DeepMind AI learns physics by watching videos that don't make sense

Teaching artificial intelligence to understand simple physics concepts such as that one solid object can't occupy the same space as another could lead to more capable software that takes less computational resources to train.

The company has created artificial intelligence that can beat the best players at chess and Go, write computer software and solve a puzzle. These models are very specialized and do not understand the world. Something fundamental is still missing according to the researchers at DeepMind.

Luis Piloto and his colleagues at DeepMind have created an artificial intelligence called physics learning through auto-encoding and tracking objects that is designed to understand the physical world.

Simulation videos of objects moving as we would expect, such as balls falling to the ground, rolling behind each other and bouncing off each other, were used to train PLATO. The data they gave to PLATO showed the exact coordinates of the objects in the frame.

To test PLATO's ability to understand physical concepts such as persistence, solidity and unchangingness, the researchers used another series of simulations. Some showed objects obeying the laws of physics, while others showed nonsensical actions, such as a ball rolling behind a pillar, not emerging from the other side, but reappearing from behind another pillar further along its route.

They asked PLATO to predict what would happen in each video, and found that its predictions were usually correct, but not always.

Read more: DeepMind has made software-writing AI that rivals average human coder

According to Piloto, the results show that an object-centered view of the world could give an artificial intelligence more flexibility. He says that an apple could be in many different scenes. There is no need to learn about an apple on a tree or an apple in the garbage. You are in a better position to know how the apple behaves in new systems if you isolated it as its own thing. Learning efficiency is provided by it.

The work could lead to new avenues of artificial intelligence research, and may even reveal clues about human vision and development. He was concerned that the paper said that the implementation of PLATO is not viable.

He says they are using an architecture that other people can't use. It is important that other people can get the same results so that they can take them to the next level.

Computational requirements for training and running artificial intelligence models could be lowered by the findings.

He says it's similar to teaching a child what a car is by first showing them what wheels and seats are. The benefit of using object-centered representation, instead of raw visual inputs, is that it makes it easier for artificial intelligence to learn physical concepts.

Nature Human Behavior was published in the journal.