Short answers, pulled from the story.
Pattern recognition is the assignment of a label to a given input value based on patterns extracted from data. In machine learning it covers tasks like classification, regression, and sequence labeling. Supervised approaches use hand-labeled training data; unsupervised approaches find inherent structure without labels.
Pattern recognition algorithms aim to find the most likely match for an input, taking statistical variation into account. Pattern matching algorithms look for exact matches to pre-existing patterns, with regular expression matching being a common example.
The first pattern classifier was the linear discriminant, developed by Fisher in the frequentist tradition. Its parameters are the mean vectors and covariance matrix, estimated directly from collected data.
Applications include computer-aided diagnosis for cancer screening (such as Papnet for cervical cancer), automatic speech recognition, face detection, license plate recognition, handwriting recognition on postal envelopes, and autonomous vehicle technology.
Supervised learning uses hand-labeled training data to build a model, while unsupervised learning finds inherent patterns in unlabeled data. The unsupervised equivalent of classification is called clustering, which groups inputs by similarity rather than assigning them to predefined classes.
Probabilistic classifiers output a mathematically grounded confidence value alongside their label choice, allowing them to abstain when confidence is too low. They can also output ranked lists of N-best labels with probabilities, and they integrate more cleanly into larger machine-learning pipelines by reducing error propagation.