Where does augmentation collide?

In machine learning, augmentation means enlarging a training set by adding transformed copies of the examples already in it, so a model sees more variety without anyone gathering new data: an image might be flipped, rotated, cropped, or recoloured, and each altered copy still carries the original label. Shorten and Khoshgoftaar's 2019 survey A survey on Image Data Augmentation for Deep Learning catalogued these transformations and showed how they curb overfitting when data is scarce. The referent here is a data-processing step that multiplies examples to train a model, and it has nothing to do with lengthening a musical theme, enlarging a body part in surgery, or adding a second drug to lift a mood.[1]

Unit

One transformed copy of a training example, still bearing its original label

Recovery means

Not applicable

Measured by

Gauged by how much a model's validation accuracy improves, and by the number and variety of transforms applied

In music, augmentation means restating a melody or theme in longer note values, typically doubling each duration, so the tune unfolds at half its former speed while keeping its shape: it is a standard device in fugue and counterpoint, often paired with its opposite, diminution, which shortens the values. The same word also names an augmented interval or chord, one widened by a semitone beyond its usual size. As set out in the article on augmentation in The New Grove Dictionary of Music and Musicians (2001), the word marks a stretching of durations or intervals. This referent is a compositional operation on rhythm or pitch, unrelated to copying training data, enlarging tissue, or combining medicines.[2]

Unit

One theme restated in proportionally longer durations, or one interval widened by a semitone

Recovery means

Not applicable

Measured by

Read from the score as the ratio of augmented to original note values, or named by the widened interval

In psychiatry and pharmacology, an augmentation strategy means adding a second agent to a treatment that has only partly worked, so the added drug boosts the effect of the first rather than replacing it: the classic case is adding lithium to an antidepressant for a patient whose depression has not lifted. De Montigny, Grunberg, Mayer, and Deschenes's 1981 report Lithium Induces Rapid Relief of Depression in Tricyclic Antidepressant Drug Non-Responders described this rapid effect and helped establish lithium augmentation as a recognised tactic. The referent is a prescribing move that combines drugs to raise a partial response, and it is unlike copying data to train a model, stretching a musical theme, or enlarging a body part.[3]

Unit

One added agent given on top of a partly effective treatment

Recovery means

Not applicable

Measured by

Judged by the improvement in symptom scores after the second agent is added, against the partial response before