Implementation of Field-Programmable Gate Array Platform for Object Classification Tasks Using Spike-Based...

Area of Science:

• Artificial Intelligence
• Neuromorphic Engineering
• Computer Vision

Background:

• Deep convolutional spiking neural networks (DCSNNs) offer a promising low-power alternative for edge AI applications.
• Training DCSNNs effectively, especially on hardware platforms like FPGAs, remains a significant research challenge.
• Spike-based backpropagation techniques are emerging as viable methods for training complex SNN architectures.

Purpose of the Study:

• To compare the efficacy of Temporal Spike Sequence Learning via Backpropagation (TSSL-BP) and Surrogate Gradient Descent via Backpropagation (SGD-BP) for training DCSNNs.
• To evaluate the performance and feasibility of deploying trained DCSNNs on low-power Field Programmable Gate Arrays (FPGAs) for object classification.
• To provide insights into the limitations and advantages of DCSNN deployment on FPGAs for researchers and industry.

Main Methods:

• Implemented and compared TSSL-BP and SGD-BP for training DCSNNs with convolutional filters.
• Developed a low-power FPGA board for deploying DCSNNs.
• Inferred TSSL-BP and SGD-BP models on the FPGA for object classification using public (MNIST, CIFAR10, KITTI) and private datasets (INHA_ADAS, INHA_KLP).

Main Results:

• Comparative analysis of TSSL-BP and SGD-BP performance across various datasets and network architectures.
• Evaluation of DCSNN inference accuracy and efficiency on the FPGA platform.
• Assessment of power consumption and configuration requirements for FPGA deployment.

Conclusions:

• Identified the most effective spike-based backpropagation technique for deeper DCSNNs on FPGAs.
• Demonstrated the feasibility of deploying trained DCSNNs on low-power FPGAs for object classification.
• Provided practical considerations for optimizing DCSNNs for edge AI hardware.