High-Fidelity ANN-to-SNN Conversion via Closed-Loop CKA Distillation

ANN-to-SNN conversion offers energy-efficient inference but faces a fidelity-latency trade-off due to open-loop error accumulation. While conversion-aware training mitigates this, it sacrifices the generality of using off-the-shelf ANNs. We propose a closed-loop fine-tuning framework that calibrates these errors without altering the source model. Our approach employs a Dual Alignment Mechanism, utilizing global Kullback-Leibler divergence for output distillation and introducing an adaptive local Centered Kernel Alignment constraint, weighted by initial conversion loss, for feature alignment. We uncover a critical time-dependent dynamic: local constraints are essential for stabilizing representations in low-latency regimes (e.g., $T=8$) where global gradients are unstable, whereas global alignment drives fidelity at higher time steps. Experiments on CIFAR-10 demonstrate that our method achieves over 99\% of source ANN accuracy at $T=32$ (e.g., ResNet-18: 96.38\% vs.\ 96.39\%). Furthermore, this fine-tuning acts as a regularizer, yielding SNNs with input noise robustness that matches or exceeds the source ANN.

Modern artificial neural networks can be accurate but computationally expensive, especially on resource-limited devices. Spiking neural networks are an alternative that communicates using short binary events, or “spikes,” which can be more efficient on suitable hardware. A common way to build a spiking network is to convert a well-trained artificial neural network, but this conversion often loses accuracy unless the spiking network runs for many time steps, which increases latency. This paper studies how to make such conversion more accurate at low latency without changing or retraining the original artificial neural network. We introduce a closed-loop fine-tuning method in which the original network acts as a teacher and the converted spiking network acts as a student. The method aligns both final predictions and internal feature representations, using stronger guidance where conversion causes larger errors. Experiments show that this approach greatly reduces the accuracy gap between converted spiking networks and their source models, achieving nearly the same accuracy on CIFAR-10 at moderate time steps. It also improves robustness to noisy inputs. These results suggest a practical path toward more reliable and efficient spiking neural network inference for future neuromorphic or resource-constrained systems.