用于双阈值脉冲神经网络的改进自适应阈值算法

脉冲神经网络(spiking neural network, SNN)由于在神经形态芯片上低功耗和高速计算的独特性质而受到广泛的关注。深度神经网络(deep neural network, DNN)到SNN的转换方法是有效的脉冲神经网络训练方法之一，然而从DNN到SNN的转换过程中存在近似误差，转换后的SNN在短时间步长下遭受严重的性能退化。通过对转换过程中的误差进行详细分析，将其分解为量化和裁剪误差以及不均匀误差，提出了一种改进SNN阈值平衡的自适应阈值算法。通过使用最小化均方误差(MMSE)更好地平衡量化误差和裁剪误差；此外，基于IF神经元模型引入了双阈值记忆机制，有效解决了不均匀误差。实验结果表明，改进算法在CIFAR-10、CIFAR-100数据集以及MIT-BIH心律失常数据库上取得了很好的性能，对于CIFAR10数据集，仅用16个时间步长就实现了93.22%的高精度，验证了算法的有效性。

Improved adaptive threshold algorithm for double threshold spiking neural network

The spiking neural network(SNN) has gained widespread attention due to its low power consumption and high-speed computing capabilities on neuromorphic chips. The conversion from deep neural network(DNN) to SNN is an effective training method for SNN. However, there are approximation errors in the conversion process, leading to significant performance degradation of the converted SNN under short time steps. Through a detailed analysis of the errors in the conversion process, this paper decomposed them into quantization and pruning errors and asymmetric errors, and proposed an improved adaptive threshold algorithm to balance the threshold of SNN by minimizing the mean square error(MMSE) to achieve a better balance between quantization and pruning errors. Additionally, this algorithm introduced a dual-threshold memory mechanism based on the IF neuron model to effectively address the asymmetric errors. Experimental results demonstrate that the improved algorithm achieves excellent performance on the CIFAR-10, CIFAR-100 datasets, and the MIT-BIH arrhythmia dataset. For the CIFAR-10 dataset, it achieves a high accuracy of 93.22% with only 16 time steps, validating the effectiveness of the algorithm.