Skip to content

Evaluate BCI IV-2a with three fixed training seeds #63

Description

@GTC2333

Problem

The BCI IV-2a grader currently trains each subject once with seed 42. Model initialization, mini-batch order, Dropout, and BatchNorm can affect the selected validation checkpoint and the final Session-E Kappa.

Trigger scenario

The same submitted architecture produces noticeably different Session-E results across training seeds, making a single-seed score unusually high or low.

Impact

  • Model rankings may depend on one initialization.
  • Stable architectures cannot be distinguished from lucky runs.
  • Small score differences between agents or harnesses may not be meaningful.

Expected fix

Keep the current within-subject cross-session protocol and the same fixed 230/58 train-validation split, but train every subject with three predefined seeds, for example [42, 123, 2026].

For each seed, continue selecting the checkpoint by validation loss and evaluating it on that subject's private Session E. Report:

  • mean Kappa across seeds as the primary score;
  • Kappa standard deviation across seeds;
  • per-seed results;
  • per-subject mean and standard deviation.

This issue does not propose K-fold cross-validation or changing the existing train-validation split.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions