AI Creates Videos of Future Events

Set 38 years in the future, the plot of 2002’s blockbuster film Minority Report revolves around Washington DC’s PreCrime unit, a police force who able to stop future murders from happening with the aid of three mutant human who are able to predict homicides before they happen.  Minority Report managed to side step the “psychic predicts a murder” cliché storyline with its innovative use of technology: not only could precogs predict future murders, but their visions could be streamed via a neural bridge in the form of a video that the police officers could watch. Fantastical? Nope, and researchers from MIT already have a jump on the technology.

Using a deep learning algorithm, scientists at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have crafted an ai system that is able generate short videos from a static photograph showing what should have happened in the following second and a half after the photo was taken. Trained on over 2 million videos, the system is composed of two neural networks. One network creates the actual video by picking out the dynamic elements of a photo and animating them based on the knowledge it has of the subjects and their motion. The second neural network is quality control. It watches the output from its counterpart, and if it believes the video is artificial, it rejects it, and tells the first neural net to try again.

As a proof of concept, the system is amazing, but it is clearly not ready for prime time. It can only generate videos that are 1.5 seconds long, and even with the quality checker in place, the results aren’t very realistic. The system does, however, excel at identifying the elements of a photo that are in motion and how they should move. This predictive ability coupled with machine vision will likely play a crucial role in the evolution of robotics and in particular, autonomous vehicles who need to know ahead of time how pedestrians and other vehicles are likely to behave.

Read more at MIT News: http://news.mit.edu/2016/creating-videos-of-the-future-1129