Nowadays, many devices can capture what is happening on camera, but what if you could record on video what still has to happen? Researchers at MIT CSAIL have created a deep learning algorithm capable of creating videos showing what is expected in the future. After intensive training (2 million videos), the artificial intelligence system began to generate frames, opposing two neural networks to each other. One created the scene by determining which objects in a given frame were moving at that time. Another, meanwhile, acted as a quality checker - it determined whether the video was real or simulated, and the artificial video was considered successful when the checking AI was tricked into believing that the footage was genuine.
However, the technology has its limitations. It cannot create video beyond a second and a half into the future, and the results are not strikingly realistic, but the technology is still good enough to predict, for example, the movement of waves on the beach or people walking on the grass.
However, if researchers manage to make predictions more realistic and longer in time, this could have far-reaching consequences. Self-driving cars will be able to predict where and how other vehicles and pedestrians are going, and security cameras will be able to mark inconsistencies in footage based on what they expect to see. The technology could also be used for day-to-day tasks like adding animation to still images or compressing video (since every video frame is no longer needed). And no matter the circumstances, predicting the future can help AI understand what's going on at the moment - it could help in any case where machine pattern recognition is essential.