Early Stopping: Watching Validation Loss to Halt Training

Training Control

Ready — epoch 0

Early-Stopping Parameters

Patience (epochs) 5

Min Δ improvement 0.0010

Overfitting onset (epoch) 20

Validation noise 0.015

Live Readouts

Epoch

0 / 60

Bad epochs

0 / 5

Train loss

1.000

Val loss

1.000

Best val loss

1.000

Best epoch

How it works

Training loss keeps falling because the model can always fit the training set better. Validation loss falls too — until the model starts memorising training-set noise instead of learning general structure, at which point it turns back up. That turning point is overfitting, and early stopping is the algorithm that catches it automatically:

best = +inf, best_epoch = 0, bad = 0
for epoch e = 0, 1, 2, ...:
  train_model_one_epoch()
  v = validation_loss(e)
  if v < best - min_delta:
      best = v; best_epoch = e; bad = 0
  else:
      bad += 1
  if bad >= patience:
      stop training at epoch e
      restore weights from best_epoch

Patience — how many consecutive non-improving epochs are tolerated before halting. Low patience stops early and safely but can quit during a temporary noisy plateau; high patience rides out noise but risks more overfitting before it reacts.
Min Δ — how large an improvement must be to "count", so the algorithm ignores noise-sized wiggles instead of resetting its patience counter on them.
Overfitting onset — the epoch at which this simulated model starts memorising the training set; the validation curve (orange) turns upward from there while the training curve (blue) keeps sinking.
Validation noise — per-epoch measurement noise on the validation curve, the reason min-delta and patience both matter: a noiseless curve would never need either.

The gold marker on the orange curve is the best checkpoint found — the one early stopping actually keeps. The red plane marks the epoch where training was halted, always patience epochs after the true minimum, because that many bad epochs had to elapse before the algorithm could be sure the minimum had passed.

Training loss Validation loss Best checkpoint Stop epoch

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.