Direct training high-performance spiking neural networks for object recognition and detection.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Spiking neural networks (SNNs) offer energy efficiency on neuromorphic hardware.
• Direct training of SNNs is challenging due to non-differentiable spiking signals and complex dynamics.
• Key issues include gradient vanishing/explosion, signal decoding, and upstream task applicability.

Purpose of the Study:

• To develop effective methods for direct training of high-performance SNNs.
• To address gradient issues and improve spiking signal decoding.
• To enable SNNs for complex tasks like object recognition and detection.

Main Methods:

• Introduced Spiking Gate (SG) ResNet with a binary selection gate to implement residual learning and mitigate gradient vanishing/explosion.
• Proposed an Attention Spike Decoder (ASD) for superior spiking signal decoding compared to rate coding.
• Integrated SG ResNet and ASD into Spiking RetinaNet for hybrid SNN-ANN object detection.

Main Results:

• SG ResNet demonstrated superior accuracy on CIFAR-10 (94.52% top-1) and CIFAR-100 (75.64% top-1) with a minimal simulation time step.
• Spiking RetinaNet, using SG ResNet34, achieved an mAP of 0.296 on the MSCOCO object detection dataset.
• The proposed methods effectively overcome gradient vanishing/explosion and enhance signal decoding.

Conclusions:

• SG ResNet and ASD are effective for direct training of SNNs, enabling high performance in object recognition.
• The first direct-training hybrid SNN-ANN detector (Spiking RetinaNet) was developed for RGB images.
• These advancements pave the way for more efficient and powerful SNN applications.