Converting High-Performance and Low-Latency SNNs through Explicit Modelling of Residual Error in ANNs

TL;DR

The paper tackles the problem of accuracy loss in ultra-low-latency ANN-SNN conversion caused by residual membrane potential misrepresentation. It proposes explicit residual error modeling by injecting layer-wise Gaussian noise into the source ANN activation, forming a Noisy Quantized activation that preserves zero-mean conversion error on expectation (Theorem 1). The method combines a quantized activation with per-layer noise calibrated via a small validation set, achieving state-of-the-art or near-ANN performance at time steps as low as $T\in\{2,4,8\}$ on CIFAR-10/100 without surrogate training. This yields practical ultra-low-latency SNNs suitable for neuromorphic hardware with minimal training overhead and strong potential for edge devices.

Abstract

Spiking neural networks (SNNs) have garnered interest due to their energy efficiency and superior effectiveness on neuromorphic chips compared with traditional artificial neural networks (ANNs). One of the mainstream approaches to implementing deep SNNs is the ANN-SNN conversion, which integrates the efficient training strategy of ANNs with the energy-saving potential and fast inference capability of SNNs. However, under extreme low-latency conditions, the existing conversion theory suggests that the problem of misrepresentation of residual membrane potentials in SNNs, i.e., the inability of IF neurons with a reset-by-subtraction mechanism to respond to residual membrane potentials beyond the range from resting potential to threshold, leads to a performance gap in the converted SNNs compared to the original ANNs. This severely limits the possibility of practical application of SNNs on delay-sensitive edge devices. Existing conversion methods addressing this problem usually involve modifying the state of the conversion spiking neurons. However, these methods do not consider their adaptability and compatibility with neuromorphic chips. We propose a new approach based on explicit modeling of residual errors as additive noise. The noise is incorporated into the activation function of the source ANN, which effectively reduces the residual error. Our experiments on the CIFAR10/100 dataset verify that our approach exceeds the prevailing ANN-SNN conversion methods and directly trained SNNs concerning accuracy and the required time steps. Overall, our method provides new ideas for improving SNN performance under ultra-low-latency conditions and is expected to promote practical neuromorphic hardware applications for further development.

Figures (7)

• Figure 1: Diagram of our proposed conversion method. State-of-the-art low-latency conversion methods requiring training ANN with quantized activation still bring about a conversion gap in activation. We propose to incorporate additive noise into ANN activation, aiming to compensate for the activation gap.
• Figure 2: The distribution of the residual error of ANN output and the average postsynaptic potential for SNN. We calculate the residual error of some activation layers during the training of VGG16 on the CIFAR10 dataset.
• Figure 3: Comparison of SNN output $\phi^l(T)$ and ANN output $\mathbf{a}^l$ with the same input $\mathbf{z}^l$. The figure shows two activation functions for source ANNs: quantization clip-floor-shift (QCFS) activation (left) and our proposed Noisy Quantized activation with residual error noise modeling (right).
• Figure 4: The effect of noise intensity. Adding a certain amount of noise to the activation function benefits inference performance at short time steps.
• Figure 5: Accuracy curves of ANN and SNN on the testing dataset during the training process

Theorems & Definitions (1)

• Theorem 1