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.

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?