Full metadata record
| DC Field | Value | Language |
|---|---|---|
| dc.contributor | Department of Data Science and Artificial Intelligence | en_US |
| dc.contributor.advisor | Tan, Kay Chen (DSAI) | en_US |
| dc.contributor.advisor | Wu, Jibin (DSAI) | en_US |
| dc.creator | Chen, Xinyi | - |
| dc.identifier.uri | https://theses.lib.polyu.edu.hk/handle/200/14496 | - |
| dc.language | English | en_US |
| dc.publisher | Hong Kong Polytechnic University | en_US |
| dc.rights | All rights reserved | en_US |
| dc.title | Advanced spiking neuron models for accurate and robust pattern recognition | en_US |
| dcterms.abstract | The human brain, the most advanced intelligent system in nature, demonstrates a remarkable ability to perform complex cognitive functions and adapt to dynamic environments with exceptional efficiency and robustness. Learning from the brain's computational principles and structural organization has long been a promising direction for advancing neuromorphic artificial systems. Spiking neural networks (SNNs), which emulate the brain's spike-driven computation, stand out as a promising approach due to their sparse, asynchronous, and robust information processing. At the core of SNNs are spiking neuron models, which go beyond the simple activation units used in traditional artificial neural networks (ANNs) by incorporating brain-inspired neuronal dynamics for spatiotemporal information processing. These properties render spiking neurons a compelling foundation for building accurate and robust pattern recognition systems. | en_US |
| dcterms.abstract | Despite these advantages, the computational capabilities of current SNNs remain significantly limited compared to their biological counterparts. A fundamental bottleneck lies in the oversimplified design of spiking neuron models, which often abstract neurons as point-like units with minimal temporal dynamics. This simplification, aiming at balancing biological realism and computational efficiency, neglects the diverse neural coding schemes, complex internal structures, and rich neural modulation mechanisms observed in biological neurons, which collectively contribute to the brain's powerful pattern recognition capabilities. | en_US |
| dcterms.abstract | To bridge this gap, this thesis proposes a brain-inspired framework for advancing spiking neuron models across four critical dimensions: information representation, structural modeling, neural modulation, and evaluation benchmarks. The main contributions are as follows: | en_US |
| dcterms.abstract | First, to enrich information representation strategies of spiking neuron models, this thesis introduces a hybrid neural coding and learning framework. Taking inspiration from the rich diversity of neural coding strategies observed in biological systems, this framework comprises a "neural coding zoo" that integrates heterogeneous spiking neurons with diverse coding strategies and supports flexible layer-wise assignment to accommodate task-specific requirements. Moreover, a layer-wise learning method tailored for the effective training of hybrid coding schemes is introduced. Experiments on image classification and sound localization tasks demonstrate the superiority of the proposed framework in terms of accuracy, inference latency, energy efficiency, and noise robustness. | en_US |
| dcterms.abstract | Second, to enhance the structural complexity of spiking neuron models, this thesis introduces the Parallel Multi-compartment Spiking Neuron (PMSN). Inspired by the compartmental organization of biological neurons, PMSN models each neuron as a set of interacting substructures, facilitating effective information exchange and integration across multiple timescales. Additionally, two parallelization strategies are proposed to accelerate training and improve scalability. Across a variety of pattern recognition tasks, PMSN consistently outperforms state-of-the-art spiking neuron models in both accuracy and training speed, while maintaining comparable computational cost. | en_US |
| dcterms.abstract | Third, to incorporate biological neural modulation mechanisms into spiking neuron models, this thesis introduces TE-SNNs. Motivated by the temporal encoding in biological neurons, TE-SNNs integrate a brain-inspired Temporal Structure Encoding (TE) module that modulates neuronal activity in response to distinct input temporal structures. TE-SNNs exhibit powerful capabilities in modeling complex temporal structures, achieving state-of-the-art performance across vision, speech, and text pattern recognition tasks. Their real-world applicability is further demonstrated in autonomous driving scenarios, where they produce optimal, robust, and efficient driving strategies under highly dynamic conditions. | en_US |
| dcterms.abstract | Finally, to systematically track the research progress on spiking neuron models, this thesis introduces the Neuromorphic Sequential Arena (NSA), a comprehensive neuromorphic benchmark that offers an effective, versatile, and application-oriented evaluation framework. NSA encompasses seven real-world tasks derived from diverse neuromorphic application scenarios, each reflecting rich temporal dynamics across multiple timescales. It supports multi-dimensional evaluation across accuracy, training speed, memory footprint, and energy efficiency, providing a standardized and extensible platform for consistent comparison and future development. | en_US |
| dcterms.abstract | Together, these contributions establish a biologically grounded and systematically structured roadmap for designing advanced spiking neuron models, significantly enhancing the accuracy and robustness of SNNs in complex pattern recognition tasks. | en_US |
| dcterms.extent | xxi, 192 pages : color illustrations | en_US |
| dcterms.isPartOf | PolyU Electronic Theses | en_US |
| dcterms.issued | 2025 | en_US |
| dcterms.educationalLevel | Ph.D. | en_US |
| dcterms.educationalLevel | All Doctorate | en_US |
| dcterms.accessRights | open access | en_US |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14496