Training-time averaging leads across most MEG tasks

Non-invasive brain recordings are noisy, and researchers can combine repeated recordings of the same stimulus to improve decoding. But should they combine the recordings before a model processes them, after it makes predictions, or as part of training?

This study tests whether that design choice changes how accurately models classify speech and imagined digits.

The authors train a shared neural network to average class scores across repeated recordings before calculating its loss, then compare it with five alternative aggregation methods.

Can the way researchers combine repeated brain recordings change how well a model decodes them? The paper tests training-time logit averaging, which combines class scores across recordings before calculating the learning loss, against five other aggregation methods.

Across seven MEG conditions, researchers found the highest mean held-out test balanced accuracy in six, averaging results across five seeds per condition. Balanced accuracy averages recall across classes.

The imagined-digit task was the exception: post-hoc probability averaging scored higher, suggesting the best aggregation method depends on the task.

These findings come from MEG recordings and classifiers trained with cross-entropy loss. One of the seven settings used MegNIST, which is not yet publicly available, so that result is harder to reproduce.