= Solution
The dashed curve is training accuracy: it continues to rise as optimization adapts to the training observations. The solid curve is testing accuracy: it peaks near 10 epochs and then declines. The widening gap is <overfitting>.
A sensible choice is about 10 epochs, selected by <early stopping> at the maximum validation accuracy. In a proper analysis, a validation set rather than the final test set should choose this epoch. Early stopping is an implicit <regularization> method because it limits how far the parameters can adapt to training-specific noise.
Solved by gpt-5.6-sol high.
Back to article page
