Stanford CS229 Machine Learning | Spring 2026 | Lecture 2: Supervised Learning Setup

Supervised learning functions by mapping input data to outputs through iterative processes that minimize prediction error. By adjusting internal parameters based on observed differences between predictions and actual labels, these models generalize patterns to unseen data.

• Supervised learning — A system learns by mapping input data (X) to known output labels (Y) provided in a training set [07:14].

• Regression tasks — These predict continuous numerical values, such as financial prices, rather than fixed categories [10:15].

• Classification tasks — These predict discrete labels, such as identifying if an image contains a specific animal [10:27].

• Linear models — Predictions rely on weighted combinations of input features to establish a trend [11:54].

• Error calculation — Performance is measured by reducing the distance between the model's prediction and the actual observed value [15:21].

• Gradient descent — Optimization involves calculating derivatives to find the direction that reduces error and adjusting parameters accordingly [35:35].

• Learning rate — This value defines the distance the model moves toward the optimal solution during each update cycle [36:32].

• Mini-batch updates — Processing small samples of data rather than the entire set improves computational performance when dealing with large datasets [43:59].

• Normal equations — A mathematical method provides a direct calculation for optimal weights in linear regression without requiring iterative steps [04:49].

• What distinguishes regression from classification in supervised learning?

• How does the use of mini-batches affect the training process?