Multimodal representation learning enables im-age–text alignment through shared embedding spaces optimized using contrastive objectives. Although Spiking Neural Networks (SNNs) provide biologically inspired temporal computation, their representational behavior under static multimodal supervision remains insufficiently explored. This study presents a controlled comparison between a Spiking Neural Network (SNN) and a Multilayer Perceptron (MLP) for multimodal image–text re-trieval. To isolate the effect of temporal spiking dynamics, both architectures were trained under identical embedding dimension-ality, optimization settings, contrastive learning objectives, and retrieval protocols. Experiments were conducted on a balanced episodic benchmark consisting of 1,000 image–text pairs. Results show that both architectures learn stable and non-collapsed embedding spaces with broad cosine similarity distributions. However, retrieval performance for both models remains close to chance under static supervision. Despite comparable task-level performance, cross-model cosine similarity analysis reveals substantial representational divergence between SNN and MLP embeddings, indicating distinct embedding geometries under identical learning conditions. The findings suggest that tempo-ral spiking dynamics alone do not improve static multimodal retrieval alignment and highlight a mismatch between spiking inductive bias and temporally unstructured supervision. Overall, the study emphasizes the importance of representation-level anal-ysis alongside conventional retrieval evaluation for neuromorphic multimodal learning systems.

Index Terms—Spiking Neural Networks, Multimodal Retrieval, Contrastive Learning, Neuromorphic Computing, Embedding Geometry, Representation Learning