Josh Bongard
Computational Synthesis Lab
Sibley School of Mechanical and Aerospace Engineering
Cornell University
Date:
Friday March 16, 2006
Time:
Location: 101 Perkins
Abstract
Artificial intelligence and machine learning have traditionally viewed computation as a passive process: a robot or learner receives some data from the outside world, transforms it, and then produces output. A common theme running through my research, active computation, inverts this view: computational structures actively explore their environment in order to gain more information about it, and, in turn, are improved by this exploration. I will present several applications of this concept, and demonstrate how it is a generalization of active learning, in which unlabelled training data is proposed to an oracle. I will present three applications from robotics, cognitive science and machine learning. The first algorithm automatically designs the bodies and brains of virtual robots in simulation, by combining ideas from biological development and evolution. The second algorithm runs onboard a physical robot and allows it to autonomously synthesize models of itself and its environment, using self-created exploratory actions. The third algorithm performs grammatical inference, and outperforms the best heuristic currently reported in the literature. Other applications-in-progress from bioinformatics, functional anatomy, civil engineering, psychology and medical diagnosis will be touched upon. Finally, I will situate these various projects within my long-term research plan.