STATES CHRONICLE – Facebook is making impressive advancements in artificial intelligence software, reaching the point where the software can be trained to know whether a block of virtual blocks will fall over.
Researchers at the company are interested if AI software can learn basic physical common sense – the kind humans learn since birth. Understanding, predicting and communicating with the world is tightly connected to the way humans are able to grasp the basic concepts, such as the fact that a larger object won’t fit inside a smaller one, or that unsupported objects fall.
According to Mike Schroepfer, Chief Technology Officer at Facebook, machines need to adopt a similar common-sense understanding, if they are to become more useful. Results of the project will be officially shared at the Web Summit in Dublin, Ireland this week.
Through games and observation, humans start understanding the basic physics of reality at a fairly young age. In a similar manner, Facebook researchers built a system based on its image-processing software that can learn and predict the fate of a stack of virtual blocks – tumble or no tumble.
A large number of images of virtual stacks were uploaded into the software’s memory, in addition to two stereo images mimicking the ones formed in a pair of eyes. During the first phase, the AI had to observe the many different stacks it was shown, and see which ones toppled and which ones didn’t.
The learning software was told during the simulation which ones did which. With enough examples, it was able to start predicting on its own with 90 percent accuracy if a certain stack was bound to tumble.
A New York group of Facebook researchers are behind the project, a team that experiments with deep learning – designing software that can understand language and images. According to Yann LeCun, head of the group and a professor at NYU, explained that the system that can predict when blocks will topple serves as baseline for more complex physical simulations.
The memory network is a previous system LeCun’s group had developed, one that can pick up some verbal reasoning skills and basic common sense through story reading. The memory network is the one powering M, Facebook’s newborn virtual assistant which has topped the performance of Siri or any other similar apps.
Seeing that human operators are now working behind stage for M, the AI software is in the learning phase. So far, M knows it has to ask two questions when someone asks for flowers to be delivered: “Where should I send them?” and “What’s your budget?” In the future, M will field queries for itself, relying less on its human team.
Image Source: Flickr