Adaptive Calibration: A Unified Conversion Framework of Spiking Neural Network

TL;DR

This work addresses the performance-energy gap in ANN-to-SNN conversion by introducing a training-free Adaptive Calibration framework. It combines AdaFire, a layer-aware adaptive-firing neuron model, with two efficiency mechanisms, SSC and IAT, to minimize the Unevenness Error while reducing spike counts and latency. The approach achieves state-of-the-art accuracy with substantial energy savings across CIFAR-10, CIFAR-100, and ImageNet (up to $70.1\%$, $60.3\%$, and $43.1\%$ respectively) and demonstrates effectiveness on 2D/3D, neuromorphic, detection, and segmentation tasks, all without retraining. This paves the way for practical, energy-efficient neuromorphic deployments and rapid adaptation to diverse workloads.

Abstract

Spiking Neural Networks (SNNs) are seen as an energy-efficient alternative to traditional Artificial Neural Networks (ANNs), but the performance gap remains a challenge. While this gap is narrowing through ANN-to-SNN conversion, substantial computational resources are still needed, and the energy efficiency of converted SNNs cannot be ensured. To address this, we present a unified training-free conversion framework that significantly enhances both the performance and efficiency of converted SNNs. Inspired by the biological nervous system, we propose a novel Adaptive-Firing Neuron Model (AdaFire), which dynamically adjusts firing patterns across different layers to substantially reduce the Unevenness Error - the primary source of error of converted SNNs within limited inference timesteps. We further introduce two efficiency-enhancing techniques: the Sensitivity Spike Compression (SSC) technique for reducing spike operations, and the Input-aware Adaptive Timesteps (IAT) technique for decreasing latency. These methods collectively enable our approach to achieve state-of-the-art performance while delivering significant energy savings of up to 70.1%, 60.3%, and 43.1% on CIFAR-10, CIFAR-100, and ImageNet datasets, respectively. Extensive experiments across 2D, 3D, event-driven classification tasks, object detection, and segmentation tasks, demonstrate the effectiveness of our method in various domains. The code is available at: https://github.com/bic-L/burst-ann2snn.

Figures (8)

• Figure 1: Comparison of the proposed method versus existing methods on ImageNet (VGG-16). 'T' refers to averaged inference time steps, and 'Time Cost' represents the total GPU hours before final inference. Our Adaptive Calibration method achieves competitive accuracy with fewer timesteps and lower energy consumption. Notably, it only requires a short setup time and eliminates the need for re-training.
• Figure 1: Performance comparison on different tasks.
• Figure 2: Optimization Process for Adaptive Calibration. The process begins within a search space containing candidates for Max Burst-firing Patterns ($\varphi$) and Threshold Ratio ($\rho$). The estimators assess these candidates to evaluate their performance and energy efficiency. Optimum configurations of each layer are then selected using the Pareto-frontier method.
• Figure 3: The unevenness error dominates conversion loss in ANN-to-SNN conversion. (a) Percentage of three main conversion errors, with the unevenness error dominating. (b) The adoption of Adaptive-firing Neurons greatly reduce the unevenness error. (c) Burst-firing mechanism in the Adaptive-firing Neuron model. The Adaptive-firing Neuron model minimizes this loss by allowing multiple spikes to be generated in rapid succession when the membrane potential exceeds the threshold.
• Figure 4: (a-b) Variation of the sensitivity of each layer with respect to: (a) different max burst-firing patterns $\varphi$ and (b) different threshold ratios $\rho$. (c) Pareto Frontier Searching. Optimizing the network configuration to reduce sensitivity improves performance, where each data point represents a distinct layer-specific configuration.