Files and links (1)
PDFOpen Access (License Unspecified), Open Access
Dissertation
Improving neural network performance with bio-inspired plasticity and unraveling interpretability with information entropy
Doctor of Philosophy (Ph.D.), Drexel University
May 2025
DOI:
https://doi.org/10.17918/00010978
Abstract
Spiking Neural Networks (SNNs) are the next-generation neural networks that mimic how biological neurons communicate with spikes (electric pulse). They are essential for understanding the brain and developing advanced artificial intelligence. This thesis aims to enhance the performance of SNNs, particularly in capturing temporal information that changes over time. Current SNNs are inefficient in processing temporal information due to the limitations of their plastic parameters in the time domain, which are modified for learning. To address this issue, we propose a new type of plastic parameter called reference spikes, which provides reference information distinct from the input. Reference spikes help neural networks process complex temporal information by modulating network activities temporally on a detailed level. Experiments show that reference spikes enable SNNs to distinguish temporal signals from noise with unsupervised learning and improve their memory and classification accuracy with supervised learning. Our second work is to enhance the interpretability of neural networks. Currently, the intermediate output of modern neural network models is difficult for humans to understand. However, the middle output from concept neurons in the human brain presents semantic information that is interpretable to humans. Inspired by biological concept neurons, we proposed a new method (loss function) called bi-scale entropy to generate simple and interpretable neural outputs called concepts. Each concept clusters objects into two classes and is distinct from the other concepts. Experiments show that the concepts calculated from the several datasets hold information related to labels and show clustered and independent distribution of objects, making them easier to interpret. Minimizing bi-scale entropy may be a general method to improve the interpretability of neural networks.
Metrics
21 File views/ downloads
10 Record Views
Details
- Title
- Improving neural network performance with bio-inspired plasticity and unraveling interpretability with information entropy
- Creators
- Zeyuan Wang
- Contributors
- Luis R. Cruz Cruz (Advisor)
- Awarding Institution
- Drexel University
- Degree Awarded
- Doctor of Philosophy (Ph.D.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xii, 155 pages
- Resource Type
- Dissertation
- Language
- English
- Academic Unit
- College of Arts and Sciences; Physics; Drexel University
- Other Identifier
- 991022057738504721