Can a robot write a symphony? The Turing Test - “You’re just a robot, an imitation of life. Can a robot beat this game? Disadvantages of VR therapy are the relatively high cost of equipment and potential physiological intolerance to virtual reality.

The development of elements of artificial intelligence systems is actively developing and becoming a trend of today, and only the lazy do not write smart bots. That's why we interviewed Dmitry shwars Soshnikov, one of the country's best AI experts. He is a Microsoft technology evangelist, author of books, articles and tutorials, as well as a teacher and student program coordinator, specialist in functional programming, neural networks and artificial intelligence.

- Dmitry, please tell us a few words about yourself and your work.

Dmitry Soshnikov: As an evangelist for Microsoft, I am involved in popularizing and implementing the company’s most modern technologies, now primarily technologies related to artificial intelligence. This includes speaking at conferences, working with students and teachers, startups, sometimes participating in hackfests, and programming prototype systems together with developers from different companies. I also popularize functional programming and the F# language, teach at MIPT, HSE and MAI, and develop online courses for MVA and Coursera.

In my opinion, neural networks and artificial intelligence in general are a very interesting area, the rapid development of which in recent years has already made it possible to solve a number of problems that previously had no solution, such as, for example, automatically determining a person’s age from his photograph. And the future holds many more interesting opportunities for us.

Neural networks - fashion or tool?

- What exactly has recently prompted the active development of neural network technology?

Dmitry Soshnikov: Several factors successfully overlapped here.
First, affordable computing power became available. Moreover, cloud services played a big role, since instead of investing in infrastructure for calculating neural networks, you can now rent it only for the duration of the calculations, subsequently refusing to rent. In addition, graphics processors, which were originally designed for computer graphics, began to be used. It turned out that they are well suited for artificial intelligence tasks.

Secondly, thanks to the Internet, gigantic amounts of data have begun to accumulate in many areas. Consider, for example, the previously mentioned task of recognizing a person’s age from a photograph. To train a neural network to solve it, several hundred thousand examples are required. Now you can take any social network where people themselves publish their photos and account information (age) - and we immediately receive data for training.
Thirdly, of course, some interesting research has emerged, including new neural network architectures that allow solving existing problems. But this point is probably a consequence of the first two. When resources and technologies are available, the area naturally begins to develop actively.

A large number of tools have also appeared here that allow these neural networks to be used. If earlier in order to solve artificial intelligence problems you had to have a lot of knowledge and a lot of programming, now there are services available that you can take and use.

The topic of artificial intelligence is very popular today. How much of this popularity is deserved? Is technology really that impressive or is it a big contribution from fashion? And isn’t this “fashion” detrimental to development?

Dmitry Soshnikov: There are indeed great successes in the field of AI, which are written about a lot, so the phrase “artificial intelligence” is widely heard. Thanks to this, new developers appear - someone goes and studies a new area for themselves, i.e. There are more people who understand this area. On the other hand, people are more carefully looking for those tasks where artificial intelligence technologies can be applied. From this point of view, all this is, in principle, good, because we have a chance to automate some areas that we could not automate before.

For example, we can solve the problem of accepting orders in the MacAuto window. They always try to make the solution to such problems cheaper. For example, in the USA, at first an American sat there, then there was an attempt to outsource this by transferring the voice to a country with cheap labor (where, again, the person sits and transcribes). And now a computer can do this.

Do market participants have inflated expectations? Are there any predictions that, in your opinion, will definitely not come true in the near future?

Dmitry Soshnikov: Of course have. First of all, the field of artificial intelligence is a bit romantic. There are quite a few films - for example, "The Matrix" or "Terminator" - where robots rebel and take control of everything. Therefore, there are a certain number of people who expect that another 5 years will pass and computers will take over the world. These expectations are apparently still far from reality. Nowadays, the solution of certain classes of problems related to image recognition, speech recognition, and machine learning is very well automated. But we still have quite a long way to go before understanding how human thinking works in general. Therefore, before creating such artificial intelligence that will think like a person and operate with accumulated knowledge, a lot of work still needs to be done. It is not yet very clear how to do this.

- What about expectations in terms of financial investments, rather than the implementation of science fiction scenarios?

Dmitry Soshnikov: It seems to me that for such a conversation it is necessary to break the topic of artificial intelligence into separate components, since it is a very broad area.

If we look at computer vision, there are already amazing advances that are now being implemented into businesses, increasing their efficiency and bringing economic benefits. Computer vision already recognizes images better than humans, not to mention being significantly cheaper.

In other areas, such as understanding natural language and the ability to reason about arbitrary topics, progress has been more modest.

- Are there factors that, in your opinion, hinder the development of the industry?

Dmitry Soshnikov: To be honest, I don’t see any obvious factors. I think this is the fastest growing area right now.
However, I would like to note that artificial intelligence is an area that still requires some qualifications. Working in this direction is a little more difficult than just learning to program. A person who graduates from school and does not receive a higher education can probably begin to work successfully in the field of standard development. With artificial intelligence, the bar for entry is higher, although it is gradually being lowered, including through the efforts of participants in this industry. In particular, one of the things Microsoft is working on is the so-called democratization of artificial intelligence. This means making the technology accessible to the widest possible segment of consumers.

In practice, not only Microsoft, but also many other companies are working in this direction, providing, for example, tools for solving intellectual, cognitive problems in the form of ready-made services. For example, services for determining the gender, age and mood of a person from a photograph; you can simply call them and get the result. The same applies to machine translation, etc. As part of the report at DotNext 2017, we will talk about this: how can you, without understanding at all how exactly it works, simply use the results.

A.NET maybe?

- Let's talk about the place of the .NET platform in the artificial intelligence segment. How suitable is it for solving such problems? Are there any features that help or, conversely, hinder working with neural networks?

Dmitry Soshnikov: AI methods can be implemented on any technology. Nevertheless, there are certain established ecosystems around similar tasks. For example, Python and R and their accompanying libraries are very popular languages ​​among data scientists. Those. there is already a lot of community work here. In terms of these developments, of course, .NET is a little behind, as are other similar platforms. However, .NET already has a certain set of tools, which I will talk about as part of my report.

In general, platforms are now in some sense integrated, including with each other. The same R language integrates very well with F#, which is native to the .NET platform. Accordingly, if we need to use some machine learning tools, we can use such a chain, using the capabilities and libraries of the R language. This will be quite transparent and easy to do.

In general, if we talk specifically about neural networks, Microsoft has a Cognitive Toolkit that allows you to train neural networks. And because it was originally built in the Microsoft ecosystem, it works very well with .NET.

- This tool somehow stands out from analogues from other manufacturers?

Dmitry Soshnikov: Essentially, Cognitive Toolkit is Microsoft’s analogue of the TensorFlow, Caffe, etc. frameworks.

All of them are, in principle, very similar ideologically. But Microsoft Cognitive Toolkit was the first to support a highly distributed training environment, where you can train a neural network not just on a single GPU, but on multiple GPUs or even multiple GPU stations. Those. You can make a neural network training farm.

As far as I know, Cognitive Toolkit beats other frameworks in terms of learning speed. Plus it is very convenient to use. Most frameworks are related to the Python language in one way or another, but Cognitive Toolkit initially took a slightly different path. It was initially possible to describe the architecture of a neural network in a special language, and then train it without building any models in Python. It was a little easier. Currently Cognitive Toolkit supports both options, i.e. is quite flexible.

- Probably, there are moments in which Cognitive Toolkit is inferior to its analogues?

Dmitry Soshnikov: In general, such frameworks are low-level tools on top of which arbitrary neural networks can be trained. Like its analogues, the Cognitive Toolkit supports a certain base layer on top of which you can build the architecture of networks of arbitrary complexity. Therefore, the range of tasks solved by different tools is approximately the same.

The choice of framework is largely determined by some personal preferences and the availability of reference materials. And here the Microsoft framework lags a little behind, since it appeared a little later, so there is not such a large amount of materials on it, in particular, online courses. But the situation, of course, is gradually leveling out.

We, together with MIPT, are planning to release an online course dedicated specifically to the problems associated with the use of artificial intelligence in practice. And some of the information on Cognitive Toolkit will also be included there.

Near future

- Is it possible to predict now in what direction the development of the artificial intelligence segment is going?

Dmitry Soshnikov: It’s probably too early to make final forecasts, since in 2011-2012 the extremely rapid development of technology began. Since then, recognition methods have been improved, neural network architectures have been improved, i.e. the accuracy of problem solving increases.

There are still many unresolved issues in this segment. The problems of image and voice recognition have already been solved at a fairly high level. Next, probably the most interesting thing is to try to extract some meaning from the text. There are some pretty stunning successes here too. For example, you can train a neural network on fragments of conversations from films and get a robot that can somehow maintain a dialogue. But at the same time, there will not be much meaning in the dialogue. It is not yet clear how to translate knowledge into meaning, how to combine the implicit representation of knowledge in neural networks with symbolic reasoning. This is the direction of research that scientists will pursue.

As for the tools, they are currently being actively developed. In a sense, tool manufacturers are trying to follow scientific advances in their respective field. New network architectures appear - their support appears in tools, i.e. The functionality is constantly expanding.

As I said earlier, from a developer’s point of view, there is a noticeable trend towards the democratization of artificial intelligence, including tools. In addition to the Microsoft Cognitive Toolkit I mentioned, there is an interesting tool called Azure Machine Learning, which allows you to apply them to real data without a deep understanding of the implementation of all machine learning algorithms and see if you can identify any patterns and use them further in your products. This tool is also being developed quite intensively - new methods and algorithms are being introduced there.

In general, technology is becoming more accessible. Complex things are simplified so they can be used in as wide a range of projects as possible.
Another point that I would like to mention is that these are still the first experiments in the cloud using more efficient hardware solutions that implement artificial intelligence algorithms. We won't talk about this at Dotnext, but the topic was discussed in detail at the Microsoft Ignite conference. In the cloud, Microsoft plans to offer not only classic computing resources, but also programmable logic integrated circuits: FPGA or FPGA. To simplify, these are chips that can be flashed to perform certain logical operations, and which will perform these operations very quickly. With such a scheme, we can calculate the neural network much faster. During the experiments, the processor was “prompted” to translate from language to language, as a result, the novel “War and Peace” is translated from language to language in 2 seconds. If you take all the many such processors that Microsoft has in the cloud, then Wikipedia can be translated from one language to another in the time it takes a person to blink an eye.

Dmitry will give more practical information about the use of neural networks and artificial intelligence technologies in real projects (including on .NET) in his report at DotNext 2017 Piter (“Available artificial intelligence on the .NET platform: from chatbots and cognitive services to deep neural networks").

Tags: Add tags

The Turing test is essentially a test that answers the question “can machines think?” That is, this is a set of tests that only a person or perfect artificial intelligence capable of thinking in the same way can pass. This is exactly what we are asked to do in the game - to prove that you are a Homo sapiens. In my opinion, this is a matter of principle!

Those who liked the puzzles The Talos Principle and Portal will definitely appreciate The Turing Test. These games really have a lot in common. You could even say that they form a separate subgenre, originating precisely from the Portal.

The passage is a set of logical tests that must be solved sequentially. The basis of game mechanics in The Turing Test make up energy spheres that need to be inserted into special cells to open doors, start mechanisms, etc. The only tool that helps in this matter is a certain magnetic gun that can absorb these spheres and shoot them at the right place.

There are more than 70 tests in total, but many of them are completed in a matter of seconds. Therefore, the entire passage does not take much time. In general, there are few difficult obstacles in the game; most of them can be completed almost on the fly. Some people may find the puzzles too simple, but I liked the balance of difficulty. On the contrary, I don’t like being “stupid” for a long time in such games.

The general idea is also similar to the games that inspired the developers. Philosophical and ethical considerations about intelligence, about the difference between man and machine.

As for the plot itself: the main character finds herself on a space station on the snowy surface of Europa, a satellite of Jupiter. The base staff has disappeared, and everything is controlled by artificial intelligence, which meets the heroine. She will have to pass this very Turing Test and find out what happened to people. The story is quite interesting, but it is presented in such small pieces (a couple of computer lines at the beginning of each level) that you don’t notice it for almost the entire game and only pay attention towards the end.

Can a robot write a symphony or turn a piece of canvas into a masterpiece of art? This question remains open. But here’s what we know for sure: modern medical nanorobots are able to “cut off the oxygen” of a tumor by penetrating the bloodstream and deliver drugs to the affected cells without harming healthy ones. Neural networks are about to surpass doctors in diagnosing many diseases, and examinations using CRISPR technologies can soon be carried out without leaving home. The future of medicine has already arrived - let’s find out what to expect from new methods and how this will affect our health.

Diagnose It: Artificial Intelligence

The success of treatment largely depends on quick and error-free diagnosis: for this, the doctor needs to accumulate a lot of practical experience and be aware of current scientific works in his field. But with so many new studies and case reports appearing in print every month, where can you find the time to study it all? This is where computers come to the aid of people, capable of processing huge amounts of information in seconds.

Today, algorithms for analyzing medical data are created by major corporations, including Microsoft, IBM and Google. Most often, their developments are based on various forms of self-learning artificial intelligence that can find patterns in large data sets, such as brain scans or images of suspicious growths on the skin. Such algorithms are trained using libraries of thousands of examples, where each image is assigned a diagnosis made by a qualified doctor.

Neural networks have already learned to identify many diseases as effectively as humans, and in some cases they even manage to outperform specialists.

Convolutional neural networks (CNN) allow you to analyze multiple images. These are deep (multi-layer) structures in which each artificial neuron receives only a small fragment of the output of the previous layer. Gradually, the network generalizes local features, recreating the complete picture. By combining all the data, the CNN can recognize various details in the original image, including characteristic elements based on which doctors make a diagnosis.

CRISPR technologies can not only eliminate the causes of diseases, but also identify diseases, for example, look for traces of DNA or RNA of infectious agents.

While the CRISPR-associated protein Cas9 is most often mentioned in connection with therapy, other proteins: Cas12a and Cas13a are usually the “diagnostic” proteins.

In 2017, researchers from MIT introduced a diagnostic technology called SHERLOCK (Specific High Sensitivity Enzymatic Reporter UnLOCKing). It uses the Cas13a enzyme, which can recognize specific RNA sequences and cut similar RNA strands located nearby, completely destroying the suspicious object. Like the book Sherlock Holmes, the medical SHERLOCK is capable of recreating a complete picture of events from the smallest evidence: the technology works with attomolar (10–18 mol per liter) concentrations of nucleic acids. The method was tested on lentiviruses containing fragments of the dengue fever virus and the Zika virus: SHERLOCK was able to detect pathogen particles and distinguish them from each other at a concentration of no more than two attomoles.

During testing, it turned out that reagents for diagnostics using SHERLOCK can be dried and then reconstituted, while the sensitivity of the method does not decrease much. For portable tests, it is suggested to use fiberglass paper. The authors of the development believe that one test system will cost approximately 61 cents.

Other researchers are also working on creating kits for CRISPR testing at home. Recently, developments in this area were started by Jennifer Doudna, one of the pioneers of the medical CRISPR revolution. Her team created a method called DETECTR (DNA endonuclease-targeted CRISPR trans reporter), using the Cas12a protein. It finds specific DNA sequences and cuts the nearest fluorescently tagged nucleic acid reporter molecules added to the sample, producing a signal. In this way, traces of pathogens of many diseases can be detected, including different strains of the influenza virus.

According to the creators of such test systems, CRISPR analysis will take no longer than a few hours, and the results can be obtained via the Internet. However, it is still unknown when such sets will appear on public sale.

It's an illusion: virtual reality

Virtual reality is often talked about in the context of computer games and “3D cinema,” but the technology also has potential in medicine, and not in the most obvious areas. For example, VR is effectively used as a pain reliever.

The burn unit at Loyola University Hospital in Illinois takes this approach.started used ten years ago: hospital patients play a simulator during painful proceduresSnowWorld.

The action takes place against the background of semi-fantastic northern landscapes with many snowdrifts and frozen rivers, the hero's task is to play snowballs with polar bears, penguins and snowmen. To complete all levels, the patient involuntarily focuses on the puzzle and is distracted from physical sensations. MRI brain scans have shown that SnowWorld actually reduces the perception of pain, so patients require fewer strong painkillers that can be harmful to the body.

VR is replacing or complementing painkillers in many areas of medicine. The technology is used to relieve pain during childbirth and during dental procedures. The pain-relieving properties of virtual reality are especially relevant in light of the “opioid crisis” in the United States - it is associated with the rise in popularity of prescription painkillers (for example, OxyContin and Vicodin) in recent decades.

VR works not only for physical pain: it can also overcome psychological trauma. The first experiments took place in the late 1990s, when psychologist Barbara Rothbaum managed to alleviate the symptoms of post-traumatic stress disorder in Vietnam veterans using virtual models of a helicopter cockpit and a clearing in the Asian jungle. This technique complemented exposure therapy - a gradual “approach” to traumatic memories that the patient’s consciousness tries to avoid. A similar scheme works in the treatment of anxiety disorders and phobias using virtual reality. Technology helps to cope with aerophobia and fear of public speaking: a simulated environment makes it possible to repeatedly “rehearse” a frightening situation.

The disadvantages of VR therapy are the relatively high cost of equipment and potential physiological intolerance to virtual reality.

Some participants in trials of new treatment methods experienced VR “nausea” (virtual reality sickness), which causes the same symptoms as motion sickness or seasickness. According to common hypotheses, both disorders arise from disturbances in the functioning of the receptors of the vestibular apparatus or a conflict between signals coming from the vestibular apparatus and the organs of vision.

Developers have already created several methods that can reduce VR “nausea”. For example, you can place a stationary object on the screen that the user's eyes will fixate constantly. The authors of the Nasum Virtualis technology propose to use a virtual image of the nose located in the center of the screen as a reference point. The player's eyes perceive it as their own nose, so the feeling of nausea and dizziness recedes.