Deep Learning (DL) · Beginner level

Beginner level · Module 5: Training Neural Networks Well · Lesson 3 of 3

Regularisation, Augmentation and Early Stopping

Prevent overfitting so the model works on new photos, not just the training set.

In this lesson you will

  • Apply dropout, weight decay and data augmentation
  • Use early stopping with validation monitoring

Key concepts

dropoutweight decaydata augmentationearly stopping

The lesson

Deep networks can memorise small data sets. Dropout randomly switches off neurons during training; weight decay keeps weights small. Both make memorising harder.

Data augmentation creates new training examples by flipping, rotating, cropping and changing the brightness of images — realistic variations of field photos.

Early stopping watches validation loss and stops when it no longer improves, keeping the best weights. Together these methods let small data sets produce useful models.

Step by step

  1. Add augmentation to the training transforms.
  2. Add dropout and weight decay.
  3. Add early stopping on validation loss.
  4. Compare the gap between training and validation accuracy.

Practice activity

Complete the training improvement report for the leaf disease classifier.

Expected result: A report showing a smaller train–validation gap and better test performance, with each change's effect.

Check your understanding

  1. Which augmentations suit leaf photos?

    Show answer

    Realistic ones such as flips, small rotations, crops and brightness changes.

  2. What does early stopping do?

    Show answer

    Stops training when validation performance stops improving.

Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.