Ken Forbus Speaks at Microsoft Research on Bespoke Data

Prof. Ken Forbus has recently presented his research in a video, titled, "Multimodal Learning from Bespoke Data", as part of the MSR Series (Microsoft Research).

Video Summary: Big Data is receiving a great deal of attention in AI, Computer Science, and the popular press. But not all learning problems are best addressed by massive amounts of data. You do not want to have to tell a new assistant one million times how to fill out a form, for example. Cognitive science results tell us that children can rapidly learn new words, sometimes even given just one exposure. Prof. Forbus calls these situations Bespoke Data, because they rely on communication between beings with substantial knowledge to greatly reduce the number of training examples. 

In the talk, he describes how ideas from analogical reasoning and learning, jointly developed with Dedre Gentner (Professor of Psychology at Northwestern), can be used to build AI systems that learn at human-like rates. This could lead to a revolution in machine learning. For example, in a head to head comparison run by Chen Liang, a graduate student in Forbus’ group, showed that their techniques could provide state of the art performance, with one thousand times less data than traditional machine learning systems.