Short answers, pulled from the story.
Supervised learning operates by feeding an algorithm labeled input-output pairs so it can map new inputs to correct outputs. This paradigm allows statistical models to learn from supervision as defined in the 2012 book Foundations of Machine Learning by Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalkar.
Researchers S. Geman and E. Bienenstone published a paper titled Neural networks and the bias/variance dilemma in 1992. Their study clarified how machines learn from examples and described the tradeoff between bias and variance in their 1992 study Neural Computation 4, pages 1, 58.
Engineers use regularization techniques to control the bias-variance tradeoff and prevent overfitting when training models on data. Structural risk minimization includes a regularization penalty that acts like Occam's razor by preferring simpler functions over more complex ones.
Support-vector machines, decision trees, and neural networks stand among the most widely used supervised learning algorithms today. Linear regression and logistic regression handle continuous values and binary classification tasks respectively while boosting serves as a meta-algorithm combining multiple weak learners into strong predictors.
Semi-supervised learning provides labels only for a subset of training data while leaving other samples unlabeled or imprecisely labeled. This approach combines with active learning strategies to maximize efficiency when perfect labeling remains expensive or impossible in real-world scenarios.