Where does feature collide?

In phonology, a distinctive feature is a minimal property of a speech sound that separates one meaningful sound from another, such as whether the vocal cords vibrate or whether the lips are rounded. Trubetzkoy's 1939 Grundzüge der Phonologie analysed sound systems as networks of contrasts between such properties, and Jakobson, Fant, and Halle's 1952 Preliminaries to Speech Analysis recast them as a small set of binary acoustic features that together describe every phoneme. The referent is an abstract contrastive property of language, defined by the oppositions it marks, and it has nothing to do with a buried trace at a dig or a measured number fed to a classifier.[1][2]

Unit

One binary or privative property of a speech sound, present or absent

Recovery means

Not applicable

Measured by

Inferred from minimal pairs and confirmed by acoustic and articulatory correlates

In pattern recognition and machine learning, a feature is a single measurable property of the thing being classified, and the ordered list of such measurements for one item is its feature vector, a point in feature space that a model reads. Marill and Green's 1963 On the effectiveness of receptors in recognition systems studied how well a chosen set of measurements separates the classes to be recognised, and Duda and Hart's 1973 Pattern Classification and Scene Analysis fixed the now standard vocabulary of features, feature vectors, and feature space. The referent is a number or attribute computed from data as input to a decision rule, not a property of a speech sound or a nonportable remain in the earth.[3][4]

Unit

One measured property of an item, one coordinate of its feature vector

Recovery means

Not applicable

Measured by

Extracted from raw data by hand-designed or learned transforms and judged by how well it separates classes