Designing a spiking neural network system for object recognition

Abstract

Widely touted as the 3rd generation of neural networks, Spiking Neural Networks were introduced as an enhanced representation of working memory. As compared to Deep Neural Networks that comprise of neurons with continuous-valued activation functions, Spiking Neural Networks demonstrate exceptional biological plausibility as discrete-valued spikes with rich information content are used. An innovative brain-inspired computational model, Spiking Neural Networks have demonstrated high energy efficiency and remarkable processing capabilities of spatiotemporal information. However, the intricately discontinuous mechanism and complex neural dynamics have posed a significant challenge in formulating efficient learning algorithms for training Spiking Neural Networks.

In this project, a detailed study was performed to understand the motivations and considerations taken by research teams to develop novel innovations throughout the years. Next, a comparison of the performance of state-of-the-art frameworks on benchmark datasets was conducted. Thereafter, an analysis was conducted to understand more about the conventional and neuromorphic datasets that are commonly used by related works. Subsequently, a total of 6 separate proposals were made and the extent of their contributions to the performance of the baseline model was tabulated. In addition, 6 sets of experiments were performed to analyse the compatibility and synergy between these enhancements. Through these experiments, it was demonstrated that the different combinations of enhancements were compatible with each other. In particular, the final overall combined model was shown to outperform present state-of-the-art models and attain remarkable accuracy rates in classification tasks.

Furthermore, the Django web framework was used to develop an interactive and secure web application that incorporates the overall SNN model. This application seeks to provide users with a seamless experience in uploading both neuromorphic event streams and frame-based images for subsequent classification.

This report contains an in-depth literature review of both classical works and recent state-of-the-art frameworks, elaborations on the underlying principles for the proposed enhancements, test results, explanations on how the improved SNN model was integrated with the web application, design considerations for the web application and a detailed guide on the process flows for using the web application. This report concludes with an analysis on the final results and plausible future extensions.