Novel channel-oriented design boosts EEG-to-music reconstruction accuracy

Most brain-computer interface (BCI) research has focused on decoding vision and language from neural signals, but music — with its weak, distributed, and noisy neural signatures — remains a far more challenging frontier. In a new paper on arXiv, Jiaxin Qing, Junwei Lu, and Lexin Li identify a critical flaw in existing EEG decoding pipelines: early channel mixing, which aggregates signals from all electrodes at the start, destroys the weak but discriminative EEG patterns needed for music reconstruction. To solve this, they propose a channel-oriented design with three novel components.

First, channel-wise tokenization treats each electrode as its own explicit token, preserving spatially localized neural evidence. Second, channel-wise multi-view self-distillation enforces consistency across temporal crops and random channel subsets, learning robust and distributed representations. Third, channel-wise data augmentation introduces structured channel dropout to improve invariance to noise, artifacts, and missing electrodes. Together, these components enable stable alignment to a semantic music representation space within an encoding-alignment-decoding pipeline. The authors provide theoretical justification for why preserving channel-level structure improves alignment, and empirically demonstrate consistent and significant performance gains over multiple state-of-the-art baselines.