Date: 22 – November – 2023 @ 03:00 PM – 04:00PM
Location : DT 2203 & online
Title: Hebbian Learning and Predictive Processing as the Foundations for a Computational Implementation of the Common Model of Cognition
Speaker : Mary Kelly
Abstract:
The state-of-the-art in artificial intelligence (AI) has been advancing rapidly, such that many problems deemed impossible a decade ago can now be performed by AI at an expert-level, from winning at complex games to creative activities like art, writing, and conversation. It is the role of the cognitive scientist to ask critical questions of the recent developments in artificial intelligence: 1. Do humans and current AI behave similarly? 2. In what ways do humans and current AI differ? 3. Where and why do current AI and humans differ? 4. How can the differences be bridged? Our focus is on the last question: Where there is a gap between current AI and human behaviour, can we bridge such a gap by exploring novel, computational, cognitive architectures? We are working towards one possible answer in the form of a new cognitive architecture built on two neurobiologically plausible variants of Hebbian learning, namely, predictive processing (the theory the brain learns by correcting predictions) and vector-symbolic models of human memory. The use of these particular building blocks facilitates online learning across tasks without the catastrophic forgetting characteristic of traditional approaches to training neural networks. We provide an overview of the architecture, discuss preliminary results on maze-learning tasks, and detail future directions. Our work is intended as a blueprint for designing a modern, deep neural cognitive architecture capable of learning to complete arbitrary tasks, online, in a manner more similar to the brain than current AI approaches.
Bio:
Dr. Mary Kelly is the principal investigator of the ANIMUS lab and an Assistant Professor in the Department of Cognitive Science. She is a cognitive scientist with a background in machine learning, memory and learning, psycholinguistics, and computational linguistics. Her research has two goals: (1) to advance the scientific understanding of the basic cognitive functions that underpin human learning, knowledge, and language acquisition, and (2) to develop biologically-inspired machine learning systems capable of achieving expert performance on arbitrary tasks through learning.