Improvement of Transfer Learning Techniques in Activity Recognition

Abstract

Activity recognition is a research area focused on identifying human actions from a collection of sensor data. Often, solutions rely on design-time techniques that remain static after being trained once and assume that further data derives from identical sources and conditions. However, real-world applications require solutions to embrace daily living’s unpredictability and allow for exposure to unexpected changes.
Transfer learning is a growing field that can address the variability of activity recognition problems by reusing the knowledge from previous experiences to recognise activities from different conditions, resulting in the leveraging of resources such as training and labelling efforts.
While integrating ubiquitous sensing technology and transfer learning seem promising, there are some research opportunities that, if addressed, could accelerate the development of activity recognition.
This Thesis reviews the current status and challenges of off-the-shell capabilities for transfer learning opportunities in modelling activities of daily living in classification problems. It starts by introducing a novel method consisting of fusing features at the hidden level of artificial neural networks derived from different activation functions to create a more informative set of features. This method’s finding shows that the fusion of two activation functions contributes to scenarios with small quantities of labelled data available in the source domain and the absence of labelled data in the targeted domain. It then builds upon the learned experience from the previous method to investigates the benefit of feature fusion as a strategy to enhance the Learner model within the Teacher/Learner approach when the feature-space is different between domains.
​​​​​​​The Thesis concludes by presenting a framework that converges the feature fusion strategy with the Teacher/Learner approach over the active learning technique to automatise the selftraining process of the Learner models. Evaluation of these studies shows promising results towards the mitigation of the insufficiency of labelled data available by enabling the Learner model to outperform the Teacher’s performance, as presented in the Chapters within this Thesis.

Date of Award	Feb 2021
Original language	English
Supervisor	Shuai Zhang (Supervisor), Christopher Nugent (Supervisor) & Ian Mc Chesney (Supervisor)