Event-Clock Hybrid Driven Reconfigurable Perception-Computation Technology

今年，我们为混合驱动的可重构感知计算平台开发并验证了以下技术：(1)脑电图信号的尖峰编码及其基于尖峰神经网络（SNN）的处理。在一些工作中，我们成功地将尖峰编码分别应用于脑电信号的自适应编码、随机编码和频率编码，并实现了基于SNN的竞争性睡眠阶段分类精度；(2)深度snn的三元权量化方法及硬件实现。在这项工作中，我们通过将snn的权重量化为{- 1,0,1}来实现节能的推理硬件。通过设计跨层连接，避免了模型训练过程中的梯度消失问题。在推理阶段，可以在三元权重snn中使用简单的逻辑运算，以减少硬件开销；(3)可重构对分神经网络（BNN）拓扑结构的训练与构建机制。提出了一种通用的神经网络构建方法及其训练机制。通过构造具有对分结构的掩模矩阵，可以自动训练具有特定拓扑结构的BNN模型。

This year, we developed and verified the following technologies for the hybrid-driven reconfigurable perception-computation platform: (1) Spike coding of Electroencephalogram (EEG) signals and its spiking neural network (SNN)-based processing. In several works, we successfully applied spike coding to adaptive, stochastic and frequency coding of EEG signals, respectively, and achieved competitive sleep stage classification accuracy based on SNN; (2) A ternary weight quantization method for deep SNNs and hardware implementation. In this work, we achieved energy-efficient inference hardware by quantizing the weights of SNNs to {-1, 0, 1}. The gradient disappearance problem during model training is avoided by designing cross-layer connections. Simple logical operations can be used in ternary weights SNNs at the inference stage, to reducing hardware overhead; (3) Training and construction mechanism of reconfigurable bisection neural network (BNN) topology. We proposed a general construction method of BNN and its training mechanism. By constructing a mask matrix with a bisection structure, we can automatically train a BNN model with a specific topology.