From 499d800ad5897554056a1f0f7198b4f9e1ccb721 Mon Sep 17 00:00:00 2001 From: Pierre Guetschel Date: Fri, 1 May 2026 09:53:58 +0200 Subject: [PATCH] fix comment --- open_eeg_bench/ridge_probe.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/open_eeg_bench/ridge_probe.py b/open_eeg_bench/ridge_probe.py index 52641d9..ddc15a7 100644 --- a/open_eeg_bench/ridge_probe.py +++ b/open_eeg_bench/ridge_probe.py @@ -246,7 +246,8 @@ def _fit_streaming_ridge( ) # tensor of shape (K,) # Among tied best scores, pick the largest λ (most regularization): - # more parsimonious, fewer effective degrees of freedom, numerically stabler. + # when val cannot discriminate (e.g. discretized scores on small val sets, + # plateaus), the most regularized solution is the safest bet. # NaN-aware: a NaN val score means the val pass hit non-finite features # or predictions — exclude it from both max and tie-breaking. finite_mask = torch.isfinite(val_scores)