Decoding silent reading from non-invasive EEG (arxiv.org)

Researchers have unveiled a groundbreaking study on decoding silent reading using non-invasive EEG technology, marking a significant advance in understanding inner speech processing. The study addresses the challenges of collecting reliable data on spontaneous inner monologue, proposing silent reading as a scalable alternative. By analyzing EEG data from 240,000 word presentations in a controlled environment, the team applied a contrastive decoder—integrating a convolutional encoder and a causal transformer—trained to align EEG signals with word embeddings from a large language model. The findings reveal that robust lexical and semantic information can be extracted from EEG during silent reading, outperforming baseline models. This capability indicates that decoding is mainly data-limited rather than saturated, paving the way for improved methodologies in brain-computer interfaces (BCIs) and enhancing our understanding of cognitive processes related to reading and language. Furthermore, the experiments demonstrated the importance of electrode placement, showing that removing certain electrodes impacted word-level gains while preserving context tracking. These insights underscore the potential of advanced machine learning techniques to decode complex brain activity and may have implications for assistive technologies in communication and rehabilitation.