by the European Human Brain Project. The most significant outcome of the SpiNNaker is that it builds a working, small-scale model of the human brain, with the manifold scientific possibilities it brings with it.
More information may be read here.

4. New ImageNet state-of-the-art with GPipe

Since its foundations, the greatest challenge in computer science has been insufficient computing power. Multiplying the number of cores is one way to solve the problem, while optimizing the software is another. The challenge comes to bear in neural networks, as training a new one requires gargantuan computing power and no less time.
Thus, Google Brain’s GPipe is a significant improvement that makes neural networks more cost-effective. By using GPipe, neural networks can process significantly more parameters. And that leads to better results in training.
GPipe combines data parallelism and model parallelism, with a high level of automation and memory optimization. In the paper, researcher expanded the AmoebaNet from 155.3 million parameters to 557 million parameters and inserted as input 480×480 ImageNet Images. The result was an improvement in ImageNet Top-1 Accuracy (84.3% vs 83.5%) and Top-5 Accuracy (97.0% vs 96.5%), making the solution the new state-of-the-art.

5. Enterprise machine learning done right – Uber shows its best practices

With machine learning being a recent development, there are no proven and tested methodologies to build new AI-based developments nor best practices with which to dive in and use in development.
Uber’s recent blogpost shared an interesting vision of incorporating AI-driven culture into the company. Instead of building one large ML project to perform one task on an enterprise scale, Uber powers its teams with data scientists and looks for ways to automate their daily work. The company has now done a dozen projects, carried out by a variety of teams – from Uber Eats menu items ranking or marketplace forecasting to customer support.
On the back of this strategy, Uber has gone from a company that did not use machine learning at all to one now heavily infused with AI-based techniques. More details about ML applications in Uber can be seen in their blogpost.

Summary

AI, machine learning and deep learning tend to be seen as buzzwords, applied only by tech giants to solve their hermetic problems. Folding proteins, deriving the laws of physics and simulating the brain as it really works all show that AI is the rule-breaker and disruptor it any field it is applied in. Even if it was dominated by bold minds physicists of all time.

AI Monthly digest #2 - the fakeburger, BERT for NLP and machine morality

AI Monthly digest #2 – the fakeburger, BERT for NLP and machine morality

/in , /by and

Fake images of hamburgers, the autonomous trolley problem, Google’s BERT for NLP and more stories from October, curated by deepsense.ai’s team, right here in our AI Monthly Digest.

October brought important developments in machine learning and sparks for interesting discussions about machine morality. deepsense.ai’s Arkadiusz Nowaczyński and Konrad Budek chose the five stories below.

1. Enter the fakeburger – DeepMind managed to produce convincing images of hamburgers, animals and landscapes

Renowned AI company DeepMind produced synthetic photos of hamburgers, landscapes and animals out of the ImageNet dataset. In most cases, as a team, it was difficult for us to determine if the pictures depicted a real or fake burger.
AI Monthly digest #2 – the fakeburger, BERT for NLP and machine morality - Fakeburgers
This was not the first time a neural network was used to create a convincing fake photo – take a look at NVIDIA’s one hour of imaginary celebrities.

We now have the ability to produce realistic images after training on the ImageNet dataset famous for advancing state-of-the-art image classification. DeepMind researcher Andrew Brock achieved his breakthrough with highly tuned Generative Adversarial Networks (GAN). The GAN uses a generator that produces artificial samples of images, and a discriminator, which distinguishes between fake and a real-world examples. The GANs here are scaled up, leading to the impressive results. Larger networks and training batch sizes (=2048) vastly improve the quality over previous work. Google’s Tensor Processing Unit (TPU) made the training feasible on such a large scale.

AI Monthly digest #2 – the fakeburger, BERT for NLP and machine morality - Fake images

Animals, nature and food look convincing, but scenes involving humans aren’t there yet.

Even though the model isn’t perfect, this is a remarkable step towards generating realistic looking photos with neural networks. To read more about the BigGAN and creating images of fakeburgers, check out the Arxiv paper.

2. AI-generated portrait sold for $432,500 at auction

Contrary to the fakeburgers and fakedogs built by DeepMind, the de Bellamy family consists of people with AI-generated eerie faces that are clearly non-human. The entire de Bellamy family images was generated by Obvious, a group of French AI engineers and artists.
Images were produced using GAN (Generative Adversarial Network) in the same manner DeepMind used it to produce an image of a hamburger. The model was fed with 15,000 portraits from last 600 years and attempted to build new ones using the data.
It is easy to see that images have no match to the master’s paintings, even considering the variety of styles represented by the artificial artist. Nevertheless, the portrait of Edmond de Belamy was sold for nearly half million dollar.
AI Monthly digest #2 – the fakeburger, BERT for NLP and machine morality - AI-generated portrait sold for $432,500 at auction

3. The machine morality – biases and automated discrimination

The transparency of machine learning models and Artificial Intelligence again in October came to the fore, igniting discussion over the Internet. The first important issue was gender bias in an Amazon-developed Artificial Intelligence model for preprocessing resumes delivered to the company. Trained with 10 years’ worth of resumes, the system was unintentionally trained to choose male candidates, as tech is dominated by men. The company changed its model but as the MIT Technology Review states, the company is no longer certain about the system’s neutrality in other areas.
Given the towering AI adoption rate, discussion about machine ethics and transparency is necessary. According to data from Deloitte, the number of machine learning project implementations was twice higher in 2018 than in 2017, and is expected to quadruple from there by 2020. As machines themselves are incapable of being racist, misogynist or biased, creating proper datasets and designing the evaluation process to spot hidden bias is crucial to building models that best and most fairly serve companies and society alike.
The data science community has not left the problem unaddressed – Google AI launched a Kaggle competition aiming to design image recognition models that could perform well on photos taken in different geographical regions than those used for training. The competition is being held as a competition track at the NIPS 2018 Conference, which is better-attended than even ComicCon.
This all goes to show that experts have been right from the start — AI is yet another tool that needs to be constantly evaluated and developed if it is to achieve its goals.

4. And the machine morality again – autonomous trolley problem

The topic of machine bias is even more important considering the rise of autonomous cars. A global study shows that people from various social and cultural backgrounds differ in their perception of “a lesser evil” when it comes to the “trolley problem” in autonomous cars. In a nutshell, if a car has to choose whether to hit an elderly person or a child, whom should it choose? A group of people or a single person? A pregnant woman or a medical doctor?
The study shows that choices vary by country, and the differences are more than significant. When facing the extreme situation of a car accident, human drivers make their mind autonomously and are solely responsible for their choices. On the other hand, every autonomous car will carry the AI model its manufacturer has provided. Are the legal system and society ready to transfer the responsibility for driving and the choices it necessitates from the driver to a non-human machine? Should the AI model fit the culture it operates in or follow some other code? These questions have yet to be answered.

5. Google’s BERT for NLP – new state-of-the-art in language modeling

Natural Language Processing may enter a new era with Google’s Bidirectional Encoder Representation from Transformations (BERT).
For now, NLP practitioners continue to use pre-trained word embeddings as initialization or input features in custom architectures for specific tasks. BERT, a model that can be pre-trained on a large text corpus and then fine-tuned for various NLP downstream tasks, may change that. It might be similar to what we have seen in Computer Vision in the last couple of years, where fine-tuning models pre-trained on ImageNet has proved a great success.
BERT is a multi-layer bidirectional encoder taken from Transformer architecture, which was introduced in Attention Is All You Need. The pre-training procedure is entirely unsupervised and includes two objectives: filling random gaps in the input sequence and classifying whether two input sentences are actually two consecutive sentences cut out from the larger text. During the fine-tuning, predictions can be made for entire sequences or each input token separately.
The study shows that fine-tuning a pre-trained BERT model has set a new gold standard in 11 benchmark tasks.
Google has released an official implementation of BERT for NLP available on github.

And now for some bonus information:

Paul Romer, winner of this year’s Nobel Prize in Economics, is a 62-year old ex-World Bank chief economist, writer and Python programming language user. He is also a firm supporter of making research available and clear, so he shares his findings via Jupyter notebooks and makes data available for everyone to process and interpret.
His example perhaps shows that combining knowledge about economics and science with a proper toolset for one’s daily work can help lead to a rewarding career.