BN-SNN: Spiking neural networks with bistable neurons for object detection

Area of Science:

• Artificial Intelligence
• Computer Vision
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer an energy-efficient alternative to traditional neural networks.
• Current CNN-to-SNN conversion methods face challenges with long temporal durations and inference latency, impacting accuracy.
• The use of SNNs in object detection tasks is an emerging and under-explored research area.

Purpose of the Study:

• To develop a novel CNN-to-SNN conversion technique for object detection.
• To address the limitations of existing conversion methods, specifically temporal duration and inference latency.
• To explore the application of SNNs in object detection using a new neuron model.

Main Methods:

• Proposed a novel approach integrating a bistable integrate-and-fire (BIF) neuron model with a single-shot multibox detector (SSD) as the detection head.
• Converted the ResNet architecture into an SNN using the proposed BIF neuron framework.
• Validated the approach on object detection tasks using the MS-COCO and Automotive GEN1 datasets.

Main Results:

• The proposed conversion technique achieved reduced temporal steps for SNN-based object detection.
• Significant enhancements in mean average precision (mAP) were observed, with mAP@0.5 scores of 0.476 on MS-COCO and 0.591 on Automotive GEN1.
• Demonstrated the effectiveness of the BIF neuron model in SNNs for object detection tasks.

Conclusions:

• This research presents the first application of bistable integrate-and-fire neurons to object detection.
• The novel conversion approach successfully reduces temporal steps and improves mAP in SNNs for object detection.
• The findings represent a significant advancement in the field of SNNs and their practical applications in computer vision.