AI Monthly Digest 20 - TL;DR

AI Monthly Digest 20 – TL;DR

/in /by

With the world spinning faster every day and delivering an insane amount of news and information to process, the temptation to cherry-pick content to consume increases. Powering summarization with AI tools is among the news we’re covering in this edition of AI Monthly Digest.

Other stories include the proof of deep learning supremacy in the world of board games, biased machines and what comes after Atari and board games.

TL;DR not working (yet)

A research team from the Allen Institute for Artificial Intelligence and the University of Washington has created a scientific summarization engine. The key goal was to check if an AI-based solution could shorten a text and capture the most essential information from a scientific paper. Including the most important information while also providing enough context to make the piece comprehensible is a challenging task even for a human. That the text needs to be grammatically correct and as interesting as possible makes the challenge all the more difficult.

TL;DR – it didn’t work. Yet. For all the details, see the Arxiv paper.

Why it matters

While summarization engines aren’t yet entirely feasible, polish applied over the next few years will lead them to eventually shine.

According to the research done at the University of Ottawa, the number of scientific papers published since 1665 surpassed the 50 million mark in 2009. Add to that the approximately 2.5 million new papers published every year, and you quickly see that staying up to date in any field, let alone many, is only getting harder.

Also, we are beset by news on all sides today. To wade through the overload and get results, people need to manage their attention carefully. That’s why a good, automated summarization engine can be a service we increasingly need not only in science but also our daily lives.

Tackling biased language models

“Thrash in, thrash out,” data scientists will say from time to time, alluding to the quality of an AI model or prediction, which can only be as good as the dataset the model was provided with.

A dataset can be full of hidden or obscure misinformation that a human supervisor usually misses or considers irrelevant given the knowledge they have already gathered. A good example of a model gone bad is the one Amazon AI used to analyze applications for engineering roles, which turned out to be biased against female engineers. Analyzing the profiles of the then current engineers, the model identified significantly fewer females among them, and concluded that women can’t code. For a human, this is obvious nonsense, but with AI there is no such thing as common sense.

Considering that, it is not surprising that research is being done to tackle bias in Natural Language Processing. A consortium from MIT, Intel, and the Montreal Institute for Learning Algorithms (MILA) has come up with a way to evaluate NLP models in a more disciplined and structured way. Their StereoSet is a dataset that measures typical biases in English to check if a model is truly neutral toward its subject. The research is available on Arxiv.

Why it matters

Apart from the fact that any discrimination is evil, a biased model comes with multiple disadvantages. Coming back to the example of Amazon, nobody knows how many talented female coders the company’s model caused to be dismissed. Biases in language processing hurt society, business, and companies in multiple ways, and reducing these biases is one of the most significant challenges to overcome in the near future.

AI playing GameBoy games

An iconic gadget of the late twentieth century, Game Boy and Game Boy Color made it into the palms and pockets of a lot of kids worldwide–118 million, to be exact. The platform handled multiple types of portable games, including simple platform games like Super Mario and more sophisticated RPG-like experiences like the Pokémon series.

Long story short, the platform delivers significantly more advanced games while keeping the graphic environment simple. Today, this is a good training sandbox for reinforcement learning models.

Why it matters

Reinforcement learning shines in multiple classes of problems, especially when there is no straightforward  way to solve them. But the point is in delivering an environment where the model can encounter various challenges while keeping the simulator relatively lightweight. Atari is currently a standard sandbox for testing various approaches, even as the needs and classes of problems to solve evolve.

Delivering a GameBoy simulator suitable for running AI experiments is an interesting way to broaden the research. The software is available on GitHub.

It’s official – deep learning supremacy in board games

The supremacy of deep learning in board games is now a fact. Leela Chess Zero has won the latest Top Chess Engine Championship, formerly known as the Thoresen Chess Engines Competition. The tournament has been run since 2010 and is considered an unofficial computer chess championship.

The winner of the 2019 edition was Stockfish, an open-source chess engine that is not powered by ML-based techniques and utilizes more traditional ways of playing chess. Here, “more traditional” doesn’t mean ineffective – no human has beaten Stockfish.

After just ten months of development with deep learning, LCZ  beat Stockfish twice – in the first half of 2019, and again this year. Stockfish was initially published in 2008, giving it 12 years of constant development, and in 2019 it remained the champ, but only barely, squeaking out a 50.5 to 49.5 advantage. During the recent faceoff, however, Leela Chess Zero defeated Stockfish 52.5-47.5 in 100 matches.

To give some scale, a typical non-professional player has a chess ranking (so called ELO rating system) of 1000 points. A talented non-professional can reach 2000 points. The current world champion, Magnus Carlsen, has about 2900 points.

Leela and Stockfish, meanwhile, are both over 3800 – far beyond human reach.

Why it matters

First of all, this isn’t only about chess. Apart from some hermetic insight on particular moves or ways to overcome interesting chess situations, most of the moves are incomprehensible for lay observers. Appreciating the strategy and tactics behind each move requires a rare level of chess mastery.

So in fact, the match was between two approaches to solving problems in computer science– traditional coding versus deep learning, the latter of which is superior in playing chess.

AI Monthly Digest #19 - when the Latin alphabet is not enough

AI Monthly Digest #19 – when the Latin alphabet is not enough

/in /by

March brought interesting publications about the limited ability of machines to read Bengali script, the challenges of generalization for machines and new materials discovered by neural networks.

These issues are connected by one crucial aspect – they are challenges for today’s AI, and come with massive consequences, from limiting the survivability of language to saving human lives and pushing progress forward.

Pushing to make Bengali script comprehensible for machines

AI-based tools not only reduce the number of mundane and boring tasks that fall to humans in daily business operations but also help to categorize documents and digitize what hasn’t yet been automated. Optical Character Recognition (OCR) recognizes printed text well when it’s done in the Latin alphabet, but it apparently fares far worse in reading the Bengali alphabet, which has 50 letters, 11 vowels and 39 consonants and is the sixth most used writing system in the world.

To tackle this challenge, a group of researchers from United International University in Bangladesh cataloged the challenges OCR faces in reading Bengali.

Why it matters

OCR technology makes digitizing printed materials–books, newspaper and old manuscripts alike–easier. It is a self-feeding mechanism, where the amount of digital texts serves as the basis for training natural language processing models. With those, AI can provide more support in performing duties in a language, so digitizing more texts widens the world of possibilities a language offers.

Also, when a language is harder for AI solutions to understand, it is less likely to be applied in new solutions, thus making solutions designed to work for that language ineffective, be that a recommender engine or other tool.

In the long term, this could result in a lower number of speakers of the language, and thus throw the language into decline, and with it that little bit of the world’s cultural heritage the language represents. So in fact, working to make the language comprehensible for machines can be seen as a struggle to keep it alive in the future.

Machine learning tackles antibiotic resistance

Antibiotics are among the most significant achievements in medicine, saving the lives of thousands of people around the world and effectively ending the era of widespread death due to common infections. Apparently that era is on its way back.

According to a WHO report on antimicrobial resistance, drug-resistant diseases already cause at least 700,000 deaths globally a year, including 230,000 deaths from multidrug-resistant tuberculosis. In the most pessimistic scenario, in the absence of action, that figure could increase to 10 million deaths globally per year by 2050.

This information is even more alarming in the light of the COVID-19 outbreak, a bacterial infraction after viral infections (especially respiratory system-effecting ones) are increasing the death toll and make the overall disease more severe.

Multiple new chemicals would be viable in antibiotic treatments, but they are the proverbial needle in a haystack. Harnessing the power of deep neural networks can make searching for them easier and faster by tossing out the limits of human perception.

An example of this approach can be found in this recent article in Cell magazine.

Why it matters

AI is a revolutionary technology that supports a massive variety of activities — in business by delivering better demand forecasting, in security with AIOps platforms and automated network traffic analytics and in manufacturing with quality control.

Using neural networks to improve healthcare is an effective way to make lives easier, increase longevity and enhance the quality of human life globally.

Using AI to deliver new materials for batteries

Harnessing neural networks in the search for new materials goes well beyond drug research. In fact, the need for new, more resilient, flexible and effective materials extends to nearly all industries. The revolution in IoT, wearables and smart appliances and electric cars is powering (pun unintended) the need for new materials for batteries.

To see how neural networks are being used to search for new materials for batteries, see this article in MIT News.

Why it matters

Human civilization is highly dependent on electricity and the ability to store it efficiently is crucial for our further development. Researching new batteries and materials to build them will (again – pun intended) power our further progress.

Can AI generalize? Apparently not

The ability to use a skill to do something new or in a new environment is a mark of intelligence, both natural or artificial. The concept of generalization is simple, yet there was no repeatable and reliable way to test if a system can do it.

This led a consortium of researchers from the University of Amsterdam, MIT, ICREA, Facebook, and NYU to produce gSCAN, a benchmark for generalization abilities. The tests they produced are simple, at least from a human point of view: drive in a direction never before taken (for example turning right when taught to always turn left).

Details about the benchmark can be found in this Arxiv paper.

Why it matters

The ability to generalize is the next step in the development of AI. The benchmark tests function similarly to the Atari games used in reinforcement learning – a technique limited enough to be applied swiftly, but applicable upon extrapolation in more sophisticated projects.

Also, the ability to generalize marks a limit of AI, which needs to be pushed forward to deliver new results. This benchmark is a way to do that, or at least test if it’s possible.

AI Monthly Digest #18 – the pixelated first step toward megastructures

AI Monthly Digest #18 – the pixelated first step toward megastructures

/in /by

As we predicted in our AI Trends 2020, NLP is the year’s leading trend. But research on self-regenerating structures has likewise been both surprising and fascinating. February was rich in news on self-regenerating machines and NLP-related events and research.

Microsoft trains T-NLG, a 17-billion-parameter language model

The tech behemoths are racing to build larger and more efficient language models, which are both costly and technologically challenging, if not daunting. Microsoft is a relatively new player in the game, and follows in the footsteps of Google with itsBERT model and OpenAI, which brought outGPT-2.

While Microsoft may not have been the first mover, its Turing Natural Language Generation (T-NLG) is now the largest model trained and published for language generation. It continues the line of work begun by BERT and GPT-2. Increasing neural network size once again proved successful. After pre-training on essentially all text from the Internet, T-NLG can be fine-tuned to solve various downstream tasks like answering questions, understanding documents or simply filling in as a conversational agent (like Meena from our last post). In order to achieve this new state of the art in large-scale deep learning, Microsoft came up with novel optimization techniques and tricks that make it feasible to train artificial neural networks with over 100 billion parameters.

To put this number into perspective: the human brain has around 100 trillion synapses (in a nutshell, their job is to connect neurons). That’s a thousand times more than 100 billion. Does this mean we can now train artificial neural networks that are 0.1% size of the human brain? Definitely not, biological neurons are much more complex than artificial ones and the recent discovery shows that a single biological neuron is able to compute XOR, a mathematical function that requires a multilayer artificial neural network to be computed.

Why it matters

With industry heavily focused and invested in NLP, the technology is one of deepsense.ai’s cutting edge trends in 2020. The new models being delivered support business and consumer services like chatbots and search engines.

Given these developments, any bigger and newer language model brings us closer to the day when natural communication between computers and humans a la Star Trek will be possible. In decidedly less galactic terms, it is also a convenient way to communicate with a computer when your hands are otherwise occupied – like when you’re cooking or driving a car. A simple “skip that song” makes a difference.

Growing Neural Cellular Automata

Apparently nature is the most brilliant engineer of all, with its unrivaled ability to reuse and repurpose building blocks to deliver a breathtaking set of incredibly diverse forms. With cells being the most basic building block of every creature alive, from the lowly germ to humans, it is both fascinating and inspiring to crack the process of making a single cell that not only functions but is specialized in a particular context.

Putting it simply – a single cell can be a separate organism or only a tiny part of a larger structure, designed to perform a specific role, without itself having the ability to survive. Apart from that, living organisms can regenerate themselves, sometimes regrowing lost limbs, sometimes just healing wounds. Cracking this process would deliver a significant boost for engineering.

What happened exactly?

Every living cell contains a DNA code that contains all the information about the organism in which it lives–its height, eye color, and other factors. In terms of programming, it can be seen as a list of instructions containing if-then statements, which tell the cell how it should operate in its given context and surroundings.

Research published on distill.pub tackles the matter by delivering images built with single “cells” that contain DNA-like information which, when rolled out using local updates, converges to a full image in global scale. Some parts of the image can be destroyed, and the image will rebuild itself with the lifelike process of spawning new cells to replace destroyed ones.

Also, more images can be spawned and built near the original one, effectively blending them to observe how cells interact with the context. In some cases, when blended and destroyed, the images rebuild in odd ways, delivering imperfect or twisted versions of previous projects. Scar tissue of some kind.

Why it matters

Currently, it is hard to come up with a practical application, but it brings science-fiction-like prospects that could result in unimaginable progress. Just to name a few:

Building megastructures – be it a Death Star, a Dyson Sphere or a planet-wide supercolony, making a centralized program to control the whole system might be challenging due to the very scale of that system and the latencies within it.

Also, when delivering a “smart gravel” that transforms into the desired type of material or form you need to build the structure, managing the construction or repairs would be much easier. A car wouldn’t have a spare tire and a toolbox, but a box containing smart-gravel (or any other currently non-existing smart material) infused with the DNA-like code of the car and used to repair what’s broken.

But that’s a distant future and practical appliances are yet to come – not unlike when neural networks and machine learning started from the creation of a computational model for neural networks in 1943. The technology is now transforming our daily lives in the form of machine learning, reinforcement learning or any other paradigm used to infuse our tools.

These lyrics don’t exist

The final news is more lifestyle related and lighter in mood than the rest. After the popularity of thispersondoesnotexist.com, the time has come for song lyrics. Pop on over to theselyriscsdontexist.com and see how you  can choose the mood, theme and musical style–then read the lyrics the AI spits out. Perhaps a happy heavy metal song about eating burritos would suit your fancy? No problem.  Or would a sad pop ballad about a steam locomotive be more to your liking? That’s no problem, either.

Why does it matter

It doesn’t, but data science is also about fun and neural networks can power both inspiring and powerful tech as well as silly toys. Because why not?

AI Monthly Digest #17 - a lovely chatbot

AI Monthly Digest #17 – a lovely chatbot

/in /by and

January Jump-started the year with news of the latest achievements in NLP, a breakthrough in neural networks for solving math problems, and AI creating artificial life.

Assuming that the world will continue changing as quickly as it is now, 2020 looks likely to bring unprecedented development to AI. Let’s take a closer look at all this exciting news.

Meena – a chatbot you would like to talk to

Chatbots are undoubtedly one of the hottest AI trends of the last few years (apart from the reinforcement learning of course), having been widely embraced by businesses seeking to streamline efficiency. Most applications are based on a trigger-and-response paradigm with little to no use of real natural language processing techniques. However, a real breakthrough may now have been made.

The GPT-2 saga was one of the most important and influential AI-related stories of 2019. That a model like GPT-2 produces convincing text, however, doesn’t necessarily spell the end of development in this area. In fact, Google joining the race in January proves the contrary.

Google Brain has created Meena, a 2.6-billion parameter end-to-end neural conversational model trained on 341 GB of text from the Internet. Meena is 1.7x larger than OpenAI’s GPT-2 and was trained on 8.5x more data for 30 days on a TPUv3 Pod (2048 TPU cores). To put it in a nutshell, it is larger and more efficient than any such model created to date, though for now, due to the potential risks that accompany releasing a demo, it remains unavailable to the public.

Why it matters

Meena is further proof that larger models, bigger datasets and superior computing power are driving progress in today’s mainstream AI research. A problem that challenges the developers in a way that has never been seen before – it is not about devising new concepts and testing them, but rather producing bigger datasets and neural networks.

A scalable pipeline for designing reconfigurable organisms

The dichotomy between machines and living organisms has long been absolute.  The key difference was in the material – machines are (were?) made of steel or plastic with no living tissue, at least not at a purpose.

Research presented in January, however, introduces new technology to the world: biological machines. They can be automatically designed and optimized in simulation for a desired locomotion behaviour and then manufactured and deployed in physical environments.

Josh Bongard, the lead researcher, said: “These are novel living machines. They’re neither a traditional robot nor a known species of animal. It’s a new class of artifact: a living, programmable organism.”

Watch the 11-min video covering the topic:

Why it matters

First and foremost, it is fascinating to watch the fantasy of science-fiction writers become reality.  The blending of organic matter and machines has long figured in fiction. Writers have seen the blending both as delivering conscious robots and living organisms made of metal (Necrons from Warhammer 40k are good example) or living organisms designed to fulfill the role of a machine (the living spaceships of Companions from Gene Roddenberry’s Earth: Final Conflict series are a good example).

Although this research shows only humble beginnings, it is fascinating indeed.

Solving maths with neural network

Math creates problems that are getting increasingly hard to tackle. And solving problems isn’t always about computing. Math requires thought, and that’s what machines are not good at, so the skyrocketing amount of computing power available has made little difference.

Artificial intelligence, of course, aims to change this.

Scientists from Facebook AI Research took a neural network originally designed for language modeling and machine translation and trained it to solve advanced mathematics equations, where the task is to predict symbolic integration of the input equation. It turned out that such neural translation was able to find correct answers more often than traditional software including Maple, Mathematica, and Matlab.

Solving equations requires symbolic reasoning, which is one of the hardest challenges for neural network-based systems. It has become an active area of research recently and we can expect more in the upcoming months. For example, a new paper from Deepmind MEMO: A Deep Network for Flexible Combination of Episodic Memories tackles the problem of reasoning over long distances.

Why it matters

Math problems are inherently abstract But solving them can fuel progress, even if it isn’t what we anticipate. Solving the Seven Bridges of Königsberg riddle was the first step toward graph theory and modern topology.

So automating the process is a first step toward speeding up the progress overall. And it is good news for us all.

AI Monthly Digest 16 - NeurIPS and AI Dungeon Master - Header

AI Monthly Digest #16 – NeurIPS and AI Dungeon Master

/in /by and

From the largest conference to a tiny-yet-significant project delivering an AI-powered role-playing game master, December delivered an interesting end to 2019.

The most important news in December was the NeurIPS conference, the most significant scientific event related to machine learning and neural networks.

NeurIPS 2019

Thanks to the Neural Information Processing Systems Conference, December is one of the most important months of the year for AI. The event draws together over 15,000 participants including scientists, researchers, business people and AI enthusiasts of all stripes. The 2019 event took place in Vancouver, BC from Sunday, December 8 to Saturday, December 14.

The conference is massive, with over a thousand papers accepted, multiple workshops addressing key trends and researchers from around the world exchanging knowledge and experiences.

Together with Volkswagen, deepsense.ai delivered a paper on training autonomous cars in a simulated environment and implementing the model in a real car. The research was a breakthrough and opens up interesting directions for further exploration. The paper was presented during a workshop at the conference.

Why it matters

NeurIPS is the largest scientific conference on machine learning and artificial intelligence, both cauldron and melting pot of new ideas and future directions for researchers and the industry as a whole.

Whether it’s reinforcement learning, demand forecasting done with the new heuristics or other ML-related solutions the market hasn’t heard of (yet), NeurIPS is where you’ll hear how ground is being broken.

AI Dungeon (Master)

With Dungeons and Dragons back in the game (pun intended), the institution of a game master (or, in this system, a Dungeon master) goes mainstream again. The phenomenon of modern role-playing pen-and-paper games is described in more detail in this New Yorker article.

For those unfamiliar with RPG nomenclature, a game master is the person who oversees the role-playing world, ultimately controlling everything apart from the players – the monsters, and non-player characters. In the pen-and-paper role-playing games, this role falls to one of the players (the “dungeon master”). In computer RPG, it’s all scripted by designers and delivered through the graphical environment.

Until now, that is. The AI Dungeon is a text-based role-playing game with AI filling the shoes of game master. The player can do basically anything – from asking questions (what is that hut made of?) performing actions (attack the ork warband) to living out far-fetched fantasies (I transform into a dragon and try to eat the moon). And the AI needs to respond.

Why it matters

Run by game masters and populated by the spawn of players’ imagination, classic role-playing games are made up of words. It is close to literature, yet more interactive, with the integration and interaction between the person behind the curtain of the world. Apart from imagination, nothing is required to play the game. Until recently, two people were needed to strike up a game – a game master and a player. Now, an AI-based game master sits ready to be used on the cell phone in your pocket or the browser on your desktop.

The experience may for now still pale in comparison to what an experienced game master will unfold for you. But it’s early days, and the technology appears to be heading in the right direction.

The top minds in AI predict what’s ahead for 2020

AI changes in a way that is nearly impossible to predict. The technology delivers significant breakthroughs and inspiring news nearly every day. But the need to make the reality less chaotic and reason at least something about the future is deep high. So VentureBeat has gathered some of the top minds in AI to tell us what may come up and go down, if you will, in the coming year.

The predictions vary from highly technical thoughts on the popularity of particular frameworks and solutions to a much broader view of the general development of machine learning.

Why it matters

It is easy to get lost in the riddled and constantly changing world of modern machine learning. VentureBeat’s guide provides some insight and delivers comments from people from top-notch AI companies.

Armed with that knowledge, one can easier filter out the noise and analyze the important trends.

AI Monthly Digest #15 - the end of two sagas

AI Monthly Digest #15 – the end of two sagas

/in /by and

November brought the release of the latest and largest version of the GPT-2 Natural Language Processing (NLP) model and the information that AlphaStar has reached the top level in Starcraft on Battle.net. There is also Lee Sedol’s bitterness to contemplate. The Go master recently retired.

This issue also chronicles a scam pulled off using a deep neural network. Researchers had warned us it would happen, and it finally has.

OpenAI releases its most recent iteration of GPT-2

The GPT-2 model was arguably the most significant Natural Language Processing news of 2019. The model’s performance made headlines, of course, but so too did OpenAI, the organization behind the model, opting not to make it public, at least not initially. And that sparked debate.

Those who opposed making the model public pointed out that a technology that produces such high-quality texts could be dangerous for society, and all the more so given the explosion of fake news in recent years. A model that can create texts that are nearly indistinguishable from human-written ones could support the creation of fake news in unprecedented ways.

On the other hand, this model represents strong progress for both NLP and text generation, and its ability to serve the public is precisely why, according to the second group, it should be made public.

The researchers malice-proofed their model, if you will, and released subsequent iterations step by step. A note on the matter can be read on the OpenAI’s website.

Why it matters

GPT-2 has been one of the dominant AI-related topics in 2019. Making it public finally enabled researchers to further expand its capabilities and companies to use it for their own benefit. How? According to The Verge, GPT-2 can write all sorts of texts – poems, stories, news, fake news and even code. It is undoubtedly a powerful tool that can support various goals, be they benevolent or malevolent.

Finally, the release brought the additional benefit of sparking open, public and debate on the ethics of publication of the AI-related research.

Google applies BERT to its search engine… and then Microsoft too!

When thinking about NLP applications, the top-of-mind usage is in improving search engines, be it an e-commerce one installed within an online shop or a general-purpose one like Google or Bing.

According to a recent Google blog post, up to 15% of everyday Google queries have never been made before. What’s more, search engine users have developed their own sub-language for building queries, which are usually based on keywords. Although effective to a degree, it is far from being a natural form of communication.

Search engines have struggled to process more complex sentence-length queries. So “good chocolate cake” is a better query than “could you please deliver me a recipe for an awesome chocolate cake, thank you in advance”. The latter could be completely misunderstood by the search engine.

BERT was added to Google Search in October. And November saw BERT installed in Bing, Microsoft’s search engine. According to the latest data from Statista, Bing commands 5.26% market share, second only to Google. Microsoft itself claims that Bing’s share is widely underestimated and reaches up to 33% on desktop.

No matter which version we take as truth, both search engines benefit from Google’s model.

Why it matters

Applying BERT to search engines delivers yet more proof that AI and machine learning improve our daily lives in unprecedented ways, even if it is often entirely “behind the scenes”. From now on, millions of users will use artificial intelligence solutions every hour of every day without even realizing it.

Last but not least, BERT will improve the way search engines work, further improving our lives.

The $243,000 deepfake scam

Experts have been warning us about deep fakes for a long time now. Indeed, such fakes allowed us to hear Barack Obama saying things he never actually said. Elsewhere, a renowned professor wrote of hearing his own voice uttering things he had never said.

But this time things got real. Using deep neural networks, cybercriminals managed to fake the voice of a company’s CEO, who needed to make an urgent transfer of $243,000 to a Hungarian supplier.

Why it matters

This is the first big scam pulled off using neural networks, and proves all too prescient the warnings that neural networks may be employed for less than noble ends. Here AI hasn’t been used to make the world safer, but to attack.

AlphaStar reaches grandmaster level

The AlphaStar saga is arguably the second most impressive story in the ML world of 2019. Using a reinforcement learning-trained agents, DeepMind achieved the level of master in the popular e-sport game StarCraft II. Deepmind’s first set of agents beat the top players TLO and MaNa. But controversy ensued when it was discovered that the Deepmind agent had access to the entire Starcraft map at once, while the human players were limited to frames seen on the screen.

The new set of agents managed to reach the top league during duels with users from all around the world on the online player-matching platform battle.net. The agent could have played with any race given – a Terran, Protoss or Zerg, against any other race. More impressive still, a single neural network handled all the matches, employing and executing different tactics suited for different races and opponents. While this may pale next to what a human player can do, the fact that neural networks traditionally handle single tasks makes this accomplishment stand out.

Why it matters

Reinforcement learning is a hot trend when it comes to delivering new, impressive results in machine learning. Training an agent that plays a popular computer game may sound like child’s play rather than serious scientific work. Yet the implications are serious – the agent acts in real time, with limited information, to solve complicated problems.

In the larger scheme of things, this is a next step in creating agents that can solve more complex (and more important) problems in real time. Building the next generation of autonomous cars is only a top-of-the-mind example.

The AlphaStar Saga also prompts us to contemplate both the future of AI and its present-day flaws.

After being defeated by AI in 2016, global Go champion Lee Sedol retired last month, claiming that nothing more can be accomplished in Go. “There is an entity that cannot be defeated,” he remarked in an interview with the Yonhap News Agency in Seoul.

“Entity” is an interesting word here. Human champions die or retire. AlphaGo will be there forever, an eternal champion to be unbeaten.

On the other hand, Sedol pointed out that he had taken a game from AlphaGo by exploiting a bug: when he made a totally unexpected move, the neural network got confused. Players who have done battle with AlphaStar frequently point out the model’s inflexibility. AlphaStar seems challenged to change its strategy once it has begun playing, allowing human players to plot their own strategies to exploit this shortcoming.

“The whole secret lies in confusing the enemy, so that he cannot fathom our real intent.”
– Sun Tzu

AI Monthly Digest #14 - quantum supremacy

AI Monthly Digest #14 – quantum supremacy

/in /by and

October brought amazing news for Google: the company has reached quantum supremacy. Today we’re going to look at just what that means.

November was no slouch, either, bringing us as it did greater clarity on the development of the most prominent frameworks as well as an interesting initiative to impose good practices in machine learning research.

To top off all the news, this issue of AI Monthly Digest reports on a funny incident related to robots. Have fun reading!

Google gains quantum supremacy

Quantum computing is not a new concept. It has been around for 30 years, even if it has spent most of that time as a concept with no practical applications. A situation comparable with the one of neural networks.

At some point, however, big data and cloud-based computing power changed the game for neural networks and artificial intelligence. Today, neural networks, the theoretical plaything of yesterday, stands behind demand forecasting systems, powers natural language processing tools and supports the daily operations of countless companies in multiple other ways with reinforcement learning on the cutting-edge.

Although a practical application of quantum machines has yet to come, Google’s recent experiment shows clear progress. Their quantum computer, operating on a 54-qubit processor called ‘Sycamore,” performed in 200 seconds computations that would take 10,000 years on the fastest supercomputer existing.

While the 10,000 years was later amended down to two days, the efficiency Google has managed to muscle into their system still boggles the mind.

Why does it matter

Or rather, why does it matter for machine learning? The answer is a bit tricky, but the question was bandied about at the most recent AI World Conference.

Quantum computing will enable a machine learning model to reflect and process more complex conditions and scenarios. According to the Boston Consulting Group, the optimal model for the upcoming future is to combine traditional and quantum computing.

Keras or PyTorch? Or maybe Tensorflow?

One of the most popular deepsense.ai blog posts examined the differences between Keras, the most popular library for Tensorflow, and PyTorch, one of the most popular deep learning frameworks.

According to the latest analysis done by The Gradient, the question of which is better is now irrelevant. Keras has recently been incorporated into TensorFlow as an official API to make the framework more convenient and easier to use. For those who love head-to-head battles for supremacy, the clash no longer pits Keras against PyTorch, but rather TensorFlow against PyTorch.

What’s even more interesting, each framework seems to push the other’s development. TensorFlow remains popular in the business community while scientist favor PyTorch.

Why does it matter

Knowledge of popular frameworks is crucial for data scientists as well as for business efficiency – choosing the right one can be essential to a project’s success.

In the end, the choice may end up being irrelevant as both deliver a comprehensive platform to deal with machine learning and facilitate the whole process. The frameworks are tools, after all, and should play a supporting role rather than a limiting one.

Robots meet reality

The film Robocop offered an iconic portrayal of machines being used in law enforcement. The dystopian future of grim Detroit is nothing like the streets of today’s Los Angeles, where the first police patrol robots are being tested.

As CNBC reports, their performance is far from perfect. For example, one robot ignored a woman who was trying to inform it of a nearby fight, asking her only if…she would move.

Why does it matter

In the larger scheme of things, it doesn’t. But it is a funny yet welcome reminder that the road toward fully autonomous machines is more tortuous than we might expect.

Sotabench – benchmarking models for everyone

The progress that has been made in machine learning is mind-boggling. AI Monthly Digest was designed to cut through the buzz and deliver reliable and trustworthy news. On the other hand, it is easy to fall into the trap of considering every new model “state-of-the-art” or “best-performing”.

Yet there are multiple benchmarks to test models and determine which one is best suited for purpose. One obstacle to that being done is that the developers community cannot, due to a lack of time, test the models.

To address this problem, the people behind Papers With Code launched the site sotabench.com, which undertakes to benchmark all open source models with tests.

Why does it matter

The initiative is the next step toward establishing a set of good practices among machine learning developers and researchers. In “Papers with code”, the group promoted the idea of publishing not only research papers, but also the code behind them. Thus, anyone who so chose to could check and reproduce the effects of a research him or herself. This step makes claims of delivering the next breakthrough more feasible.

In fact, any effort to make the development of AI more transparent and easier to reproduce is more than welcome.

AI Monthly Digest #13 – an unexpected twist for the stock image market

AI Monthly Digest #13 – an unexpected twist for the stock image market

/in , /by and

September brought us two interesting AI-related stories, both with a surprising social context.

Despite its enormous impact on our daily lives, Artificial Intelligence (AI) is often still regarded as too hermetic and obscure for ordinary people to understand. As a result, an increasing number of people use Natural Language Processing-powered personal assistants, yet only a tiny fraction try to understand how they work and how to use them effectively. This makes them somewhat of a black box.

Making the field more comprehensible and accessible is one aspect of  AI researchers’ mission. That’s why research recently done by OpenAI is so interesting.

Hide-and-Seek – the reinforcement learning way

Reinforcement learning has delivered inspiring and breathtaking results. The technique is used in the training models behind autonomous cars and the controlling of sophisticated devices like automated arms and robots.

Unlike in supervised learning, a reinforcement learning model learns by interacting with the environment. The scientist can shape its behavior by applying a policy of rewards and punishments. The mechanism is close to that which humans use to learn.

Reinforcement learning has been used to create super killing agents to go toe-to-toe against human masters in Chess, Go and Starcraft. Now OpenAI, the company behind the GPT-2 model and several other breakthroughs in AI, has created agents that play a version of hide-and-seek, that most basic and ageless of children’s games.

OpenAI researchers divided the agents into two teams, hiders and seekers, and provided them a closed environment with walls and movable objects like boxes, walls and ramps. Any team could “lock” these items to make them unmovable for the opposing team. The teams developed a set of strategies and counter-strategies in a bid to successfully hide from or seek out the other team. The strategies included:

  • Running – the first and least sophisticated ability, enabling one to avoid the seekers.
  • Blocking passages – the hider could block passages with a box in order to build a safe shelter.
  • Using a ramp – to overcome the wall or a box, the seekers team learned to use a ramp to jump over an obstacle or climb a box and see the hider.
  • Blocking the ramp – to prevent the seekers from using the ramp to climb the box, the hiders could block access to the ramp. The process required a great deal of teamwork, which was not supported by the researchers in any way.
  • Box surfing – a strategy developed by seekers who were basically exploiting a bug in the system. The seekers not only jumped on a box using a ramp that had been blocked by the hiders, but also devised a way to move it while standing on it.
  • All-block – the ultimate hider-team teamwork strategy of blocking all the objects on the map and building a shelter.

The research delivered, among other benefits, a mesmerizing visual of little agents running around.

Why does it matter?

The research itself is neither groundbreaking nor breathtaking. From a scientific and developmental point of view, it looks like little more than elaborate fun. Yet it would be unwise to consider the project insignificant.

AI is still considered a hermetic and difficult field. Showing the results of training in the form of friendly, entertaining animations is a way to educate society on the significance of modern AI research.

Also, animation can be inspiring for journalists to write about and may lead youth to take an interest in AI-related career paths. So while the research has brought little if any new knowledge, it could well end up spreading knowledge on what we already know.

AI-generated stock photos available for free

Generative Adversarial Networks have proved to be insanely effective in delivering convincing images of not only hamburgers and dogs, but also human faces. One breakthrough is breathtaking indeed. Not even a year ago the eerie “first AI-generated portrait” was sold on auction for nearly a half-million dollars.

Now, generating faces of non-existent people is as easy as generating any other fake image – a cat, hamburger or landscape. To prove that the technology works, the team behind the 100K faces project delivered a hundred thousand AI-generated faces to use in any stock usage, from business folders, to flyers to presentations. Future use cases could include delivering on-the-go image generators that, powered by a demand forecasting tool, provides an image that best suits demand.

More information on the project can be found on the team’s Medium page.

Why does it matter

The images added to the free images bank are not perfect. With visible flaws in a model’s hair, teeth or eyes, some are indeed far from it. But that’s nothing a skilled graphic designer can’t handle. Also, there are multiple images that look nearly perfect – especially when there are no teeth visible in the smile.

Many photos are good enough to provide a stock photo as a “virtual assistant” image or to fulfill any need for a random face. This is an early sign that professional models and photographers will see the impact of AI in their daily work sooner than expected.

The full year of AI Monthly Digest

The full year of AI Monthly Digest

/in /by

The AI Monthly Digest is deepsense.ai’s attempt to filter out the buzz and chaos from the news ecosystem and provide readers with the news that matters. The team also strives to make the information understandable for lay readers, to make knowledge on developments in Artificial Intelligence more accessible and the domain itself less esoteric.

While keeping an eye on global developments in AI, deepsense.ai’s team behind the digest got not only fresh and up-to-date information on state-of-the-art technologies, but also significant insight into the ongoing changes in the industry.

So what has changed? What has happened since the first issue of AI Monthly Digest came out? In this text we will cover:

  • Natural Language Processing improvements
  • Image processing breakthroughs
  • Reinforcement Learning transfer from research to business applications
  • Societal impact of Artificial Intelligence (AI)

Natural language processing

Two major breakthroughs have brought about rapid progress in Natural Language Processing (NLP).

BERT, XLnet and the speed of change

The first breakthrough was Google delivering BERT, a model that improves pre-trained word embeddings (a vector representations of words that enable computers to process the text) and enables data scientists to further fine-tune their networks to fulfill specific roles, such as automated chatbots or document processing support tools. BERT has been around since October 2018.

Less than a year later, XLNet outperformed BERT and delivered the new state-of-the-art in NLP. Indeed, one’s cutting edge can get rusty in no time flat.

GPT-2 – a long story

The next breakthrough witnessed by AI Monthly digest readers was the GPT-2 model, which excels in natural text generation. The model introduced in February delivered texts that were nearly indistinguishable from ones delivered by a human writer. To see just how indistinguishable, have a look at the examples below.

Example:

Human written:

In a shocking finding, scientists 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.”

The effects were so impressive that OpenAI, the company behind the creation, decided NOT to go public with the model, flouting established industry and research standards. But they believed they were acting responsibly.

Over the past year we have also reported on research on the business efficiency of automated translations. A key fact that has come to light is that there is a 10.9% increase in trade when automated translations are used in the e-commerce. That’s fairly convincing evidence for the business efficiency of AI-powered translations.

Image Recognition

In NLP, state-of-the-art technology is struggling to match human accuracy, whereas machines have already achieved superhuman performance in image recognition. And to say that advancement has been rapid would be an understatement.

In the second edition of AI Monthly Digest, news about delivering a convincing image of a hamburger, a cat or a dog was a thing indeed. Mere months later, neural networks were delivering very realistic faces–and not just of fake people. Researchers generated a network that makes a video of person speaking just from a single frame. As a part of the research, the team delivered speaking paintings, including the Mona Lisa and Van Gogh. Admittedly, the hamburger paled in comparison.

Elsewhere, a neural network developed the ability to spot early signs of depression in humans. According to WHO data, depression is a leading cause of disability in the world and when not cured can lead to suicide. The software was accurate in spotting the signs of depression in more than 80% of cases.

Thus neural networks and the data scientists behind them can boost healthcare not only by employing machine learning for drug discovery, but also to make early diagnoses. And in the case of spotting depression, they managed to do this using just the information they could gather from the cameras on mobile devices.

Reinforcement learning

One of the most important areas of research work, reinforcement learning was a frequent area of focus in the first 12 months of AI monthly Digest.

An important development was Deepmind’s AlphaStar beating human champions in Starcraft 2, The game poses multiple challenges for human players and artificial intelligence alike. It is played in real time, requiring excellent reflexes. Every unit has a strategy it works with and counter-strategies that make it useless.

Also, due to the fog of war (the darkness covering the map outside of the player’s units sightline) the player has incomplete information on their opponent’s position and actions. Compare that with the complete openness of chess and Go.

The Deepmind-designed AlphaStar model vanquished top StarCraft pro-players TLO and MaNa from TeamLiquid, though not without a catch: the model had an unfair advantage over a human player by being able to “see” all the visible battlefield at the same time. Humans are limited to the frame they can see on their screen, so the ability to switch between multiple spots on a world map is limited. This advantage proved so crucial that when the model was stripped of it, MaNa won the match.

While gaining superiority over human players is one thing, being a worthy and entertaining opponent is something else entirely. The AI in gaming is currently based on endless if-then loops that are prone to failure, especially in the more complicated, open worlds seen in The Witcher 3 or Grand Theft Auto.

The first game to feature a reinforcement learning-trained agent as an opponent is the MotoGP 19, a racing game.

So, during the first 12 months of AI Monthly Digest, reinforcement learning moved beyond the realm of pure research and into real-life business applications.

Society

Like the steam engines of the past, Artificial Intelligence is poised to transform society in unprecedented ways. The first signs of change are being seen now – sometimes for the better and sometimes for the worse.

Amazon’s AI-based recruitment solution was biased against women in the company’s recruitment process. The main reason was that male coders were overrepresented in the dataset the recruitment model was trained on. Amazon the model down, but not before giving us a glimpse of the possible effects of deploying insufficiently supervised models.

To avoid such situations and make AI technology more comprehensible for non-skilled users, the Finnish government launched an AI-popularization program. The grassroots movement of providing the uninitiated– barbers, bakers and car mechanics, for example – with basic AI training, gained the attention of the Finnish government, which sought to make machine learning a commodity. An AI-powered demand forecasting for a small, family business? Why not?

A humorous but thought-provoking example of the social impact AI research can have is the panic Starcraft players exhibited when Deepmind revealed that AlphaStar was lurking in Battle.net, and was primed to do battle with random players.

Given the superiority AlphaStar had shown over the MaNa and TLO, it comes as no surprise that players were at once thrown back on their heels and put on their toes:killer AI had come to humiliate them and turn the Battle.net rankings on its head. Thus, users exchanged tips on spotting the non-human opponent. So, while AlphaStar no longer wields an unfair advantage, the players remain thrilled.

Summary

In 12 short months, we’ve gone from being amazed by a convincing image of a hamburger to delivering images of fake people. The world came through three breakthroughs in natural language processing and players have witnessed turmoil around reinforcement learning-powered agents.

The world of AI is changing fast. AI Monthly Digest will be keeping its A-eye trained on the ball.

AI Monthly Digest #12 - the shadow of malicious use

AI Monthly Digest #12 – the shadow of malicious use

/in , /by and

With this edition of AI Monthly Digest, we have now for a full year been bringing readers carefully selected and curated news from the world of AI and Machine Learning (ML) that deepsense.ai’s team considers important, inspiring and entertaining.

Our aim is to deliver information that people not necessarily involved in AI and ML may find interesting. Also, the digest is curated by data scientists who ensure that the information included isn’t just hot air or marketing mumbo-jumbo, but significant news that will impact the global machine learning and reinforcement learning world.

This edition focuses on natural language processing, as the GPT-2 model is still an important element of AI-related discourse. This edition also contrasts the enthusiasm of ML-developers with concerns expressed by a renowned professor of Psychology.

1200 questions to ask

With natural language processing, a computer needs to generate natural texts in response to a human. This is at least troublesome, especially if a longer text or speech is required.

While these problems are being tackled in various ways, the gold standard is currently to run the newest solution on a benchmark. Yet delivering one is another challenge, to put it mildly.

To tackle it, researchers from the University of Maryland created a set of over 1200 questions that are easy to answer for a human and nearly impossible for a machine. To jump from “easy” to “impossible” is sometimes a matter of very subtle changes. As the researchers have said:

if the author writes “What composer’s Variations on a Theme by Haydn was inspired by Karl Ferdinand Pohl?” and the system correctly answers “Johannes Brahms,” the interface highlights the words “Ferdinand Pohl” to show that this phrase led it to the answer. Using that information, the author can edit the question to make it more difficult for the computer without altering the meaning of the question. In this example, the author replaced the name of the man who inspired Brahms, “Karl Ferdinand Pohl,” with a description of his job, “the archivist of the Vienna Musikverein,” and the computer was unable to answer correctly. However, expert human quiz game players could still easily answer the edited question correctly.

Capitalizing on this knowledge, researchers will be able to deliver better benchmarking for models and thus determine which part of the question confuses the computer.

Why does it matter

With each and every breakthrough, researchers get closer to delivering human-level natural language processing. At the same time, however, it is increasingly hard to determine if the neural network is understanding the processed text, or is just getting better fitted to the benchmark. Were the latter the case, the model would outperform existing solutions but register no significant improvement in real-life performance.

An example with detailed explanations are available in the video below.

A benchmark updated with those 1200 questions delivers significantly more precise information on the model’s ability to process the language and spot the drawbacks.

Large GPT-2 released

GPT-2 is probably the hottest topic among AI Trends 2019, especially considering the groundbreaking effect and controversial decision to NOT make the model public. Instead, OpenAI, the company behind the model, decided to cooperate with chosen institutions to find a way to harden the model against potential misuse.

And the threat is serious. According to research published in Foreign Affairs, readers consider GPT-2-written texts nearly as credible and convincing as those written by journalists and published in The New York Times (72% compared to 83%). Thus the articles are good enough to be especially dangerous as a weapon of mass disinformation or fake news factory – AI can produce a limitless amount of credible-looking texts with no effort.

To find the balance between supporting the development of the global science of AI and protecting models from being used for maleficent ends, OpenAI is releasing the model in iterations, starting a small one and ultimately aiming to make the model public but with the threat of misuse minimized.

Why does it matter

As research published in Foreign Affairs states, the model produces texts that an unskilled reader will find comparable to journalist-written ones. Image recognition models are already outperforming human experts in their tasks. But with all these cultural contexts, humor and irony, natural language once seemed protected by the unassailable fortress of the human mind.

The GPT-2 model has apparently cracked the gates and with business appliances it may be on the road to delivering a model that can provide human-like performance. The technology just needs to be controlled so as not to fall into the wrong hands.

What is this GPT-2 all about?

A GPT-2 model is, as stated above, one of the hottest topics of AI in 2019. But even the specialist can find it hard to understand the nitty-gritty of how the model works. To make the matter more clear, Jay Alammar has prepared a comprehensive guide to the technology.

Why does it matter

The guide is good enough to allow a person who has limited to no knowledge on the matter to understand the nuances of the model. For a moderately skilled data scientist given sufficient computing power and a dataset, the guide is sufficient to reproduce the model for example to support demand forecasting with NLP. It enables a data scientist to broaden his or her knowledge with one comprehensive article – a convenient way indeed.

Doing research is one thing, but sharing the knowledge it affords is a whole different story.

Malicious use, you say?

Jordan Peterson is a renowned professor and psychologist who studies the structure of myth and its role in shaping social behavior. If not a household name, he is certainly a public person and well-known speaker.

Using deep neural networks, AI researcher Chris Vigorito launched a notjordanpeterson.com website that allowed any user to generate any text that was later read with the neural network-generated voice of Jordan Peterson. As was the case with Joe Rogan, the output was highly convincing, mirroring the manner of speaking, breathing and natural pauses.

The networks was trained on 20 hours of transcripted Jordan Peterson speeches, an easy number to obtain where a public speaker is concerned. The amount of work was considerable, but not overwhelming.

Why does it matter

The creation of the neural network is not as interesting as Jordan Peterson’s response. He has written a blogpost entitled “I didn’t say that”, where he calls the situation “very strange and disturbing”. In the post, he notes that while it was fun to hear himself singing popular songs, the prospect of being an unwitting part of a scam is more than real. Due to the rising computing power available at affordable prices and algorithms getting better and less data-hungry, the threat of this technology being used for malicious ends is rising. If you’d like to know just how malicious he means, I’ll leave you with this to consider.

I can tell you from personal experience, for what that’s worth, that it is far from comforting to discover an entire website devoted to allowing whoever is inspired to do so produce audio clips imitating my voice delivering whatever content the user chooses—for serious, comic or malevolent purposes. I can’t imagine what the world will be like when we will truly be unable to distinguish the real from the unreal, or exercise any control whatsoever on what videos reveal about behaviors we never engaged in, or audio avatars broadcasting any opinion at all about anything at all. I see no defense, and a tremendously expanded opportunity for unscrupulous troublemakers to warp our personal and collective reality in any manner they see fit.