Despite the capacity of artificial intelligence for learning which has enabled advanced in various technologies, AI is still not fully able to imitate our brain’s power of imagination. An AI can have a hard time forming images, sensations, and ideas if they a lack a direct input. However, this may very well change thanks to MIT researchers who developed AI who is able to imagine the future of still images by producing videos.
A team of researchers from the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT, led by Carl Vondrick, managed to create a deep learning algorithm that after examining still images of various objects, people, and environments, it creates videos showing what it thinks might happen in the future of each still image like the picture of train station expecting a train to arrive.
Vondrick also collaborated with MIT professor Antonio Torralba and Hamed Pirsiavash from the University of Maryland, Baltimore. Although they are not the first ones to try this sort of AI innovation, their algorithm adds new techniques like the capacity to process an entire scene at once.
Their work is detailed in a paper which will be presented at the Conference on Neural Information Processing Systems in Barcelona, taking place next week. They reveal that developing the AI’s capacity for imagining the future was possible after exposing the algorithm to over 2 million videos of simple activities and scenes found throughout our daily life. Additionally, all the videos were found in the wild with no label or description to help the Ai figure out what was happening in each one.
Afterward, researchers used a method known as adversarial learning to pit an AI that could generate videos featuring a probable future of a scene based on still images, with another AI that was meant to determine and discriminate between what seems real and what is a fake video.
After repeated experiments, the video generator AI could create more increasingly realistic videos of what might happen in certain images, so much so that it was able to fool the second AI. Human subjects even deemed the imagined videos to be more realistic in 20 percent more cases.
The ability to imagine the future through videos can be used for smarter video recognition of various objects in security footage, adding animation to still images, and more.
Image credit: MIT
Roxanne Briean
Latest posts by Roxanne Briean (see all)
- Former Virginia Tech Freshman Sentenced to 50 Years in Prison for Stabbing a Girl to Death - June 28, 2018
- San Francisco Woman Threatened to Call Police on Girl Who Sold Ice Water for Disneyland Trip - June 25, 2018
- Epping Woman Sentence to 20 Years of Prison for Mutilating Homeless Woman’s Children - June 12, 2018