Repository logo
 

Can a Spiking Neural Network Predict Unknown Learning Histories Solely on a Snapshot of Binary Choice Data?

Files

Thesis(13.33 MB)

Date

2023

Authors

Supervisor

Landon, Jason
Dave, Parry
Cowie, Sarah

Item type

Degree name

Doctor of Philosophy

Journal Title

Journal ISSN

Volume Title

Publisher

Auckland University of Technology

Abstract

Behaviour is influenced by past experiences, allowing us to make predictions based on current behavioural patterns. The aim of this doctoral project was to predict (more precisely, to retrodict) unknown learning histories based on small-sized datasets derived from binary choices made by pigeons. These choice behaviour datasets were extracted as discrete events from published research papers and each data included a 5-second window of observable behaviour after the delivery of a reinforcer. A spiking neural network (SNN) with a single Leaky Fire-and-Integrate neuron was developed to process these data and generate retrodictions of learning histories. Experiment 1 showed that retrodictions can be made by approaching behavioural data differently, without manipulating the reinforcer-behaviour relationship. Pigeons' binary choices provided sufficient information to the SNN model, which successfully matched the manual analysis of their actual choice behaviour patterns. The SNN's ability to make retrodictions relied solely on the pigeons' unique response patterns and the SNN's learning capacity. Retrodictions were successful even when the SNN was tested with new samples of various sizes from the same datasets. Thus, the SNN model demonstrated its capability to learn and make accurate predictions from behavioural data. In Experiment 2, the effectiveness of the SNN was validated by comparing its performance with that of other artificial neural networks. Three deep learning models were developed. The retrodictions made by these models were then compared to the performance of the SNN. The results showed that all models were able to accurately retrodict the pigeons' learning history. However, when additional performance measures such as F1 and precision were taken into account, the SNN outperformed all other deep neural network (DNN) models. Experiment 3 showcased the SNN's capability to work with novel small-sized datasets consisting of choice behaviours of other pigeons, both individuals and groups, who had slightly different learning histories. By recalculating the SNN's firing rates in a personalised manner, better predictive performance was achieved compared to conventional approaches used in generalisation tests, despite variations in the pigeons' learning histories. To confirm the method's reliability, Experiment 4 involved retraining the SNN with new datasets by using two transfer-learning techniques (fine-tuning and feature extraction) and then testing it on small new datasets. Both strategies yielded robust retrodictions, demonstrating the advantages of the applied methodologies. This thesis may become a valuable addition to the behaviour-analytic toolbox by providing a tool capable of retrodicting unknown learning histories from limited behavioural datasets containing a small window of binary choices. Collectively, the experiments demonstrate that the SNN is an effective tool for understanding the connection between learning histories and behaviour in behavioural research. The SNN exhibited adaptability and responsiveness to relatively small amounts of data from observable behaviour and produced retrodictions, thus demonstrating its potential to replace labour-intensive manipulations of the reinforcer-behaviour relationship or lengthy common training procedures. It makes personalised predictions possible and facilitates the study of differences in individual learning patterns shedding light on the relation between learning history and behaviour. By retrodicting learning histories, this work establishes a foundation for exploring the use of new training methodologies, using optimal training conditions tailored to individual organisms and specific learning tasks.

Description

Keywords

Source

DOI

Publisher's version

Rights statement

Permanent link

http://hdl.handle.net/10292/16704

Collections

Doctoral Theses