Small children understand what will happen if you turn a glass of juice. But no car. Despite the variety of contemporary computer algorithms are not able to predict the consequences any of his actions. Of course, if this computer is not specifically to teach.
A group of researchers from the Institute for artificial intelligence Paul Allen (Allen Institute for Artificial Intelligence (AI2) has developed a program that helps a weak form of AI to “understand” what happens when you perform a particular action. “Predicts” the future, showing how you may behave this or that object under certain conditions. This, scientists say, will help the AI make fewer mistakes. For example, the Autonomous vehicle will be able to “predict” the consequences of the development of a given situation on the road.
The system developed Ruzbehan Mottaghi (Roozbeh Mottaghi and his colleagues, shows a number of objects. In the software platform created by scientists who use machine learning and 3D simulation. Researchers have converted more than 10,000 images in the volumetric scene. For this purpose we used a specialized 3D engine.
After conversion of the original image and their bulk analogues were loaded in the neural network. As a result of computer system is gradually studied on the example of simple 3D objects and movements of these objects. After the system has received a certain amount of steam “image — 3D model”, she is able to suggest what forces can be applied to three-dimensional object, and what the consequences might be manifestations of these forces.
Not to say that the work went very smoothly. But situations in which the computer can predict the behavior of three-dimensional object when interacting with it, was more than situations, when the machine is not understood that can happen. The AI “understand”, for example, that if the stapler lying on the edge of the table to push, he will fall. The system also successfully demonstrated the situation with a coffee table and sofa. The AI could “understand” that if the coffee table moving to the sofa, the table will rest on the sofa, and then could not move.
“Our goal is to study the dynamics of the physics engine. You must learn to predict the possible behavior of objects represented by scene,” says the project Manager.
The results of this work can be useful for many areas. So, not everywhere applicable method of trial and error. In the case of robomobile it is absolutely impossible. The computer must predict the possible development of the situation, and do it very quickly. The collection and analysis of data in such situations is very important. You can give another example. System customer service in the store also can not push from the shelves of products, in order to understand the consequences of their actions. This will cost the owner of the shop and scare away customers.
The work done by scientists, is part of project Plato.His goal is to give robotic systems ability to understand the consequences of their actions without testing of this in practice. In particular, systems with a weak form of AI, which is used in the project, understand what will move the skier going down the mountain. They understand how to move a soccer ball that you just sent in flight footballer. Such events and their consequences, the AI needs to learn to understand in real-time.
In recent years, computer systems markedly “wised up”. They can analyze the images, followed by category, identification of image elements and tagging. Brendan lake Brendan Lake), a specialist from new York University, believes that the project Ai2 is important for the modern world. “Understanding scenes is much more complicated than object recognition,” says lake. “When a person sees the frame of any scene, he is able to tell the whole story about what is happening or what could happen on the frame. Ideally, the machine should be able the same.
Of course, while that person greatly exceeds the capabilities of the machine in the example. But the task of scientists is to teach a computer system to analyze the possible consequences of their actions. That computers are in this sense equal the man or superior to him, while there is no question. But at this stage it can be very useful for a number of areas.
Scientists participating in the project do not hide the results of their work. Source code, dataset, and everything else you can get at this link for self-study.
