Overfitting

• The model memorizes the training data and fails on new data.
• Looks great on training set, bad on validation/test set.
• Sign: training loss keeps going down but validation loss starts rising.
• Common cause: model too complex or too little / unrepresentative training data.
• Quick fix ideas: more data, simpler model, regularization, early stopping, cross-validation.

Underfitting

• The model is too simple and cannot learn the training data well.
• Bad performance on both training and validation sets.
• Sign: both training and validation loss are high and don’t improve much.
• Fix: use a more powerful model, add relevant features, train longer.

Good fit / Generalization

• The model performs well on new, unseen data (validation/test).
• Training and validation loss are similar and both low.