PRODUCTION SCENARIO
A grocery chain trains an image classifier to spot damaged produce. Over 60 epochs, training accuracy rises from 82% to 99%. Validation accuracy peaks at 88% near epoch 20 and then falls to 81%.
Which action should the team take?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Increase regularization and keep the model from the validation peak.
Training accuracy that keeps rising while validation accuracy starts to decrease is the signature of overfitting, and AWS documents increasing regularization as the adjustment; the epoch-20 model is the one that generalizes best. A lower learning rate changes how fast the weights move, not the model's freedom to memorize, so the curves diverge again later. Batch size changes gradient noise and throughput, not the generalisation gap. Adding layers gives the model more room to memorize when it is already at 99% on the training set.