News and events
22 October 2014
Hypothesis testing in plastic neural systems: from cognition to machine learning and robotics
Presented By Dr Andrea Solgoggio, Department of Computer Science
- N.1.12, Haslegrave Building
About this event
Abstract: Understanding the world in terms of causes and effects is a fascinating aspect of human and animal intelligence. Two learning modalities that are linked to the emergence of primitive cause-effect concepts are called classical and instrumental conditioning. They are ubiquitous in animal behaviour and, the latter in particular, was inspirational in the foundation of a machine learning paradigm called reinforcement learning. Recent robotic experiments have recognised the complex, holistic and embodied nature of such learning mechanisms, which are often characterised by large and ambiguous information streams, active and passive learning, and unknown delays of effects, feedback and rewards. This talk explains why conditioning is gaining attention in the artificial intelligence community, what are the possible neural mechanisms that implement it, and how those can be exploited to create intelligent systems. The talk focuses in particular on a recent model, published in the journal Biological Cybernetics, that explains how neural systems can test hypotheses to learn from ambiguous information streams.