Spiking Neural Networks with Consistent Mapping Relations Allow High-Accuracy Inference

TL;DR

This work addresses the challenge of efficiently converting trained ANNs to SNNs without sacrificing accuracy or incurring large inference delays. It identifies the core source of conversion errors as the mismatch between ANN activation mappings and the spike-based input-output dynamics, and introduces the Consistent ANN-SNN Conversion (CASC) framework, combining Consistent IF (CIF) neurons and Wake-Sleep Conversion (WSC) to achieve near-lossless conversion. Across classification and object detection tasks, CASC delivers high accuracy with substantially fewer time steps and improved energy efficiency, while demonstrating robustness to quantization and generalization across architectures. The results suggest CASC as a practical pathway to deploying accurate, low-power SNNs on neuromorphic hardware in real-world vision tasks.

Abstract

Spike-based neuromorphic hardware has demonstrated substantial potential in low energy consumption and efficient inference. However, the direct training of deep spiking neural networks is challenging, and conversion-based methods still require substantial time delay owing to unresolved conversion errors. We determine that the primary source of the conversion errors stems from the inconsistency between the mapping relationship of traditional activation functions and the input-output dynamics of spike neurons. To counter this, we introduce the Consistent ANN-SNN Conversion (CASC) framework. It includes the Consistent IF (CIF) neuron model, specifically contrived to minimize the influence of the stable point's upper bound, and the wake-sleep conversion (WSC) method, synergistically ensuring the uniformity of neuron behavior. This method theoretically achieves a loss-free conversion, markedly diminishing time delays and improving inference performance in extensive classification and object detection tasks. Our approach offers a viable pathway toward more efficient and effective neuromorphic systems.

Figures (6)

• Figure 1: Statistics of the mapping of total input and output of spiking neurons and the output of SNN and ANN on CIFAR10 with VGG16. (a) Traditional methods using IF neurons (orange points) fail to achieve a one-to-one mapping relationship in total input-output as ANNs (pink line), (b) whereas our method (purple points) can achieve fitting with CQReLU (blue line). (c) Visualizing the outputs of the three models, we find that our method accurately identifies the maximum values and closely approximates the ANN outputs without loss.
• Figure 2: Impact of stable points on achieving consistent input-output mapping with ANN. We show how $Y$ varies with four different cases of inputs $X$, where the red squares are stable points. (a) The stable point does not cross the upper and lower bounds for real-valued inputs. (b) As long as the stable point does not cross the upper and lower bounds, the SNN can obtain the expected output. (c) If the stable point exceeds the upper bound, the SNN cannot recover the spikes already issued, thus causing more spikes. (d) If the stable point is below the lower bound, the SNN does not have the extra time step to emit the spikes that should have been transmitted.
• Figure 3: Illustration of our proposed method. In contrast to IF neurons, the CIF model also determines whether the current membrane potential is less than zero to ensure that neurons with a total historical output that is not zero satisfy $0\leq V_i^{(\ell)}[t]\leq V_{th}$, thus relieving the effect of the upper bound on accurate conversion. With WSC, the neurons only receive external information before $\mathcal{Q}$ time steps, ensuring the network has the same inputs as the ANN. Furthermore, WSC gives the SNN more time to emit unissued spikes and correct the more spikes. Using the proposed method, the SNN can accurately approximate the ANN activation value layer by layer.
• Figure 4: Effect of the proposed methods on CIFAR100 with VGG16. (a) Ratio of neurons with more spikes in the first layer. (b) Ratio of neurons with more spikes in the second layer. (c) Ratio of neurons with fewer spikes in the second layer. (d) MSE of SNN output and ANN output. (e) Top-1 accuracy of our methods.
• Figure 5: Measuring the relative error on the CIFAR100 and ImageNet with VGG16 and ResNet34. Our method demonstrates the ability to achieve nearly loss-free conversion. The horizontal line represents the mean, the triangle represents the median, and the box range is 5% to 95%.