Look at these blocks. Do you think they’re going to fall?
That simple question drives a new experiment from Facebook’s artificial intelligence research lab – an attempt to create software that can observe a simple version of the world and predict what will happen.
Facebook’s researchers wanted the software to intuit what was about to happen in the physical world just like humans do, not to make judgements based on rules written by engineers. “Making such judgements does not require us to invoke Newton’s laws of mechanics – instead we rely on intuition, built up through interaction with the world,” they write. What if the computer, like an infant, could play with wooden blocks and watch how they moved?
First, the researchers created 180,000 computer simulations of two to four coloured blocks stacked in random configurations. They also took 493 videos of wooden block towers, filmed as they fell or stayed in place.
What happens next?
Some of the simulations and videos were fed to artificial. These are software models that can learn and change based on the data they processes, the basis of modern artificial intelligence. After this training, the networks were presented with a previously unseen scenario and made a prediction as to whether the blocks would fall.
The best neural net accurately predicted the fall of the simulated blocks 89.1 per cent of the time. The AI fared less well with real blocks, with the best system only getting it right 69 per cent of the time. That was still enough to outcompete human guesses on the same data for the virtual blocks, and to be level with humans for predicting what would happen to real blocks. The AI also gets good results on problems involving numbers of blocks it hasn’t seen before.
points to Facebook’s bigger dreams for artificial intelligence. , company AI director Yann Lecun said that one of his goals is to build algorithms that can make common-sense predictions “by just observing the world”.
More on these topics:
This entry passed through the Full-Text RSS service – if this is your content and you’re reading it on someone else’s site, please read the FAQ at fivefilters.org/content-only/faq.php#publishers.