A model that scores 99 percent on training data and 60 percent on new data has learned the wrong lesson. Overfitting means memorizing noise and idiosyncrasies instead of the underlying pattern, and it is the most common failure in applied machine learning.

Signals are easy to spot. Training loss keeps falling while validation loss rises. The model performs well on examples it has seen and poorly on anything else. Complexity, small datasets, and long training all increase the risk.

Common countermeasures

• More training data
• Regularization such as weight decay
• Dropout during training
• Early stopping based on validation loss
• Simpler model architectures
• Data augmentation

Underfitting is the opposite problem, where the model is too simple to capture the pattern at all. The goal is the middle ground, where the model generalizes without memorizing.