EEG-based visual decoding tries to map neural signals to visual semantics, but faces a granularity mismatch: EEG is noisy and coarse, while visual features are often treated as static and fine-grained. In a recent arXiv preprint, Yin et al. propose CAIA, a framework that adaptively blurs visual input to mimic human selective attention and simultaneously cleans EEG representations using neural oscillation priors.
The visual side uses a cognitive-dynamics-based mechanism that combines center-biased and saliency-guided cues through cross-modal attention, dynamically reducing redundant information. On the EEG side, a spectral sparsity regularizer inspired by the information bottleneck principle suppresses noise. A distribution-aware boundary calibration loss further stabilizes alignment when outlier samples would otherwise skew the learned mapping.
Experiments show that CAIA improves both subject-dependent and subject-independent Top-1 and Top-5 accuracy in zero-shot brain-to-image retrieval, significantly outperforming prior methods. The authors argue that optimizing visual information density to match neural granularity offers a more interpretable and robust path for neural decoding, rather than treating visual features as fixed targets.