Why it matters

Reading loss curves shapes training decisions. Standard practice.

Advertisement

The architecture

Track: training + validation loss.

Watch: spikes, plateaus, gap.

Loss curve readingTraining losssmooth desiredValidation lossgap = generalizationAnomaliesspikes / plateausSpikes = instability; plateaus = LR issue or capacity limit
Loss curves.
Advertisement

How it works end to end

Spikes: instability (grad clipping? LR too high?).

Plateaus: LR schedule + data issue.

Gap: overfitting.

Log at multiple scales.