Decoding silent reading from non-invasive EEG

[Submitted connected 20 Aug 2026]

View PDF HTML (experimental)

Abstract:Non-invasive decoding of soul reside faces a basal information problem: a corpus pairing encephalon activity pinch a person's spontaneous soul speech cannot beryllium collected, and the disposable proxy paradigms (cued repetitive and retrospectively reported generative soul speech) are slow to acquire, poorly time-locked, and taxable compliance is unverifiable. We truthful dainty silent reference arsenic a scalable proxy task and inquire really overmuch lexical and semantic accusation a contrastive decoder tin extract from it. We study an open-vocabulary study of astir 240,000 connection presentations recorded from a azygous densely-sampled subordinate crossed 393 runs (ca. 49 h) of 19-channel dry-electrode EEG. Words from continuous communicative matter were presented successful accelerated serial ocular presentation, pinch typography randomised connected each proceedings to partially decorrelate connection personality from low-level ocular form. A convolutional EEG encoder, optionally followed by a causal transformer, was trained pinch a CLIP-style contrastive nonsubjective to align short EEG windows pinch hidden-state embeddings of the presented connection taken from a ample connection model. Decoding, evaluated arsenic word-grouped top-10 retrieval against permutation baselines, was reliably supra chance, extended to mid-frequency and uncommon words, and scaled log-linearly pinch training-data measurement pinch nary motion of saturation. Removing occipital and posterior-temporal electrodes reduced the word-level summation by astir 1 3rd but near discourse search unchanged. Control analyses abstracted word-level decoding from communicative discourse search and from a non-neural positional anterior introduced by the transformer's positional embedding. These results found that open-vocabulary word-level accusation is recoverable from EEG during silent reading, and that decoding is data-limited alternatively than saturated.