These Machines are Learning to use Their Imagination

0
357
To be as effective as possible at specialized tasks and everyday interactions, intelligent computers of the future will need to be able to correctly project and predict logical next steps.
charles taylor | Shutterstock.com

Interestingly, creating the kind of super intelligent robots we see in science fiction will, as Facebook’s head of AI Yann LeCun notes “require knowing everything about the world and human nature.”

Last summer, Google’s AlphaGo won against the game’s grandmaster in a demonstration of advances made in machine learning. In the fixed context of game rules, it is easy for a computer to logically predict the next best move.

Evidently, we are learning how to create smarter computers, and smarter computers are starting to learn like us.

But how does AI fair with the very variable rules of life?

For a person, it’s pretty easy to guess what might happen next if you see a girl in a bikini standing on a diving board. Based on what we know (i.e. our past experiences, direct or indirect) the logical conclusion is that the girl will jump into the pool.

For most types of AI, it’s still difficult to make accurate predictions based on real-time events, like figuring out who the killer is in a “whodunit” murder mystery.

This is because machines don’t necessarily have past experiences upon which to draw a conclusion.

A team of MIT researchers, however, is working on a solution by training AI with neural networks and videos.

Phase I: Passive Absorption

The researchers began by basically letting the AI binge watch 2 million videos, effectively exposing the AI to the necessary experience and information it would need in order to predict what will happen next. Think of the scene from “The Matrix” where Neo receives his combat training:

Morpheus: How is he? Tank: Ten hours straight. He’s a machine.
Neo: I know Kung Fu. Morpheus: Show me.

 

Phase II: Show Me

The team continued by feeding the computer a sequence of still photos, like our previous example of the girl in a bikini on the diving board.

Then, the asked the AI to predict the next frame, essentially asking it to complete the photo sequence by predicting what might happen next.

Amazingly, the computer succeeded in synthesizing what it knows from the videos with the still photos, and was able to compile the images to create short (and choppy) videos as its prediction of what will happen.

For example, when shown an image of a train, the computer predicted that the train would continue moving past the “camera”– just as a human might guess.

But, it’s predictions weren’t perfect; Sometimes, the AI was spot on, and other times it failed miserably.

At any (frame) rate, the results by the MIT researchers are encouraging. To be as effective as possible at specialized tasks and everyday interactions, intelligent computers of the future will need to be able to correctly project and predict logical next steps.

Interestingly, creating the kind of super intelligent robots we see in science fiction will, as Facebook’s head of AI Yann LeCun notes “require knowing everything about the world and human nature.”

banner ad to seo services page