Training accuracy is sitting at 98% but validation is stuck at 71% and not moving, even with a 0.3 dropout rate on the dense layers.
Dataset is a bit small (around 4,000 images) — could this just be a data volume problem rather than a regularization problem?
Vichet M.
ML practitioner
With only 4,000 images that gap is very likely a data problem before it's a regularization problem. Try aggressive augmentation first (flips, crops, color jitter) — it's usually cheaper than collecting more data and often closes most of that gap.
Sophea K.
Algorithms mentor · answered 22h ago
Also worth checking if your validation split is stratified the same way as training — an uneven split can create a gap that looks like overfitting but isn't.
Ratanak S.
Community moderator · answered 18h ago