Deep Learning (DL) · Intermediate level

Intermediate level · Module 5: Evaluating Deep Learning Models · Lesson 3 of 3

Diagnosing Underfitting and Overfitting

Use training and validation curves to decide whether to make the model bigger, regularise it or get more data.

In this lesson you will

  • Diagnose underfitting and overfitting from curves
  • Choose the right remedy

Key concepts

training vs validation curvesunderfitting remediesoverfitting remediesdata vs model problems

The lesson

Underfitting — poor training and validation performance — calls for a larger model, longer training or better features.

Overfitting — high training but lower validation performance — calls for more or better data, augmentation, regularisation or a smaller model.

If validation is good but the field test is poor, the problem is the data's coverage, not the model: collect data from the missing conditions.

Step by step

  1. Plot training and validation curves for both models.
  2. Diagnose each model.
  3. Apply one remedy and retrain.
  4. Compare the curves.
Troubleshooting tree

My deep learning model is disappointing — what now?

Use training, validation and field results to find out whether the model, the training or the data is the problem.

Is training accuracy low too?
  • Yes
    Underfitting: train longer, raise the learning rate a little, use a larger or pre-trained model — and check the labels and data loading for bugs.
  • No
    Is validation accuracy much lower than training accuracy?
    • Yes
      Overfitting: add realistic augmentation, dropout or weight decay, use early stopping, or collect more varied data.
    • No
      Is the field or test performance much lower than validation?
      • Yes
        Data mismatch or leakage: check for near-duplicates across splits and collect training data from the real conditions (phones, lighting, places).
      • No
        Performance is consistent. Use error analysis to find the most common remaining mistakes and target them with data or labels.

Notes

Always look at examples, not just numbers: a few dozen misclassified images often explain more than any metric.

Practice activity

Apply the diagnosis to the project's models and complete the evaluation report.

Expected result: A diagnosis per model, one tested remedy and a final recommendation.

Check your understanding

  1. Which remedy suits overfitting?

    Show answer

    More data, augmentation, regularisation or a smaller model.

  2. What does good validation but poor field performance suggest?

    Show answer

    The training data does not cover real conditions.

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.