The Application of Classical Conditioning to the Machine Learning of a Commonsense Knowledge of Visual Events
Year: 2013
Type: dissertation
Abstract: In the field of artificial intelligence, possession of commonsense knowledge has long been considered to be a requirementto construct a machine that possesses \nartificial general intelligence. The conventional approach to providing this commonsense knowledge is to manually encode the required knowledge, a process that is both tedious and costly. After an analysis of classical conditioning, it was deemed that constructing a system based upon the stimulusstimulus interpretation of classical conditioning could allow for commonsense knowledge to be learned through a machine directly and passively observing its environment. Based upon these principles, a system was constructed that uses a stream of events, that have been observed within the environment, to learn rules regarding what event is likely to follow after the observation of another event. The system makes use of a feedback loop between three sub-systems: one that associates events that occur together, a second that accumulates evidence \nthat a given association is significant and a third that recognises the significant associations. The recognition of past associations allows for both the creation of evidence for and against the existence of a particular association, \nand also allows for more complex associations to be created by treating instances of strongly associated event pairs to be themselves events. Testing the abilities of the system involved simulating the three different learning environments. The results found that measures of significance based on classical conditioning generally outperformed a probability-based measure. This thesis \ncontributes a theory of how a stimulus-stimulus interpretation classical conditioning can be used to create commonsense knowledge and an observation that a significant sub-set of classical conditioning phenomena likely exist to aid in the elimination of noise. This thesis also represents a significant departure from existing reinforcement learning systems as the system presented in this thesis does not perform any form of action selection.
Author: Timothy Andrew Furze
Institution: University of Leeds
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1557741327
