Robot Fingertips and Human Minds

: Investigating Embodied Intelligence through Learning and Decision-Making

Abstract

Learning and decision-making are fundamental processes for intelligent behaviour in organic and artificial systems. A more complete theoretical framework for these processes is required to enable better understanding of human cognition in health and disease, and the development of embodied artificial intelligence. This thesis employs an interdisciplinary approach to this challenge, drawing on results and methodology from neuroscience, psychology, artificial intelligence, and robotics. By making a distinction between decision- making during well-practised behaviour and learning during task acquisition, these two integrated processes are dissociated.
In the first part of this thesis, an embodied, biomimetic model of tactile perception provides the foundation to investigate decision-making. A statistical, biomimetic evidence accumulation model of decision-making is developed in this context. It is shown to provide robust and accurate performance on a grating discrimination task. This classic measure of tactile acuity has been well-studied and enables comparisons between the model and human behaviour. The model is further developed by novel adaptations which implement context-sensitive learning of decision parameters. This enables optimal task performance and is a crucial feature for biological and autonomous agents.
In the second part of this thesis, a novel experimental paradigm that uses an online behavioural task is developed to investigate learning remotely. Task manipulations are designed to quantify the contributions of learning, working memory, and memory con- solidation. This task is deployed remotely on a cohort of 158 healthy adult participants to validate these manipulations, with computational models of learning used to support this. Participants were also genotyped to investigate how natural variation in dopamine receptor expression affected performance on the task.
As such, this thesis contributes towards our understanding of the mathematical and behavioural framework in which learning and decision-making can be measured, modelled, and implemented in silico.

Date of Award	1 Oct 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Matt W Jones (Supervisor) & Nathan F Lepora (Supervisor)