Introduction

The same table is a different problem under a different label regime, and there is no reinforcement-setting page here. The chapter first names the task, then walks the learning settings from weak and scarce labels to streams, few-shot data, and the choice of a model family. After this chapter the reader can name the task and the learning setting, and reinforcement learning is in Reinforcement Learning.

Tasks

The first step is to see where a problem sits among machine learning setups. The Overview maps how those setups differ by how much supervision and feedback they use, so you can place a problem before you pick a method. When one example carries more than one label, Multi Label Classification covers the main ways to assign several labels per example, with problem-transformation recipes, metrics, and tooling.

Learning settings

Once the task is named, the label regime decides the setting. Weakly Supervised is for labels that are incomplete, inaccurate, or inexact rather than fully supervised. Label Algorithms handles unbalanced labels, label propagation and spreading, and label noise, for when labels are scarce, noisy, or need to be spread across a graph of related examples. Semi Supervised is training with lots of unlabeled data and only a little labeled data. When labels are expensive, Active Learning chooses which examples a human labels next, with uncertainty and diversity sampling and query-by-committee. Online Learning is for models that update as labeled examples arrive in a stream, including non-stationary settings. N-Shot Learning covers zero-, one-, and few-shot learning, adapting with only a handful of labeled examples per class. With the task and the setting known, Model Families maps the problem shape to an algorithm family.