Proceedings of the Annual Meeting of the Cognitive Science Society

Rapid Serial Visual Presentation (RSVP) is a critical cognitive paradigm for investigating visual attention dynamics and developing high-speed Brain-Computer Interfaces (BCIs). However, real-world deployment faces challenges from environmental noise, particularly cross-modal auditory perturbations that degrade neural decoding. This work proposes the Time-Frequency Dual-View Network (TFDV-Net), which fuses time-domain and frequency-domain features via a global-local interaction backbone to capture weak ERP signals. Furthermore, a Dual-Stream Knowledge Distillation strategy is introduced to enhance robustness, enforcing the model to reconstruct task-invariant cognitive patterns from noise-corrupted signals by learning from a "clean-state" teacher. The proposed method achieves a state-of-the-art balanced accuracy of 83.02% on a dataset containing ecological noise. Notably, we find that semantic conversational noise induces a significantly larger performance drop compared to non-semantic urban noise. This result quantitatively validates the hypothesis that high-level semantic processing imposes a greater cognitive load on the visual attention system than low-level acoustic interference.