Machine Learning

Machine learning (ML) is the branch of artificial intelligence that develops algorithms able to learn patterns from data and generalize to unseen data, solving tasks without having been explicitly programmed for them. A classic operational definition is Tom Mitchell's: a program learns from experience—the data—with respect to a task and a performance measure if its performance at that task, as measured by that metric, improves with experience.

The main paradigms

Machine learning is traditionally organized into three paradigms. In supervised learning, the model learns from labeled examples—input and desired-output pairs—to map new inputs to their correct output. In unsupervised learning, without labels, it discovers hidden structure in the data, such as groupings or patterns. And in reinforcement learning, an agent learns by trial and error by interacting with an environment and maximizing a reward. To these is added today self-supervised learning, which generates its own signal from the data and is the engine behind the pre-training of large models.

Where it fits

Machine learning is a subset of artificial intelligence, and deep learning—based on neural networks with many layers—is in turn a subset of machine learning. That is: deep learning is within machine learning, which is within artificial intelligence.

The process and its limits

The basic flow goes from the data to training the model and then to evaluating it on unseen data, because what matters is not being right on the training data, but generalizing. Its limits are real: overfitting, when the model memorizes the training noise and fails on new data; bias, which a model learns from biased data and amplifies at scale; and dependence on data quality, per the old principle that garbage in, garbage out.