Generalization: Difference between revisions

Generalization in machine learning refers to how accurately can a trained model correctly predict new, unseen data. A model that generalizes well is the opposite of one that overfits on [[training data]]. It's an essential concept in machine learning since it allows models to be applied in real-world problems where input data may change frequently.

Machine learning models are trained by optimizing their parameters to minimize the difference between predictions and actual outcomes in training data. If the model becomes overfitted to this training data, it may become complex and unable to generalize well to new information. Overfitting occurs when the model fits noise rather than underlying patterns in the training data; consequently, it becomes too specialized for new data sets and performs poorly when given new ones.

On the contrary, underfitting occurs when a model is too simplistic and fails to capture underlying patterns in training data. An underfit model will also perform poorly on new data. Thus, machine learning seeks a balance between overfitting and underfitting; whereby it can reliably capture patterns while also avoiding fitting noise into its predictions.