New method cuts labeled data needs for spiking neural networks

Researchers have introduced a new method that lets spiking neural networks adapt to new data domains while requiring far fewer labeled examples, according to a study published in Applied Intelligence.

They have shown success in both theory and applications, but the study’s authors note a persistent obstacle: training deep spiking models still demands a large amount of labeled data. Existing state-of-the-art methods have tried to ease this burden by applying either domain adaptation or active learning on their own, rather than integrating the two.

SADA addresses that gap with two main components. The first is a long-term sample valuation strategy designed to identify which target-domain samples are the most informative, so that a limited annotation budget is spent on the examples most likely to improve performance. The second is a spike-domain distance loss, a training objective that jointly considers intra-class and inter-class distributions to support more effective knowledge transfer from the source domain to the target domain.

The team reports extensive experiments on multiple domain adaptation benchmarks. According to the abstract, SADA substantially reduces annotation requirements while maintaining competitive or superior performance compared with existing methods, achieving state-of-the-art results on most evaluation tasks. The authors include Qing Cai, Cheng Zeng, Xiurui Xie, Guisong Liu and Qiugang Zhan, and the article appears in Volume 56 of the journal as article number 523.

The published abstract describes the method’s scope as domain adaptation benchmarks evaluated under limited annotation budgets; detailed performance figures and the full experimental analysis are available in the complete article, which is accessible through the publisher.

Subject of Research: Sample-efficient active domain adaptation for spiking neural networks

Article Title: A sample-efficient active domain adaptation method for spiking neural networks

Article References: Cai, Q., Zeng, C., Xie, X., Liu, G., & Zhan, Q. (2026). A sample-efficient active domain adaptation method for spiking neural networks. Applied Intelligence, 56(16), Article 523. https://doi.org/10.1007/s10489-026-07547-w

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07547-w

Keywords: spiking neural networks, neuromorphic computing, domain adaptation, active learning, sample efficiency, machine learning, annotation budget, spike-domain distance loss, Applied Intelligence, sample-efficient, active, domain