A computational account of contingency learning
Loading...
Date
2019-08
Authors
Smith, Bradley
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Associative learning is a fundamental concept that underlies the explanation of many psychological phenomena. Recent adaptations of the classic Stroop (1935) task offer insight into how associations are formed and augmented as a function of experience (Lin & Macleod, 2018). In this thesis, I collected data using word-colour contingency learning tasks that were designed to measure the ebb and flow of associations over the course of a controlled experiment. In the experiments, non-colour related words were presented in coloured text. I manipulated the contingency structure between the words and colours (i.e., words appeared in a particular colour more often than others). The participants’ task was to identify the colour of the word. Associative learning was measured by observing increasingly quick responses when a word was presented in the colour in which it is most frequently paired. I then adapted an existing instance-based account of performance, based on the MINERVA 2 model of memory, to explain these results. The theory works by storing events into episodic memory and retrieving those traces in parallel. Unlike standard explanations for associative learning, it does not represent associative links directly. The model replicated the critical behaviours in the previously published results including simple acquisition and participants’ responses to changes in the contingency structure mid-task. It also predicted behavioural differences that had not been considered in previous research. These behavioural differences were observed in the experimental data when the experiment was replicated with a larger sample size. Once the model was developed against known results, I used it to predict people’s behaviour in novel experiments; exploring how participants would behave under more extreme mid-task changes to the contingency structure and how resilient learning for a contingency structure would be when all word-colour pairs were presented equally often. The model predicted results that contradicted conventional wisdom from prior experiments. The experimental data agreed with the model’s predictions. The goal of this research was to define and articulate an instance-based theory explanation of how people come to learn and exploit the contingency structure of events in the environment. Because the theory was able to capture the critical behaviours exhibited by participants, and makes novel predictions about participants’ behaviours in other tasks, it is a viable explanation for these associative learning behaviours. Being able to replicate human associative learning behaviour with a computational model has implications for cognitive technologies. If this theory can be generalized (MINERVA has already been shown to replicate other human behaviours like memory and decision making) and implemented at a larger scale then it can be used as a more intelligent, human-like machine learning algorithm. This algorithm would be inspired by human cognition and would theoretically behave and make decisions similar to humans. This algorithm could be utilized so that computers could learn and interact with humans in a comparable fashion to how humans interact with each other. More work will be needed to confirm that the theory is a good approximation for human cognition and to implement it into a more general machine learning context.
Description
Keywords
Instance theory, Minerva, Associative learning, Stroop, Contingency learning, Computational modelling