Where does dropout collide?

A training trick for artificial neural networks in which a random fraction of the network's units is temporarily switched off, set to zero, on each training step, so that the network cannot lean too heavily on any one unit and is pushed to learn features that work in many combinations. Srivastava, Hinton, Krizhevsky, Sutskever, and Salakhutdinov's 2014 Dropout: A Simple Way to Prevent Neural Networks from Overfitting introduced the method as a cheap way to stop a model from memorising its training data, showing it behaves like averaging over a huge collection of thinned networks. The referent here is a deliberate computation inside a learning algorithm, chosen by the engineer to improve a model, and it has nothing to do with a person leaving school or a patient leaving a study.[1]

Unit

A per-step random mask that zeroes a fraction of units, set by the drop probability

Recovery means

The dropped units are restored on the next step, and at test time all units are kept with their outputs rescaled

Measured by

Set by the drop probability and judged by its effect on validation error and the gap between training and test accuracy

A student who leaves a school or degree programme before completing it, and, as a group measure, the rate at which students do so: dropping out is treated as an outcome of a gradual process in which a student's ties to the institution weaken until departure follows. Tinto's 1975 Dropout from Higher Education: A Theoretical Synthesis of Recent Research built an influential model in which academic and social integration shape whether a student stays or leaves, and framed the dropout rate as a central measure of an institution's success. The referent is a person and a life decision, tracked as a headline education-policy statistic, not a computation in a model or a gap in a dataset.[2]

Unit

One student who departs before completion, or the share of a cohort who do

Recovery means

A dropout may re-enrol later, and re-entry and completion are tracked separately from the initial departure

Measured by

Counted as a dropout rate over a cohort, from enrolment and completion records, or by longitudinal tracking of students

A participant who leaves a trial or study before it finishes, so that their planned measurements are never taken: such losses, also called attrition or loss to follow-up, matter because the people who leave often differ from those who stay, which can bias the result. Dumville, Torgerson, and Hewitt's 2006 Reporting attrition in randomised controlled trials showed that dropout is frequently under-reported and mishandled, and argued that how missing participants are counted can change a trial's conclusion. The referent is a gap in the evidence created when a real subject stops taking part, a threat to a study's validity, not a design choice in a neural network or a young person quitting school.[3]

Unit

One enrolled participant lost before study completion, or the share of a sample lost

Recovery means

Some dropouts are recovered by continued follow-up, and their missing outcomes are addressed by intention-to-treat handling and imputation

Measured by

Counted as an attrition or dropout rate, reported by group in a flow diagram, and probed for whether the losses differ between arms