SpikeHSA: Heterogeneous Space-Aligned Spiking Hashing for cross-modal retrieval

Spiking Neural Networks (SNNs) offer a promising paradigm for cross-modal hashing (CMH) due to their biological plausibility and energy efficiency. However, sparse spike-driven representations can limit semantic encoding, while conventional single-space alignment is insufficient to capture complementary fine-grained cross-modal semantics. To address these issues, we propose SpikeHSA, a Heterogeneous Space-Aligned Spiking Hashing framework for Cross-Modal Retrieval. Inspired by neuronal threshold diversity, SpikeHSA introduces the Heterogeneous Leaky Integrate-and-Fire (H-LIF) neuron, which assigns dimension-specific thresholds to diversify spiking responses and improve the information density of spike-based representations. We further develop a Heterogeneous Space Alignment (HSA) strategy that projects image, text, and fused multimodal representations into multiple heterogeneous spaces and aligns each unimodal representation with the fused semantic reference. Extensive experiments on three widely used benchmarks demonstrate that, with convolutional image encoders, SpikeHSA improves mAP by up to 4.6% over ANN-based baselines and 4.3% over the SNN-based method, while requiring as little as approximately one-eighth of the estimated inference energy of non-spiking baselines Code will be available at https://github.com/whisper233/SpikeHSA .