Title:
Design and Optimization of Heterogeneous Feedforward Spiking Neural Network For Spatiotemporal Data Processing
Design and Optimization of Heterogeneous Feedforward Spiking Neural Network For Spatiotemporal Data Processing
Author(s)
She, Xueyuan
Advisor(s)
Mukhopadhyay, Saibal
Editor(s)
Collections
Supplementary to
Permanent Link
Abstract
Over recent years, deep neural network (DNN) models have demonstrated break-through performance for many computer vision applications. However, such models often require a large amount of computation resources to operate, creating limiting factors for energy-constrained hardware platforms. Training DNN models also requires a high amount of labeled data, which could be difficult or expensive to acquire.
The biologically inspired model of spiking neural network (SNN) is another type of network that is capable of processing computer vision data, and has the potential to achieve higher energy-efficiency than DNN due to its event-driven operations. SNN also has the capability to learn with biologically inspired algorithm that does not require training labels, i.e. unsupervised learning. While good performance has been shown for datasets with spatial-correlation, such as those in image classification tasks, the accuracy of SNN is still below that of DNN when the dataset has a higher level of complexity. This includes spatiotemporal tasks such as video classification and gesture recognition.
In this research, we tackle this problem by proposing a design methodology for feedforward SNN that can be trained with either biologically inspired unsupervised learning algorithm or supervised statistical training algorithm, to achieve spatiotemporal data processing. The proposed model shows performance that is parallel to or better than DNN when the amount of labeled training data is limited. We derive theoretical analysis for the proposed design to help optimize network performance, and demonstrate with experimental results that the proposed design can achieve improved performance while using less trainable parameters. For event-based spatiotemporal data, we demonstrate that the efficiency of the proposed network can be further improved with a fully event-driven processing method.
Sponsor
Date Issued
2022-01-31
Extent
Resource Type
Text
Resource Subtype
Dissertation