ML

Machine learning is the branch of computing that writes a function from examples instead of from first principles: the engineer picks a family of candidate functions, a loss that says how wrong each one is on the data and a price for complexity, and an optimizer chooses the member of the family that best explains the examples. It is used wherever a rule exists but nobody can write it down (a camera telling a scratched part from a clean one, a vibration signature that announces a failing bearing, the next word of a sentence), and it is paid for in data, compute and the upkeep of a model that ages as the world changes.

Machine learning is the branch of computing that writes a function from examples instead of from first principles: the engineer picks a family of candidate functions, a loss that says how wrong each one is on the data and a price for complexity, and an optimizer chooses the member of the family that best explains the examples. It is used wherever a rule exists but nobody can write it down (a camera telling a scratched part from a clean one, a vibration signature that announces a failing bearing, the next word of a sentence), and it is paid for in data, compute and the upkeep of a model that ages as the world changes.

Almost all of it fits one line, empirical risk minimization: average the loss over the examples, add a weighted penalty for complexity, minimize over the family. What the line leaves out is the actual goal, a model that is right on data that do not exist yet, which is why generalization and honest evaluation are half of the subject. Beside a physical model it sits at the black end of the white, grey and black box scale: it needs data covering the whole range of operation and says little outside it (model versus data).

The branches split by what the examples carry. Supervised learning has the right answer for each example (a label) and learns to predict it; unsupervised learning has none and looks for structure (groups, rarities, principal directions); reinforcement learning has only a reward for its actions and learns a policy by trying. Anomaly detection borrows from the second to answer a question that looks supervised, with almost no failures to learn from.

The families split by how they generalize. Linear models, trees and their ensembles rule tables of features; kernel methods predict by resemblance to stored examples; neural networks win when the data are images, sound, raw signal or text, and each architecture (CNN, RNN, transformer) assumes something about the data it reads.

Using it in a plant or a vehicle is a discipline of its own. Applied ML covers what to ask before any model, where labels come from and which method to try first, and MLOps covers keeping the model right in production. Industrial production runs mostly on thresholds, regressions, PCA and gradient boosting, and what separates the projects that reach the plant from the demos is rarely the algorithm.

Training runs on GPU clusters or TPU accelerators for large networks, and in minutes on a laptop CPU for most classical methods. Inference can happen on cloud GPUs, edge NPU cores or even MCU devices through TinyML: a linear regression is a dot product that fits a one-euro microcontroller running at 1 kHz.