Trustworthy and Efficient Knowledge Sharing across Tasks in Deep Neural Networks

Deep neural networks (DNNs) have revolutionized various fields with their superior performance on particular tasks. However, these tasks in real-world applications are often interrelated, raising the necessity for DNNs to share and transfer knowledge reliably and effectively between tasks. To solve this problem, we first develop a human understandable NeuroChains from the pre-trained model for each task, aiming to illuminate the reasoning process of the model for the given task and to foster trust in the shared knowledge. Next, based on the interpretable sparse network of each task, we enhance the efficiency of DNNs by leveraging the shared knowledge of existing tasks, thereby enabling the rapid and reliable creation of well-performing sparse models for new tasks. Finally, we further investigate the knowledge sharing of parameters in DNNs when tackling sequential tasks. We propose a novel approach to mitigate forgetting and enhance the fairness of DNNs across tasks. We conduct extensive experiments to demonstrate the efficacy of our proposed methods on convolutional neural networks and vision transformers across multiple datasets. Our proposed methods consistently outperform popular baseline methods in terms of accuracy and efficiency.

Author dc:creator

• Zhao, Haiyan

Statement dc:rights

• info:eu-repo/semantics/openAccess
• The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
• © 2023 Haiyan Zhao
• au.edu.uts.lib/cph

Language dc:language.iso

en_US

Handle dc:identifier.uri

http://hdl.handle.net/10453/173604

OAI identifier oai:identifier

oai:opus.lib.uts.edu.au:10453/173604