Skip to main content

Behavioural Genetics inspired framework for evolving populations of artificial neural networks

Venue: Birkbeck Main Building

Neuroevolution (NE) is a nature-inspired approach for creating artificial intelligence. Maitrei Kohli will describe collaborative research at Birkbeck's Departments of Psychology and Computer Science using principles of behavioural genetics to evolve artificial neural networks (ANNs) capable of exhibiting intelligent behaviours. The first phase of work involved designing and implementing a hybrid NE approach that evolves ANNs' general 'ability to learn' and combines evolution and learning within a single framework. In the second phase of work, the neuroevolutionary approach was adapted to model children's acquisition of English past tense verbs and to capture individual differences. Finally, the approach was extended to transfer learning with a special focus on heterogeneous transfer scenarios. The transfer model used ANNs as computational models capable of learning heterogeneous tasks in an evolutionary setting. From a neuroevolution perspective this represented a scenario wherein the population members are capable of learning tasks different from those they have been selected for. Overall, the hybrid framework and its applications help in understanding and synthesising the evolutionary pressures (genetic or environmental) leading to high-level intelligence, scaling the neuroevolutionary approach to evolve cognitive behaviours such as language acquisition, and also enabling lifetime learning.

Contact name: