Obtaining Optimal Spiking Neural Network in Sequence Learning via CRNN-SNN Conversion

• DOI: 10.48550/arxiv.2408.09403
• Publication Date: 2024-08-18
• Authors: Jiahao Su; Kang Soo You; Zekai Xu; Weizhi Xu; Zhezhi He
• Keywords: Sequence (biology)

Spiking neural networks (SNNs) are becoming a promising alternative to conventional artificial (ANNs) due their rich dynamics and the implementation of energy-efficient neuromorphic chips. However, non-differential binary communication mechanism makes SNN hard converge an ANN-level accuracy. When encounters sequence learning, situation becomes worse difficulties in modeling long-range dependencies. To overcome these difficulties, researchers developed variants LIF neurons different surrogate gradients but still failed obtain good results when became longer (e.g., $>$500). Unlike them, we optimal learning by directly mapping parameters from quantized CRNN. We design two sub-pipelines support end-to-end conversion structures networks, which is called CNN-Morph (CNN $\rightarrow$ QCNN BIFSNN) RNN-Morph (RNN QRNN RBIFSNN). Using pipelines s-analog encoding method, error our framework zero. Furthermore, give theoretical experimental demonstration lossless CRNN-SNN conversion. Our show effectiveness method over short long timescales tasks compared with state-of-the-art learning- conversion-based methods. reach highest accuracy 99.16% (0.46 $\uparrow$) on S-MNIST, 94.95% (3.95 PS-MNIST (sequence length 784) respectively, lowest loss 0.057 (0.013 $\downarrow$) within 8 time-steps collision avoidance dataset.