/ EN
2026-07-07 00:00 Germany Papers Foundations & Methods Translated from EN

Deep Learning Decodes Imagined Sounds and Images From MEG, Topping 70% for Visual Imagery

Summary Researchers recorded magnetoencephalography (MEG) from 18 right-handed participants as they imagined sounds and pictures, then compared two decoders: a convolutional neural network (CNN) and a linear logistic regression model. The CNN decoded both tasks above chance and exceeded 70% accuracy for visual imagery. It still decoded significantly when trained only on cortical regions unrelated to the task, suggesting imagined content is spread across partially overlapping networks rather than confined to a single sensory area, an experimental basis for feeding auditory and visual information into BCI decoders together.
Why it matters By comparing imagined sounds and imagined images within one MEG dataset, the study shows their cortical representations overlap, with decodable information even in task-irrelevant regions; for BCI, that means future decoders need not fixate on a single sensory area and can let modalities complement each other, though with only 18 participants a usable system remains some way off.

BCIwiki (bciwiki.com) — A convolutional neural network decoded visual mental imagery from magnetoencephalography (MEG) with a mean accuracy above 70%, outperforming a linear logistic regression model, in a study of 18 right-handed participants reported in the Journal of Neurophysiology on July 7, 2026. The authors are from the Department Artificial Intelligence in Biomedical Engineering (AIBE) at Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) in Germany.

Participants underwent MEG recording during auditory and visual imagery tasks, and the signals were source-reconstructed within modality-specific cortical regions of interest. The team compared the CNN and the linear logistic regression model in a subject-specific classification framework. Both approaches decoded above chance, and the CNN beat the linear model in both tasks.

A key finding was cross-modal performance: the CNN still decoded significantly when trained only on cortical regions not relevant to the task. The authors interpret this as evidence that imagined stimuli are represented in distributed, partially overlapping neural networks across modalities rather than in a single sensory area.

The paper says this cross-modal decoding capability highlights the potential of deep learning models to capture complex, multimodal neural patterns, and suggests future brain-computer interfaces could benefit from integrating auditory and visual information, pointing toward more flexible and personalized BCI designs. The findings speak to both cognitive neuroscience and BCI research, though the small sample means they need validation at larger scale.

Compiled by BCIwiki from public sources

Sources · 1
Read original ↗
Suggest a correction Revisions · none / EN
© 2026 BCIwiki.com Digest Topics Tips Subscribe Revisions About