Where does propagation collide?

The travel of a disturbance through a medium, carrying energy and shape from one place to another at a speed set by the medium's properties rather than by any motion of the medium itself. Whitham's 1974 Linear and Nonlinear Waves is a standard treatment of how such disturbances move, governed by the wave equation, and how their speed and form depend on the medium. The referent here is a physical wave moving through matter or a field, timed by a propagation speed, and it has nothing in common with a plant grown from a cutting, an error signal sent backward through a network, or a habit spreading between people.[1]

Unit

A disturbance moving through a medium, scaled by its propagation speed

Recovery means

The medium returns toward rest after the wave passes, and damping sets how quickly it does so

Measured by

Characterised by measuring wave speed, wavelength, and how amplitude changes with distance

In training a neural network, the rule that sends the output error backward through the layers to work out how much each internal weight contributed to it, so the weights can be adjusted to reduce that error. Rumelhart, Hinton, and Williams's 1986 paper Learning representations by back-propagating errors made this procedure widely known and showed it could learn useful internal representations. The referent here is a computation over a network's weights, measured by how fast the error falls, and it is unrelated to a wave in a medium, a plant grown from a cutting, or a behaviour catching on across a population.[2]

Unit

An error-driven weight update passed backward through a network

Recovery means

Not applicable

Measured by

Tracked by the training loss and how quickly it falls over repeated passes through the data