In 2016, we saw huge innovation, a lot of investment in artificial intelligence from large companies and startups, and a lot of public interest. What will 2017 bring?

1. Democratizing tools will give more companies access to artificial intelligence.

A recent Forrester study of organizations and technology professionals found that 58% are exploring artificial intelligence (AI), but only 12% are using AI systems. This is partly because applications for AI are only now beginning to emerge, but also because the technology is still in its infancy and difficult to apply. Working with them requires a set of certain skills and attitude.

Frameworks like Facebook's Wit.ai and Howdy Slack are trying to become a kind of Visual Basic of artificial intelligence, promising simple development of intelligent conversational interfaces without the highest level of developer training. Tools like Bonsai, Keras, and TensorFlow make it easy to implement deep learning models. Cloud platforms such as Google and Microsoft Azure interfaces allow you to create intelligent applications without having to worry about setting up and maintaining the associated infrastructure.

2. We'll see more goal-oriented AI systems

We don't expect large general purpose AI systems to emerge just yet. However, we can expect the emergence of targeted systems such as:

• Robotization: personnel, industry and retail
• Autonomous vehicles (cars, drones, etc.)
• Bots: CRM, consumer (like Amazon Echo) and personal assistants
• Industry-specific AI systems: finance, healthcare, security and retail

3. The economic impact of increased automation will be a topic of debate.

In 2017, the hope is to hear less about how evil artificial intelligence is going to take over the world and more about the economic impact of AI. Concerns about job losses due to AI are no longer new, but we should expect deeper and more detailed conversations about the economic impact of such developments.

4. Systems to prevent information overload will become more complex.

Interesting developments are underway in the AI ​​segment that will help analyze information and prevent its excess, especially in areas such as:

• Natural Language Understanding
• Structured Data Mining
• Mapping information
• Automatic summarization (text, video and audio)

5. AI researchers will study and sometimes solve fundamental problems

In 1967, one of the founders of the artificial intelligence laboratory at the Massachusetts Institute of Technology, Marvin Minsky, said that within the lifetime of one generation the problem of creating artificial intelligence would be solved. Was he a visionary or was he wrong? It is impossible to say yet, but fundamental problems remain to be solved. However, progress does not stand still, here are some examples:

• Research is underway on building blocks that include natural language recognition, vision, speech, advanced retrieval learning, and optimization of hardware systems (including cost) for AI flows.
• Systems with improved attention and memory will be able to solve more than one problem at a time or more complex problems such as guessing and reasoning. For example, recent progress has been made at DeepMind in the area of ​​differentiable neural computers.
• Algorithms will require less and less labeled data; uncontrolled or semi-controlled learning develops.
• Inspired by key aspects of the human brain systems, including intuitive physics and psychology, rapid model building, and cause-and-effect inference.
• AI systems for building reliable software. Peter Norvig spoke about this in detail at the O’Reilly AI Conference.

6. The interaction between people and machines will become more intense

There is diversity in machine intelligence that ranges from pure machine intelligence to augmenting human capabilities. Developments in emotional intelligence and detection, and in human-powered solutions, will enable richer interactions between humans and machine intelligence.

7. Expect increased attention to ethics and privacy issues

Most AI systems can be described as black boxes and are extremely complex. The ethical and privacy risks associated with AI are real and require careful consideration. These problems will not be solved in 2017, but we can at least hope for progress in this area.

They are developing more and more quickly, and new applications are being found for them. People are inventing new ways to teach machines, and machines are learning to do without people at all.

What happened?

Major breakthroughs in this direction are now based on deep reinforcement learning. For example, it showed excellent results of self-training: in three days the neural network went from the level of a novice Go player to the level of a professional who only wins victories. The new AI model AlphaGo Zero was trained from scratch without human intervention, playing only with itself.

It is no coincidence that the first prize at the ICML 2017 conference was given to the work Understanding Black-box Predictions via Influence Functions, the authors of which do not try to build new models, but explain why existing models give certain results. Online access to deep learning educational programs is also expanding. As part of the AI ​​Experiment project, in collaboration with Støj, Use All Five and the Creative Lab and PAIR teams at Google, a machine learning platform has been created. It allows everyone to try their hand at self-training a neural network and understand how machine learning works.

Research universities and institutions such as Oxford, MIT Boston, and GE Avitas Systems have invested in supercomputers designed for deep learning of neural networks. And began work in the USA. He will study social connections and the interaction of artificial intelligence and various areas of human activity.

Nvidia researchers indicate that new learning methods such as generative adversarial networks (GANs) will take AI capabilities to previously unseen heights. Not long ago, we told how a group of scientists from Nvidia taught artificial intelligence quickly and almost imperceptibly.

Neural networks have also been used for more “mundane” tasks. For example, Carlsberg launched a study. Artificial intelligence developed by Microsoft helps to create new types of beer. In New Zealand, there has appeared, making the “smartest” election promises. And the famous manufacturer of realistic sex toys, RealDoll, has launched a new one, which will specialize in robotic sex dolls with artificial intelligence.

At the same time, beauty is no stranger to neural networks. For example, he created beautiful special effects for a video of a Chinese singer. And the Russian voice assistant with artificial intelligence “Alice” recorded for the New Year.

What will happen?

Despite the fears of opponents of AI, the field will undoubtedly develop actively in the coming year. Artificial intelligence should make robots smarter, medicine more accessible, and recognition systems commonplace. Nvidia has collected forecasts from researchers and experts from around the world to understand how all this will develop in the future.

Experts are confident that AI will find even wider practical application in medicine and will become an integral part of the healthcare sector. For example, Mark Mikalski, executive director of the Center for Clinical Data Analysis at Massachusetts General Hospital and Brigham and Women's Hospital, believes that the medical field will move to creating real products for clinical practice. “AI will begin to penetrate medicine from the diagnostic side, and other segments will not be long in coming - it will be adopted by disease prevention specialists, surgeons and doctors of other specialties,” Mark Mikalski is quoted as saying in the Nvidia study.

Luciano Prevedello, a physician in the Department of Radiology and Neuroradiology at The Ohio State University Wexner Medical Center, is confident that “in 2018 and the next few years, AI will be so deeply embedded in medicine that it will be perceived as an integral part of it.” From research laboratories, AI will move into the patient room.

Experts have no doubt that artificial intelligence will enter deeper into everyday life. And AI development will occur in a variety of areas. In particular, such technologies will be used in smartphones. And new deep learning methods will add transparency to data processing.

Orange Silicon Valley CEO and President of the Orange Institute research laboratory Georges Nahon believes that biometrics will replace credit cards and driver's licenses this year. “Facial recognition has already revolutionized security through the use of biometrics, and seeing how technology and retail are merging, such as Amazon with Whole Foods, I think that in the near future people will not have to stand in lines,” he said.

The entire technology industry will change under the influence of artificial intelligence, Nicola Morini Bianzino, Managing Director for AI and Head of Strategic Planning at Accenture, is confident: “AI will account for 25% of technology spending. The key question will be how organizations and employees respond to the changes that AI technologies will bring.”

Also, artificial intelligence will be increasingly and more effectively used to create content - music, images, games, texts. And “smart” things will become smarter and more personalized. The time is not far off when it will be enough to think about something and get what you want, says Nvidia senior researcher Alejandro Troccoli.

Kai Fu Lee, chairman of SinovationVentures, believes that AI is "aimed at large-scale job losses" while concentrating wealth in the hands of companies that develop or adopt AI. Others believe that similar fears were present with the advent of all world-changing technologies, right back to the printing press in the 15th century.

The Economist reassures readers that "AI is creating demand for jobs" and a growing number of people around the world are "providing digital services online." Which companies and countries will thrive in the era of AI? Which segments will disappear, change, or be created? How will the nature of work change?

Military affairs

Proponents of armed drones argue that such weapons can hit targets with much higher accuracy than humans; and the larger the role they play in the theater of operations, the less often technicians will use them to harm.

But what if such weapons become independent and work independently, without human intervention? Will removing people from the military personnel list lead to an even more severe and unstoppable arms race?

An open letter published during the 2015 International Joint Conference on Artificial Intelligence warned that autonomous weapons "require no expensive or hard-to-find raw materials and will therefore become ubiquitous and cheap for all significant militaries to mass produce." Will an era with automated weapons be more peaceful or more militant?

RAND researchers are calling for an analytical framework and international effort focused on the use of long-range armed drones in counterterrorism and targeted assassinations.

Decision making

Politicians are constantly faced with a huge number of choices and motivations - many more in the days of social media than twenty years ago. Such information overload makes it difficult to cope during a crisis, let alone multiple crises.

Recently, a proposal arose to pass “all the decisions that the president makes through a computer - not to make the final choice, but to help the leader in the person of a person.”

But while AI is now largely blameless, the RAND study highlights the risks of algorithmic biases in filtering news, influencing criminal justice, and even the delivery of Social Security benefits and visas. What decisions should be entrusted to AI? What should remain in the hands of man? In the hands of a team of people?

Creation

The world has become accustomed to AI that can perform breathtaking feats of computation and beat humans at popular board games (it's been just over 20 years since the IBMDeepBlue supercomputer famously defeated chess grandmaster Garry Kasparov). How will it further progress in people's creative space?

Artificial intelligence researcher Jesse Engel believes it will “transform the creative process...by augmenting it with smart tools that provide new possibilities for expression.” Others are not so optimistic. Journalist Adrienne Lafrance notes that AI can already “flirt,” “write novels,” and “fake famous paintings with amazing accuracy.” What does it mean to be creative? Moreover, what does it mean to be human?

Discussions of AI often veer to extremes, be it the promise of a utopia free of human suffering or the danger of a dystopia where robots enslave their human creators. More balanced and rigorous analysis is needed to help shape policies to mitigate risks and maximize benefits. Certain steps need to be taken to overcome fears that AI will overwhelm the state and society.

How can AI impact a country's national interests? What types of AI, if any, can be considered strategic technologies based on government criteria? Where should market forces play a role, and where should politics play? While AI remains largely the stuff of science fiction, these questions are becoming more and more important.

The topic of Artificial Intelligence (AI) dominated the media news feed throughout the year. The tone is set by the main newsmakers - Elon Musk and Mark Zuckerberg, discussing the dangers and benefits of using Artificial Intelligence in human life. Russia and China have declared the development of AI as a priority direction in the digital economy. 2018 will be a year of development and further study of the possibilities of using AI, especially the method of deep learning, as the most promising branch of Artificial Intelligence. I’ll tell you more about this trend in the field of high technology using the example of the use of AI in marketing.

The essence of Artificial Intelligence is to create machines so smart that they will surpass the thinking and analytical abilities of humans. Machine learning, a basic AI method, has such capabilities and is already widely used in many sectors of the economy and areas of human life. However, other, more advanced technologies are rapidly developing.

This is especially noticeable in the pace of development of deep learning, which almost completely replicates the principle of operation of the human brain in data processing and decision-making modeling. In 2017, deep learning became an integral part of technology processes in healthcare and automotive manufacturing. Marketing, as the most dynamic component of every business, has also not remained aloof from the use of advanced technologies. Deep learning has had a revolutionary impact on the entire advertising industry.

The technology used in the deep learning method is based on the principles of interaction of biological neurons. With the help of self-learning algorithms, marketers now obtain descriptions of a customer's buying potential without human assistance. For example, RTB House recently analyzed a huge amount of data, clearly demonstrating that using Artificial Intelligence instead of recommendations from experienced marketers in retargeting campaigns can improve conversion results by 35%. And that's not all. Using the deep learning method, advertisers receive a forecast of user actions based on an analysis of his behavioral characteristics and desires. This greatly simplifies the work of a marketer by offering the best options for targeted advertising messages containing products that the user did not even know about or had not yet seen.

Many major brands have already seen the benefits of implementing deep learning solutions into their products or marketing tools. In 2018, we expect widespread use of deep learning and increased investment in developing its potential.

From “supervised learning” to new horizons

In 2017, there was a move away from the so-called “supervised learning” typical of the machine learning process towards a more complex system of “transfer learning”. This technology is based on transmitting human instructions to a computer: analyze existing decision-making models, examples, data sets and their subsequent analysis.

The way transfer learning works is the ability of a computer to process data from simulations rather than from reality. This process is much simpler and cheaper, as well as faster, which is very important when analyzing huge amounts of data. Using this method, the machine learns to make decisions on its own: with logical conclusions, analogy or deductive method.

For example, using an older machine learning model, a self-driving car could take a person millions of miles while data is being recorded. This data is transmitted to the car, which understands how to drive the car based on the driver's decisions. Thanks to “learning transfer,” there is no longer any need for a real driver. Instead, data can be taken from various driving simulations. By simulating millions of hours of driving, the car itself understands where it needs to go, and it already translates the knowledge into the real world.

The second approach is called “reinforced learning.” Its goal is to train a computer to make the best decisions based on feedback from the environment and the actions taking place in it. For example, how this happens when participating in bidding for the purchase of advertising space. Auction systems are very complex. Even experts often have problems determining the optimal rate that will allow them to achieve the desired results at minimal cost. The car will encounter the same obstacles at the beginning of its movement. However, unlike a person, a car can operate 24 hours a day in a simulation environment. And it can also learn a set of actions, much faster than a human. Returning to our example of buying advertising space, the computer learns from simulating auctions, receiving data on how to act most efficiently and thus win the auction.

New jobs and new challenges

Indeed, the operating principle of deep learning algorithms is absolutely identical to the functioning of the human brain. But, unlike people, computers learn much faster and can analyze enormous amounts of data. Computers don't fall asleep and make a lot of mistakes. This is where super performance comes into play. In a very simple way, AI will strive to surpass human abilities in many areas. Currently, self-learning algorithms are able to recognize actions and images much more accurately than humans.

Does this mean that there is a danger of people being completely replaced by robots? Not really. According to the World Economic Forum, 65% of children entering primary school today will be given jobs that do not currently exist. The current level of AI development allows companies to look for more IT specialists, data analysts, and programmers. Next year we'll likely see a boom in new job offers for data scientists. Although now such a proposal is not yet popular.

Innovations of 2017 will receive a powerful impetus for development in 2018

The goals pursued by the implementation of the deep learning method are to simplify our lives and increase the efficiency of human activity. This is why the use of AI is no longer a standard, but a necessity for companies that want to be competitive in the global market. This is not about the ability to personalize or improve the capabilities of the final product, but also about a number of other indirect activities such as data collection and analysis. Already, companies have such a large amount of data to analyze that they cannot cope with its processing.

This situation directly affects the decisions made by their employees and therefore their financial results. Companies whose business specializes in collecting and analyzing data for various enterprises will be increasingly in demand. Businesses with larger budgets will use AI to classify: what to offer to customers, what terms to recommend to suppliers, how to instruct employees, what to say and do in real time. It should also be assumed that many new startups will soon emerge offering solutions based on self-learning algorithms as this technology becomes widespread.

Artificial intelligence in 2017 has become part of our daily lives and public discussions. In the coming years, the focus will be on developing various AI-based technologies that will replace humans in many complex industries, ultimately making our lives much easier. But this will require a lot of work.

The WILDML project, dedicated to machine learning technologies, has published a brief overview of the most important achievements in the field of artificial intelligence over the past year.

Machine learning technologies surpass human capabilities

Perhaps the biggest public success of the year was the release of a new version of AlphaGo, a program that, thanks to reinforcement learning, defeated the world champions in the game of Go. Due to the huge number of possible moves, it was believed that artificial intelligence would not be able to win this game for at least a couple of years.

Initially, AlphaGo neural networks were trained on human moves, after which they began to play with themselves, relying on a method called Monte Carlo Tree Search. The new AlphaZero model learned to play better than previous versions without any training data. By the end of the year, an even more improved AlphaZero algorithm was released, which, in addition to Go, also plays chess and shogi. The program’s abilities amaze even the most advanced players; they themselves are ready to learn from it and borrow clever combinations. To help them, DeepMind released special training software AlphaGo Teach.

However, Go is not the only game that the computer has mastered. The Libratus program, developed at Carnegie Mellon University, managed to compete with top poker players at the 20-day Texas Hold'em Championship. A little earlier, the DeepStack poker bot beat professionals for the first time. It was created by scientists from Charles University (Czech Republic), Czech Technical University and the University of Alberta (Canada). In both cases, the game was heads-up, where generating decisions is much easier than in a full-table game.

So the next big thing for reinforcement learning technologies to achieve is more complex games with more players. DeepMind is actively working to turn the Starcraft 2 environment into a test bed for artificial intelligence, and the OpenAI bot that defeated the strongest players in Dota 2 one-on-one will in the near future be able to take on the pros in a five-on-five game.

The Return of Evolutionary Algorithms

In supervised training of neural networks, the backpropagation method is successfully used, and a replacement for it will not be found soon. And in reinforcement learning, evolutionary methods based on different principles than gradient algorithms seem to be becoming relevant again. This way, neural networks can be trained in parallel and at very high speed, on thousands of computers. In this case, expensive graphics chips are not needed - you can use a large number (from hundreds to thousands) of relatively cheap central processors.

In early 2017, researchers at OpenAI also showed that evolutionary strategies can achieve results as good as conventional reinforcement learning algorithms such as Q-learning. By the end of the year, the Uber team blogged about a number of scientific papers demonstrating the potential of genetic algorithms and further research. Using a simple genetic algorithm without any gradients, their neural network learns to play complex Atari games, and performs ten times better than DQN, AC3 or other evolutionary strategies.

New speech generation models, image recognition neural networks, and attention mechanisms

In addition to the general deep learning frameworks, many reinforcement learning frameworks have emerged:

Frameworks that run directly in the browser, such as deeplearn.js from Google and WebDNN from MIL, are designed to make deep learning more accessible. At least one fairly popular framework did not survive this year - Theano. The library developers reported that version 1.0 will be the last.

Educational Resources

With the rise in popularity of deep learning and reinforcement learning in 2017, more and more online courses, lectures and events began to appear. Here are the highest quality ones according to wildml:

• Lectures on Reinforcement Learning Fundamentals and Advanced Research Deep RL Bootcamp, co-provided by OpenAI and UC Berkeley;
• Spring course Stanford on the application of convolutional neural networks to computer vision. It’s worth taking a look at the official website of the course;
• Winter course Stanford on deep learning in natural language processing. And the course page;
• Course on deep learning theories from Stanford;
• New DL specialization on Coursera;
• Materials of the Summer School on DL and RL in Montreal;
• UC Berkeley Fall Course on Machine Learning ;
• Developer Conference Tensorflow Dev Summit and talks on deep learning fundamentals and API changes;
• Materials from many scientific conferences are now being posted online. You can learn all about cutting-edge research from recordings from NIPS 2017. ICLR 2017 and EMNLP 2017;

Researchers also post free educational materials and scientific articles on arXiv. Here are some interesting examples:

• Deep Reinforcement Learning: An Overview - an overview of the latest advances in RL;
• A Brief Introduction to Machine Learning for Engineers - a brief introduction to machine learning for developers;
• Neural Machine Translation - about neural machine translation;
• Neural Machine Translation and Sequence-to-sequence Models: A Tutorial - a guide to neural machine translation and sequence-to-sequence learning models.

AI in medicine

In 2017, there were many bold claims that deep learning technologies in medicine would surpass human abilities. But despite the hype, assessing the significance of discoveries for a person far from medicine is not at all easy. He talks about them quite comprehensively in his blog The End of Human Doctors Luke Oakden-Rayner. Here is a summary of the most important points.

One of the biggest news was the creation by researchers at Stanford University of a deep learning algorithm that diagnoses skin cancer no worse than qualified dermatologists. You can read about this on the Nature website. Another team from Stanford has developed a machine learning model that can identify signs of arrhythmia from an ECG more effectively than doctors.

2017 was not without its punctures. Many questions have arisen about the deal between DeepMind and the UK's National Health Service, which opened up access to patient data. The US National Institutes of Health provided the scientific community with more than 100 thousand chest X-rays, but they later turned out to be useless for training neural networks.

AI in art

Generative modeling is increasingly being used to generate images, sketches, music and video. At the NIPS 2017 conference, a master class on the topic of ML in creativity and design was held for the first time.

A real breakthrough in 2017 was made by generative adversarial networks (GANs). Impressive results were shown by the CycleGAN, DiscoGAN and StarGAN models, which can, for example, draw faces. Typically, GANs have had a hard time producing realistic high-resolution images, but pix2pixHD can fix that soon.

AI for self-driving cars

The largest developers of self-driving cars are taxi services Uber and Lyft, Waymo (a subsidiary of Alphabet) and Tesla. Uber started the year not entirely successfully: in San Francisco, their drone ran red lights several times due to a software glitch, and not due to the driver’s fault, as previously reported. Uber also shared some details about its data visualization platform. By December, Uber cars had covered more than 3.2 million km autonomously.

Meanwhile, in April, the first customers were given a ride by Waymo's self-driving cars, and in Phoenix the company completely abandoned “live” drivers in testing. The company also talks about how it trains in simulation mode and tests its cars.

Lyft has announced that it is working on its own processors and related software and is already testing it in Boston. Tesla's Autopilot hasn't changed significantly, but it has a new competitor: Apple. Tim Cook has confirmed that his company is developing software for self-driving cars; some of the researchers' work can already be found on arXiv.

Research projects

A huge number of impressive projects and demos were released in 2017; it is impossible to mention them all in one review. Here are just a few developments:

• Neural networks that can independently change the background of images;
• Model to create