DT-SCNN: dual-threshold spiking convolutional neural network with fewer operations and memory access for edge...

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Spiking convolutional neural networks (SCNNs) offer high accuracy and power efficiency for visual tasks on neuromorphic hardware, making them suitable for edge applications.
• Implementing SCNNs on resource-constrained edge devices is challenging due to extensive convolutional operations and membrane potential (Vm) storage requirements.
• Existing optimization methods like timestep reduction, pruning, and quantization overlook redundancy in spiking feature maps (SFmaps).

Purpose of the Study:

• To introduce a dual-threshold spiking convolutional neural network (DT-SCNN) that reduces computational operations and memory access by exploiting similarities within SFmaps.
• To decrease the number of operations and memory footprint by halving Vm and convolutional weight storage.

Main Methods:

• The proposed DT-SCNN utilizes dual firing thresholds to generate two similar SFmaps from a single Vm map.
• A modified spatio-temporal backpropagation (STBP) training method with a two-stage strategy is employed to train DT-SCNNs for single-timestep inference.
• The approach focuses on reducing redundancy in SFmaps to optimize SCNN implementation.

Main Results:

• The DT-SCNN achieves a 50% reduction in operations and data storage for convolutional layers compared to standard SCNNs.
• Accuracy loss is minimal, not exceeding 0.4% on CIFAR10, MNIST, and Fashion MNIST datasets.
• The DT-SCNN demonstrates the lowest operational count among previous works due to its lightweight design and single-timestep inference.

Conclusions:

• The DT-SCNN effectively addresses the computational and memory constraints of implementing SCNNs on edge devices.
• The dual-threshold mechanism and optimized training significantly reduce resource requirements without compromising accuracy.
• This work paves the way for developing low-latency, power-efficient AI applications for edge computing environments.