Exploring Optimized Spiking Neural Network Architectures for Classification Tasks on Embedded Platforms.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Spiking neural networks (SNNs) offer low power consumption on event-based neuromorphic hardware, particularly with sparse data.
• Training deep SNNs remains challenging, with existing Artificial Neural Network (ANN) to SNN conversion methods requiring excessive time-steps and yielding suboptimal performance.

Purpose of the Study:

• To propose and evaluate customized VGG and ResNet architectures for training deep convolutional SNNs.
• To reduce the number of time-steps required for SNN training using surrogate gradient descent.
• To address overfitting issues in SNN training through a refined dropout technique.

Main Methods:

• Developed customized VGG and ResNet architectures for deep convolutional SNNs.
• Employed surrogate gradient descent backpropagation for training.
• Implemented a refined SNN-based dropout technique to mitigate overfitting.
• Conducted experiments on embedded platforms (NVIDIA JETSON TX2) for performance validation.

Main Results:

• Achieved good classification results on public (CIFAR-10, MNIST, SVHN) and private datasets (KITTI, Korean License plate).
• Demonstrated feasibility of proposed customized SNN models and training techniques on embedded platforms.
• Validated performance in terms of processing time and inference accuracy compared to PC.

Conclusions:

• Customized SNN architectures and training methods, including fewer time-steps and refined dropout, enhance performance.
• The proposed approach is effective for deploying SNNs on embedded systems for real-world applications.
• This work contributes to efficient and high-performing SNN training and deployment.