Proceedings of the Annual Meeting of the Cognitive Science Society

Synthetic Aperture Radar (SAR) recognition is frequently impeded by severe speckle noise and data scarcity. In contrast, human vision excels in these challenging conditions due to robust top-down cognitive modulation. This work investigates whether human cognitive representations can be transferred to SAR recognition models. We introduce the first paired fMRI–SAR dataset designed to capture the cognitive neural responses of human viewing SAR images. Building upon this dataset, we propose a cognitive fusion–distillation framework. This framework integrates fMRI-derived cognitive representations into a teacher network and distills the cognitive capability into a lightweight student model via prompt tuning. Extensive experiments demonstrate consistent performance improvements in few-shot, cross-domain, and out-of-distribution conditions. Furthermore, Representational similarity analysis reveals that cognition-guided models learn representation patterns distinct from purely data-driven models. These results suggest that neuro-cognitive signals can serve as transferable inductive biases to enhance robustness and generalization in specialized vision tasks.