Skip to main content

Machine Learning

Machine learning studies systems whose behavior is fitted from data or interaction rather than fully specified by hand. The algorithm is only one part of the system: the question, data-generating process, objective, evaluation design, and deployment conditions determine whether the result is useful.

Problem Families

FamilyLearning signalTypical question
Supervised learningpaired inputs and targetspredict an outcome for a new case
Unsupervised learningobservations without task labelsdiscover structure or a useful representation
Self-supervised learningtargets constructed from the data itselflearn reusable representations or conditional predictions
Reinforcement learningactions, state transitions, and rewardschoose behavior that improves long-run return

These families can overlap. A deep neural network describes a model class, not a separate source of supervision.

Working Sequence

  1. Define the decision or prediction target and an explicit baseline.
  2. Inspect collection, sampling, labels, missingness, and possible leakage.
  3. Choose a loss that represents the training objective and metrics that represent actual use.
  4. Fit only on training data; use validation data for choices and hold the test set for final evaluation.
  5. Compare errors across relevant groups, time periods, and operating conditions.
  6. Monitor distribution shift, feedback loops, privacy, and failure cost after deployment.

Site Path

Berkeley CS 189 is the primary external course path. Product services, current model APIs, and benchmark snapshots belong elsewhere because they change faster than this foundation.