MIT Algorithm: AI Can Predict Immediate Future Using Still Images, Here's How
(Photo : Joe Raedle/Getty Images)
There are several mysteries associated with the future. It has captured the attention of civilizations since time immemorial. However, researchers at Massachusetts Institute of Technology (MIT) Computer Science and Artificial Intelligence Laboratory (CSAIL) have devised an incredible system. The MIT algorithm uses still images to predict the immediate future in the form of a short video.
While distant future cannot be predicted accurately, humans have a knack of determining immediate consequences of our actions. For example, if we make a sudden move in front of our pets, they will move backward. We have become extremely adept at such quick deductions of cause and effect. This is because we have the capacity to learn from repetitive actions. A similar activity can be difficult for artificial intelligence to conclude. Will the pet react or not? Will it run away or move back?
The system developed by MIT is quite revolutionary because of its simplicity, accuracy and method. The MIT algorithm is deep-learning in nature and is based on 2 million videos. These videos are almost equivalent to a year-long footage data.
This idea has been floating around the tech world for quite some time. However, the innovative aspect of the MIT algorithm is that it is capable of projecting new videos not seen before. An entire scene is processed in a go, unlike its predecessors which used to generate it in a frame by frame manner. This helps in reducing major errors in prediction. The creators believe that building an entire scene at once is more accurate, albeit more complex.
How MIT Algorithm Works-
The MIT algorithm has been programmed to distinguish background from foreground images and moving from still images. Based on an image fed to it, the system creates short videos of all possible immediate events. Thereafter, adversarial learning helps in filtering videos to choose the best possible outcome.
"The idea behind adversarial learning is to have two neural networks compete against each other," Carl Vondrick, Ph.D. at MIT CSAIL and lead author of the paper told Digital Trends. "One network tries to decide what is real versus fake, and another tries to generate something that fools the first network."
The results? Human subjects testing the MIT algorithm were fooled 20 per cent more often than the most basic model. However, these videos are quite short in duration in order to make them consistent. Longer videos may require human intervention at regular intervals. The creators want to use it for determining short term possibilities with possibilities of long-term artificial intelligence assistance.
Stay tuned to SWR for more updates and latest news on MIT Algorithm.