Where does training collide?

The deliberate teaching of a skill so that it can later be applied, and the degree to which what is taught actually carries over to new situations. Thorndike and Woodworth's 1901 study The influence of improvement in one mental function upon the efficiency of other functions tested whether practising one task improves unrelated ones and found gains transfer only where the tasks share elements, which founded the study of transfer of training. Baldwin and Ford's 1988 review Transfer of training gathered the later evidence on how trainee, design, and workplace factors govern that carryover. Here the referent is a person acquiring a competence and applying it, measured by later performance on the job, not the fitting of a machine's parameters or the reshaping of a body's tissue.[1][2]

Unit

One trainee's acquired skill and its later application, per program

Recovery means

Skills can fade without practice or reinforcement on the job, a decay studied as transfer loss

Measured by

Assessed by tests of learning, on-the-job performance ratings, and behavior change after the program

The structured application of physical load so that the body adapts and its capacity rises, driven by repeatedly stressing a tissue beyond its habitual demand. Selye's 1936 paper A syndrome produced by diverse nocuous agents described how a living organism responds to any sustained stress in stages of alarm, resistance, and exhaustion, the general adaptation on which the overload principle and the idea of supercompensation rest: load, recover, and the system rebuilds slightly stronger than before. The referent is a physiological change in muscle, heart, or metabolism produced by exercise, measured in the body itself, and it is distinct from a skill carried to a new task or the numerical fitting of a model.[3]

Unit

One organism's capacity change in response to a training load

Recovery means

Adaptations reverse through detraining when the load is removed, and gains depend on recovery between bouts

Measured by

Tracked by markers such as maximal oxygen uptake, strength, and tissue and biochemical change over a program

The process of adjusting a model's internal parameters so that its outputs match example data, by repeatedly measuring error and shifting the parameters to reduce it. Rosenblatt's 1958 The perceptron introduced a machine that learns a task by correcting its connection weights from labelled examples, the first clear statement of training a device by parameter adjustment. Rumelhart, Hinton, and Williams's 1986 Learning representations by back-propagating errors gave the algorithm, backpropagation, that makes this practical for layered networks by propagating the error backward to update every weight. The referent is numerical optimization of a computational model against data, quantified by a loss function, and it shares nothing with a body adapting to exercise or a worker applying a taught skill beyond the word.[4][5]

Unit

One optimization run fitting a model's parameters to a dataset

Recovery means

Not applicable

Measured by

Gauged by the loss on training data and by error on held-out validation and test data