Neural decoding of autobiographical mental image features with a general semantic model

• Autobiographical mental imagery can be decoded from brain signals using a general semantic model.¹
• The work links computational neuroscience with potential for communication BCIs.¹
• Decoder generalizability across semantic content is a focus of the approach.^{1 1}

Weekly enrichment (2026-07-20)

• The study (Anderson, Fernandino, Gross, Rey & Binder, Nature Communications 2025) tested whether language comprehension and self-generated autobiographical mental imagery share neural representations, and found the largest overlap in midline cortical regions, present across both midline and lateral surface cortex.^{2 3}
• Fifty participants underwent fMRI while imagining their own experiences of twenty common scenarios, and their self-reported experiential feature ratings of those mental images were reconstructed from brain activity.^{2 3}
• The general semantic decoding model was trained on a separate sentence-fMRI dataset of 14 participants (none overlapping with the imagery group) who read 240 sentences describing simple situations.^{2 3}
• All fMRI data were mapped into a common Schaefer-1000 cortical parcellation, and 20 experiential features (e.g., social, auditory, motor sensations) were used to model both sentence meaning and mental image content.^{2 3}
• The fMRI-to-feature mapping was fit with ridge regression on concatenated data, then transferred to reconstruct imagery feature ratings via matrix multiplication of the imagery fMRI with the learned decoder.^{2 3}
• The result demonstrates zero-shot decoding: participant-specific autobiographical feature ratings were recovered across people and across cognitive tasks despite the decoder never being trained on imagery data.^{2 3}
• Reconstruction accuracy was quantified with Spearman correlations between reconstructed and observed feature ratings across the 20 scenarios per participant, evaluated with signed-rank tests against zero and FDR correction across the 20 features.^{2 3}
• Implication for BCIs: a semantic decoder built from one group reading sentences can generalize to another group’s internally generated imagery, suggesting a shared, transferable representational code that could support communication interfaces for imagined content.^{2 3}

Footnotes

1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE8tVV9sOTJOc0NjTWhNaTNyVmZRM0Zfb1loZy0yVjgtVXNQaF9TM0x0cHBzX2k0SnptSTFxQ1dYd0N1a3ZZeEROMWthYlNMWGdVREcwOFVKTGs1bnduRHVZ?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.nature.com/articles/s41467-025-65541-1 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. https://doi.org/10.1038/s41467-025-65541-1 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸