What is pattern recognition?

Pattern recognition is the discipline of having a machine assign a category or a structure to what it observes: an image, a signal, a document, a series of transactions. In practice it answers questions such as "is this part conforming or defective?" or "does this transaction look like fraud?". It is a task, not a technique: the goal is to decide which class an input belongs to, and the methods for doing so have changed over time. At first they were rules written by hand by an expert; today the dominant method is machine learning, in which the system derives the rules from examples. That is why the two expressions overlap in everyday use without being the same thing: machine learning is the most common tool, pattern recognition is the problem that tool solves. The field's international association, the IAPR, places it alongside computer vision and image processing.

See also: Machine Learning vs Deep Learning · Machine learning is the family of models that learn from data; deep learning is the neural-network subset powering modern AI.

From hand-written rules to learned models

The classic pipeline has three steps. First acquisition: a camera, a sensor, a file. Then feature extraction, meaning the measurements that summarize the input, such as the area of a stain or the frequency of a vibration; this is the work described in the feature engineering entry. Finally classification, which decides which class those measurements belong to. Historically there were two approaches: statistical, which treats the features as numbers and computes which class they most likely belong to, and structural (or syntactic), which describes the input as a composition of parts and rules.

Today the last two steps are often merged into a neural network that learns by itself which features matter; the distinction from deep learning is covered in machine learning vs deep learning. Learning is supervised when the examples are already labeled ("defective", "conforming"), and unsupervised when the system must group by itself what looks alike. Reading printed text, OCR, is just one application.

An enterprise example

A company that molds plastic components wants to stop parts with micro-burrs before packaging. A camera at the end of the line photographs every part, and the system compares the image with thousands of examples already judged by an inspector and flags the suspicious ones. The same logic reads the vibration signals of a motor to anticipate a failure, as described in predictive maintenance, or sorts incoming emails by type of request. In the first months the system errs in instructive ways: it rejects good parts because the glare of a light looks like a burr. The longest part of the project is controlling the capture conditions and building an archive of examples judged by consistent criteria.

Why it matters for decision makers

A pattern found by a machine is not proof. With rich enough data a model always finds some regularity, and some of it is accidental: it is the automated version of apophenia, the human tendency to see designs in noise, which in everyday language is the best-known sense of "recognizing patterns". Its technical counterpart is overfitting: the model memorizes the examples and then fails on new cases. The defense is validation on data the model has never seen, and it is worth asking for before approving the budget. Three questions are useful: what a false alarm costs, what a missed defect costs, and who rechecks the doubtful cases. The decision threshold depends on the answers far more than on the choice of algorithm.