Artificial intelligence has learned to predict the future
The person has intuition or a sixth sense. So why not teach the car the same thing?
Researchers at the Massachusetts Institute of Technology were pazzled by this question. They set themselves the goal of teaching artificial intelligence the skills of prediction.
As a result of long work, they managed to create an algorithm that, on the basis of just one picture, is able to make a mini-video, which tells what, according to the “personal” opinion of the AI, should happen in the future. This technology is based on neural networks, machine learning and computer “vision”.
To learn how to predict the future in the frame, the neural network had to deeply analyze 2 million video clips, including a beach, sports fields, railway stations, children’s hospitals, etc. At the same time, the videos were not subscribed and were not marked with various markers that could help the AI understand what is shown in the frame. The neural network had to learn by itself without human prompts.
In 2017, this innovative technology generates one and a half second clips. So far, even the most advanced AI systems do not have the ability to predict the near future. But research groups at New York University and Facebook are actively working on this. However, these neural networks are now only producing a few frames from the future. Or the video quality is too low.
In the near future, AI prediction skills will come in handy for driving self-driving cars. They will predict the trajectory of pedestrians and other vehicles to avoid accidents. Also, this technology may well become the beginning of a revolution in the field of video compression without loss of quality. Due to predictive neural networks, there will no longer be a need to store all frames completely – at any time, the algorithm will be able to recreate them on its own, and in good quality.