FTBC: Forward Temporal Bias Correction for Optimizing ANN-SNN Conversion

TL;DR

The paper tackles the bottleneck in ANN-to-SNN conversion: preserving accuracy while exploiting temporal spiking dynamics. It proposes Forward Temporal Bias Correction (FTBC), a forward-pass, layer-, and channel-wise bias calibration that aligns SNN outputs with their ANN counterparts at every time step, supported by a theoretical result guaranteeing zero expected conversion error under proper temporal bias. A practical heuristic estimates these biases without backpropagation, enabling efficient post-conversion calibration. Extensive experiments on CIFAR-10/100 and ImageNet show that FTBC surpasses state-of-the-art conversion methods and synergizes with other techniques (e.g., QCFS, RTS) to achieve high accuracy at low latency, demonstrating strong practical impact for energy-efficient neuromorphic inference.

Abstract

Spiking Neural Networks (SNNs) offer a promising avenue for energy-efficient computing compared with Artificial Neural Networks (ANNs), closely mirroring biological neural processes. However, this potential comes with inherent challenges in directly training SNNs through spatio-temporal backpropagation -- stemming from the temporal dynamics of spiking neurons and their discrete signal processing -- which necessitates alternative ways of training, most notably through ANN-SNN conversion. In this work, we introduce a lightweight Forward Temporal Bias Correction (FTBC) technique, aimed at enhancing conversion accuracy without the computational overhead. We ground our method on provided theoretical findings that through proper temporal bias calibration the expected error of ANN-SNN conversion can be reduced to be zero after each time step. We further propose a heuristic algorithm for finding the temporal bias only in the forward pass, thus eliminating the computational burden of backpropagation and we evaluate our method on CIFAR-10/100 and ImageNet datasets, achieving a notable increase in accuracy on all datasets. Codes are released at a GitHub repository.

Figures (4)

• Figure 1: Overview of our proposed Forward Temporal Bias Correction (FTBC) method for ANN-SNN conversion. This approach calibrates time-based channel-wise bias terms ($b_t$) by dynamically adjusting membrance potential based on the temporal activation patterns observed in the pre-trained ANN. These adjustments are distributed across timesteps from $t = 1$ to $t = T$, ensuring the temporal precision of spike dynamics in SNNs is maintained.
• Figure 2: A heuristic algorithm to iteratively find $b^*$: At each iteration, one estimates the difference between the desired $\alpha$ and current $E_i$ expected values, and estimates the next $b_{i+1}$ with $b_i+c(\alpha-E_i)$, where $c$ is some positive constant.
• Figure 3: Change of accuracy at various time steps with batch iteration. The reported accuracies are for a VGG16 model trained on CIFAR100 dataset and $\alpha=0.5$ and $\alpha=1.0$ as a hyperparameter in calibration (see Algorithm \ref{['algo:bias-correction']}).
• Figure 4: Membrane potential distributions before firing for VGG16 model (pretrained on CIFAR100 dataset). In both cases, the SNN converted model has been calibrated for all the previous layers and for all the previous time steps (in (b)). A random IF neuron in the layer has been chosen for the presentation of the distributions.

Theorems & Definitions (6)

• proposition thmcounterproposition
• theorem thmcountertheorem
• proposition thmcounterproposition
• proof
• theorem thmcountertheorem
• proof