15 December 2023 in IEEE Transactions on Neural Networks and Learning Systems
Spiking neural networks (SNNs) operating with asynchronous discrete events show higher energy efficiency with sparse computation. A popular approach for implementing deep SNNs is artificial neural network (ANN)–SNN conversion combining both efficient training of ANNs and efficient inference of SNNs. However, the accuracy loss is usually nonnegligible, especially under few time steps, which restricts the applications of SNN on latency-sensitive edge devices greatly. In this article, we first identify that such performance degradation stems from the misrepresentation of the negative or overflow residual membrane potential in SNNs. Inspired by this, we decompose the conversion error into three parts: quantization error, clipping error, and residual membrane potential representation error. With such insights, we propose a two-stage conversion algorithm to minimize those errors, respectively. In addition, we show that each stage achieves significant performance gains in a complementary manner. By evaluating on challenging datasets including CIFAR-10, CIFAR-100, and ImageNet, the proposed method demonstrates the state-of-the-art performance in terms of accuracy, latency, and energy preservation. Furthermore, our method is evaluated using a more challenging object detection task, revealing notable gains in regression performance under ultralow latency, when compared with existing spike-based detection algorithms. Codes will be available at: https://github.com/Windere/snn-cvt-dual-phase.
13 February 2024 in IEEE Transactions on Emerging Topics in Computational Intelligence
Spiking neural networks (SNNs) are increasingly applied to deep architectures. Recent works are developed to apply spatio-temporal backpropagation to directly train deep SNNs. But the binary and non-differentiable properties of spike activities force directly trained SNNs to suffer from serious gradient vanishing. In this paper, we first analyze the cause of the gradient vanishing problem and identify that the gradients mostly backpropagate along the synaptic currents. Based on that, we modify the synaptic current equation of leaky-integrate-fire neuron model and propose the improved LIF (IM-LIF) neuron model on the basis of the temporal-wise attention mechanism. We utilize the temporal-wise attention mechanism to selectively establish the connection between the current and historical response values, which can empirically enable the neuronal states to update resilient to the gradient vanishing problem. Furthermore, to capture the neuronal dynamics embedded in the output incorporating the IM-LIF model, we present a new temporal loss function to constrain the output of the network close to the target distribution. The proposed new temporal loss function could not only act as a regularizer to eliminate output outliers, but also assign the network loss credit to the voltage at a specific time point. Then we modify the ResNet and VGG architecture based on the IM-LIF model to build deep SNNs. We evaluate our work on image datasets and neuromorphic datasets. Experimental results and analysis show that our method can help build deep SNNs with competitive performance in both accuracy and latency, including 95.66% on CIFAR-10, 77.42% on CIFAR-100, 55.37% on Tiny-ImageNet, 97.33% on DVS-Gesture, and 80.50% on CIFAR-DVS with very few timesteps.