Why On GTA Trained Neural Network Cannot Be Applied In Real Tasks

CEO of Cognitive Technologies, Andrey Chernogorov about the shortcomings of the training of artificial intelligence systems on the generated data. Recently, Harrison Kinsley, a programmer from the United States, announced the Charles neural network that learns to drive without human help. The unique artificial intelligence system written in Python, not in how well she copes with its main task — it is just not so — and that the training is the universe of GTA 5 from Rockstar Games.

Those who know anything about computer games, have already guessed that Charles is not the most accurate driver. The neural network sees only a picture of a city, in the same way as a regular player. Smart algorithm learns from not the most peace-loving and law-abiding “residents”, which regularly crash, explode, leave the road and overturn. Adventures Charles is more like survival than Autonomous driving, which it would like to see in real life.

Watch as the car is under AI control famously undercuts the real players in the virtual scenic highway of Los Santos, a lot of fun, but the idea looks at this as an opportunity to train the network to “respond to the most unpredictable and exotic situations that only can occur in the driving process, using a fully Autonomous and customizable gaming environment”. A week later, with the beginning of the test the average cycle of the “reset” Charles is about 15 minutes, that is, every day the machine is broken or into the ditch about 100 times. That, however, is several times better than the initial result.

Another example of the use of the universe Rockstar for teaching artificial intelligence have demonstrated Artur filipowicz, an it specialist from Princeton University. He used in-game image intelligent learning algorithm, achieving an acceptable level of recognition of virtual characters in conditions of poor visibility (70%), fog, and potholes (75%). The researcher explained his decision by saving resources and time. “To search for road signs in the archives and to photograph them would be too long and expensive”.

A year and a half ago, a similar thought occurred to the leading expert of group of scientific research for Xerox Adrian of Haydon. On a previous job he had the opportunity to observe how Google is spending considerable sums of money on testing and debugging prototypes of the unmanned vehicle in terms of deterrent laws and strict rules applicable to security testing. This, according to him, is that tens of millions of dollars annually.

Such conditions are definitely not suitable for companies who do not have “limitless” resources of Google and Baidu, so Haydon was thinking about how whereby it would be possible to minimize the cost of the data-gathering phase for training the AI. The scientist came to the idea of using the game engine Unity accidentally. “On the Internet I came across the trailer for the latest spin-off of Assassins Creed, and for a minute really thought that watching a movie trailer, although it was, in fact, the graphics. This is the first time when CGI had me wrapped around your finger.”.

Haydon decided that since modern game engines can so easily fool people, they might be able to “cheat” and artificial intelligence. His team spent the year 2016, by teaching the neural network recognition of virtual cars, road markings and pedestrian crossings through technology deep learning. “Our ultimate objective is to teach the AI the correct recognition of the same objects in the physical world”.

Have to admit, Adrian was able to make progress, during the year, bringing the figure up to 78% of correct positives in the detection of pedestrians. With a partially hidden difficulties the AI of Palo Altos doing a little worse, the number of successful recognitions up to 70%. Nevertheless, receiving in 2016, the grant of $1.5 million for the development of the initiative, the American team is determined in the next 2-3 years to bring the accuracy to 90%, using only synthetic training data.

Most large-scale and fundamental crowdsourcing project on gamification of learning artificial intelligence now — Project Malmo from Microsoft, which offers to the specialists to conduct their experiments in the framework of the Minecraft universe. The initiative, launched in early 2016, according to the plan of curator Dr. Hoffman, should stimulate the learning process of artificial intelligence basic concepts of interacting with the virtual reality of Minecraft, with maximum number of degrees of freedom, which in the future will allow the AI to transfer the acquired experience to the real world and more efficient to learn highly specialized tasks like Autonomous driving and intelligent recognition. Unlike colleagues who think the world of Minecraft is too “pixeltown” and therefore not very suitable for teaching AI, Hoffman believes that patterns of learning and perception of artificial intelligence do not necessarily coincide with human.

“In an attempt to learn to fly we are inspired by birds and insects, however, the ultimate realization of the concept of flight in relation to a person only vaguely resembled what we peeped in nature. Computers from the very beginning “perceive” the world does not like people, so there is nothing wrong with the internal representation of the environment the AI units were formed by non-human algorithms and parameters,” she says. With certain assumptions, I can call Kinsley, Haydon and Dr. filipowicz colleagues.

The above cases demonstrate the use of the so-called “synthetic big data” for the calibration of intelligent algorithms in this field. In the General case is used for training a significant sample of structured data on the road surface, lighting, climatic conditions, static and dynamic objects, their mutual position, and so on. The better and more varied the data are, the better the AI will cope with unexpected situations.

The quality here, including the implied authenticity of the data relative to the real conditions in which they plan to apply. Speaking of “long and expensive”. In our project the preparation and approval of training videos with specific traffic situations for the AI do 90 people. The selection of data is, unfortunately, one of the most time-consuming and expensive process faced by professionals in the field of artificial intelligence.

At this stage the temptation is to delegate the selection process, replacing “real” data simulation, saving tens of thousands of man-hours. Can the computer now to take on the role of a generator of learning sample for artificial intelligence. I dont think. I say “temptation”, because I think the use of synthetic data in training intelligent systems, the maximum critical error, which in term can lead to casualties.

Synthetic data in the General case, is any production data applicable to a specific situation, which were not obtained by direct measurement. Training data for AI is formed not on real traffic situations and simulated. Sources of simulation algorithms of the game or programme, generating the necessary information. But how do they take the data.

The answer is simple. Programs are written by humans, and therein lies the main problem. When it comes to training the neural network on these data within a virtual reality, it seems very reliable, whats more, you can “nakodit” a universe where the AI can learn the correct driving within the limits of generally accepted rules of the road.

However, the ultimate goal of learning safe driving in the real world, on real roads, and the “jump” from the virtual into the real world, the synthetic data do not yet. No game, no matter how realistic she seemed keen on the player, they couldnt (and still a very long way) to convincingly replicate real road conditions. The number of degrees of freedom and, respectively, combinations of critical parameters for inorganic sampling will be limited to the diversity algorithm, which was provided.

Unfortunately, having achieved 70% accuracy in recognition of traffic signs Professor Filipovic is still very far from 99%, providing the theoretical possibility of commercial exploitation of such development. The required accuracy is achieved including through the use of integrated technologies, implying a coherent processing of information about speed, types and physical characteristics of objects entering the on-Board computer cameras and multiple types of sensors. Testing and calibration of such a hybrid hardware-software systems on the synthetic data it is not possible.

It is obvious that any, even minor differences when working with synthetic data on a virtual polygon is the inconsistency of the gradient of brightness, illumination, light angle — can lead to completely unpredictable consequences when you transfer in real terms. Feel the difference. Synthetic data, a virtual ground:

Real data: It would be desirable to automate and ramificati learning AI now, organic data in the foreseeable future will remain the most effective way to correctly train artificial intelligence, and therefore, the tasks of data collection and verification, with the direct participation of specialists will also remain relevant. Despite the significant gain in time and resource savings, the synthetic data is unrepresentative of intelligent systems, able to make decisions that affect human life and health.

However, I can assume that in the long term 7-10 years, fully Autonomous vehicles are under the control of artificial intelligence trained on the synthetic data, will appear on the streets of the city, if it is, for example, Gotham. Send your column about how our world will change, [email protected]

Leave a Reply