Since recently, started approximately following the nomenclature used by Torch neural network package, since the names appear to me to be:
- concise
- simple
- easy to understand
Concretely, the names for some things, nouns are:
- input: the input to the layer, coming from previous layer
- output: the output from the layer, going to the next layer
- gradInput: the partial gradient of the loss with respect to the input to the layer
- gradOutput: the partial gradient of the loss with respect to the output of the layer
- gradWeights: partial gradient of the loss with respect to the weights (renaming not done yet)
Some verbs are:
- forward: calculate the output, given the input
- backward: calculate gradInput, given gradOutput
(Note: these names are a work in progress, there are still plenty of old names about, like:
- errorsForUpstream => gradInput
- outputFromUpstream => input
- errors => gradOutput
- propagate => forward
- backprop => backward
- etc ... )