Self-cross Feature based Spiking Neural Networks for Efficient Few-shot Learning

TL;DR

This work addresses the challenge of energy-efficient few-shot learning by leveraging Spiking Neural Networks (SNNs) and introducing a Self-cross Feature Network (SSCF) that combines a Self Feature Extractor (SFE) with a Cross Feature Contrastive (CFC) module. The backbone is a VGGSNN, and learning is guided by two complementary losses: a Temporal Efficient Training (TET) loss that propagates supervision through time, and an InfoNCE-based contrastive loss that enforces cross-set discriminability, with the total loss defined as $\mathcal{L}_{Total} = \lambda \mathcal{L}_{TET} + (1-\lambda) \mathcal{L}_{infoNCE}$. Empirically, SSCF achieves state-of-the-art-like performance among SNN-FSL methods on N-Omniglot (up to $98.9\%$ accuracy on 5-way 5-shot) and competitive results on CUB and miniImageNet, while maintaining significantly lower energy consumption due to the event-driven nature of SNNs (e.g., $E_{SNN} \ll E_{ANN}$ given comparable FLOPs). These results, along with ablations showing the crucial roles of SFE and CFC and the robustness conferred by InfoNCE, demonstrate the practicality of energy-efficient, cross-feature-aware SNNs for real-world few-shot tasks.

Abstract

Deep neural networks (DNNs) excel in computer vision tasks, especially, few-shot learning (FSL), which is increasingly important for generalizing from limited examples. However, DNNs are computationally expensive with scalability issues in real world. Spiking Neural Networks (SNNs), with their event-driven nature and low energy consumption, are particularly efficient in processing sparse and dynamic data, though they still encounter difficulties in capturing complex spatiotemporal features and performing accurate cross-class comparisons. To further enhance the performance and efficiency of SNNs in few-shot learning, we propose a few-shot learning framework based on SNNs, which combines a self-feature extractor module and a cross-feature contrastive module to refine feature representation and reduce power consumption. We apply the combination of temporal efficient training loss and InfoNCE loss to optimize the temporal dynamics of spike trains and enhance the discriminative power. Experimental results show that the proposed FSL-SNN significantly improves the classification performance on the neuromorphic dataset N-Omniglot, and also achieves competitive performance to ANNs on static datasets such as CUB and miniImageNet with low power consumption.

Figures (3)

• Figure 1: Architecture of the proposed few-shot learning framework based on SNNs. It combines self-feature extractor module and a cross-feature contrastive module to further refine feature representation and greatly reduce power consumption.
• Figure 2: T-SNE visualization on N-Omniglot in different time steps.We can intuitively see the clustering effect in the feature space and the impact of the time step on the feature distribution. At a shorter time step (such as T=4), the feature distribution is more dispersed and the distinction between categories is low. As the time step increases (such as T=12), the feature distribution gradually becomes more compact and the distinction between categories is significantly improved.
• Figure 3: Visualization of spiking activities at different time steps.In the above picture, we can clearly see the changes in the activity of the spiking. The features at time T=2 are significantly more than those at time T=1. This shows that as the time step increases, more and more features are activated and spikings are emitted, thereby capturing more features.