基于对偶误差的脉冲神经网络目标检测方法

脉冲神经网络(SNN)是一种模拟大脑神经元动力学的低功耗神经网络，为高计算效率、低能源消耗环境部署目标检测任务提供了可行的解决方案。由于脉冲的不可微性质导致SNN训练困难，一种有效的解决方法是将预训练的人工神经网络(ANN)转换为SNN来提高推理能力。然而，转换后的SNN 经常遇到性能下降和高延迟的问题，无法满足目标检测任务对高精度定位的要求。该文针对ANN转SNN过程中产生的误差问题，引入对偶误差模型降低转换性能损失。首先，该文分析误差产生原因，构建对偶误差模型来模拟ANN到SNN转换误差。进一步地，将对偶误差模型引入到ANN训练过程，使转换前后的模型在训练和推理过程中误差保持一致，从而降低模型的转换性能损失。最后，利用轻量化检测算法YOLO在数据集PASCAL VOC和 MS COCO上验证了对偶误差模型的有效性。

Spiking Neural Network for Object Detection Based on Dual Error

A Spiking Neural Network (SNN) is a low-power neural network that simulates the dynamics of neurons in the brain, providing a feasible solution for deploying object detection tasks in high computational efficiency and low energy consumption environments. Due to the non-differentiable nature of spikes, SNN training is difficult, and a practical solution is to convert pre-trained Artificial Neural Networks (ANNs) into SNNs to improve inference ability. However, the converted SNN often suffers from performance degradation and high latency, which can not meet the high-precision localization requirements for object detection tasks. A dual error is introduced to reduce the loss of conversion performance. To simulate the ANN to SNN conversion error, the causes of errors are analyzed, and a dual error model is built. Further, the dual error model is introduced into the ANN training process so that the errors of the models before and after conversion remain consistent during training and testing, thereby reducing the loss of conversion performance. Finally, the lightweight detection algorithm YOLO is used to verify the effectiveness of the dual error model on the PASCAL VOC and MS COCO datasets.