OM DENNE EPISODE
Deep learning often beats older methods at telling which hand you are imagining moving, but a higher score does not show what the software was paying attention to.
Researchers at Inria and the University of Bordeaux took recordings from 139 people who imagined moving a hand whenever an on-screen arrow told them which one, and tested two deep learning models and the most widely used older method on them. Then they started every clip half a second later, after the arrow had appeared.
On the larger dataset, the best deep learning model dropped from 88.7% to 76.1%, while the older method barely moved. Heat maps of what each model relied on explain the gap, and the answer has more to do with the screen than the hand. That matters outside the lab, where there is no arrow, and in stroke rehabilitation, where a system trained this way could end up rewarding the wrong signal. The work used previously recorded data and compared left hand against right hand only.
FutureProof: In Plain English is a daily science podcast hosted by Brett Beckermann for curious people without a science background.
Source: Trocellier, D., Kojima, S., N'kaoua, B., and Lotte, F. (2026). Visual cues in MI-BCI induce biases in EEG classification models with deep learning but not with standard machine learning. Frontiers in Neuroergonomics, 7, 1804143. https://doi.org/10.3389/fnrgo.2026.1804143
