Home / Summaries / COMP 3711 / validation-training-accuracy

Learning Goal - Overfitting and Detection

Q: Define overfitting in the context of neural networks

Overfitting occurs when a network has so many free parameters that it fits the specific noise and quirks of the training data rather than the underlying true pattern. Result: very high training accuracy but poor test accuracy — the model fails to generalise .

Q: How do you detect overfitting during training?

Split data into training , validation , and test sets. Monitor: - Training accuracy/cost : typically improves continuously - Validation accuracy/cost : improves, then plateaus or degrades The gap between training and validation performance is the overfitting signal. When validation accuracy stops improving, the network is memorising training-specific noise.

Q: What is early stopping and why must it use the validation set, not the test set?

Early stopping : halt training when validation accuracy stops improving (saturates). Must use validation set because: if you use the test set to decide when to stop, that test data has effectively entered the training process — you're now choosing a stopping point that fits the test set's peculiarities. Your final accuracy estimate would be overoptimistic and biased , no longer a true measure of generalisation to unseen data.

3 important questions on Learning Goal - Overfitting and Detection

Define overfitting in the context of neural networks

Overfitting occurs when a network has so many free parameters that it fits the specific noise and quirks of the training data rather than the underlying true pattern. Result: very high training accuracy but poor test accuracy — the model fails to generalise.

How do you detect overfitting during training?

Split data into training, validation, and test sets. Monitor:
- Training accuracy/cost: typically improves continuously
- Validation accuracy/cost: improves, then plateaus or degrades

The gap between training and validation performance is the overfitting signal. When validation accuracy stops improving, the network is memorising training-specific noise.

What is early stopping and why must it use the validation set, not the test set?

Early stopping: halt training when validation accuracy stops improving (saturates).

Must use validation set because: if you use the test set to decide when to stop, that test data has effectively entered the training process — you're now choosing a stopping point that fits the test set's peculiarities. Your final accuracy estimate would be overoptimistic and biased, no longer a true measure of generalisation to unseen data.

The question on the page originate from the summary of the following study material:

COMP 3711

View summary