An event-based prediction suffix tree

Abstract

This thesis introduces the Event-based Prediction Suffix Tree (EPST), a biologically inspired, event-based prediction algorithm. Similar to Variable order Markov Models (VMMs) and Hidden Markov Models (HMMs), the EPST creates a model based on the statistics of an input sequence to make predictions. Unlike these models, the EPST uses a representation specific to event data, defined as a portion of the power set of event subsequences within a short context window. Despite its high time and space complexity, the EPST possesses many promising properties such as fault tolerance, the capability for one-shot and online learning, simultaneous learning and prediction of multiple overlapping patterns, and explainability. The explainability property distinguishes the EPST from models with iterative learning algorithms such as those employed in Artificial Neural Networks (ANNs) and Hidden Markov Models (HMMs). The non-destructive learning method of the EPST enables links to be inferred from its training data to its predictions, making it transparent and interpretable. The EPST is also shown experimentally to be especially resistant to event noise types such as random event addition and removal. The efficacy of the EPST is examined in two distinct applications: anomaly detection during a simulated cyber security compromise on a robot and keystroke biometric authentication. In the first instance, the EPST model was shown to perform as well as anomaly detection measures designed specifically for features in the dataset. In the second application, the EPST’s anomaly detection capabilities were tested against a keystroke biometric authentication benchmark, where the model demonstrated comparable performance to the state-ofthe- art, achieving an Equal Error Rate (EER) of 0.074. Through experimentation, optimisations and extensions are developed for the EPST to address some of the challenges of practical applications. A significant limitation of the EPST is the high run time complexity associated with denser spike trains. To mitigate this, we developed optimisations such as model pruning and pre-prediction spike sampling. These techniques help to decrease complexity while maintaining performance. Furthermore, the EPST performance was found to be sensitive to small random variations in event timing, or ‘jitter’. This was addressed by including an ‘equivalence interval’ parameter, which controls the width of the time interval for which two events are considered a match. Finally, inhibition extends the original model to allow the creation of additional associations to rectify false positive predictions made by the base model.

Date of Award	2023
Original language	English
Awarding Institution	Western Sydney University
Supervisor	Andre van Schaik (Supervisor)