Package crino :: Module criterion

Module criterion

The criterion module provides some differentiable loss functions in order to perform a gradient descent on a hand-crafted neural network.

The currently implemented criterions are :

See their respective documentations for their mathematical expressions and their use cases.

See Also: module

Classes

[hide private]

Criterion
The Criterion class handles the loss computation between ŷ (the outputs vector of a Module) and y (the targets vector).

CrossEntropy
The cross-entropy criterion is well suited for targets vector that are normalized between 0 and 1, used along with a final Sigmoid activation module.

MeanSquareError
The mean square error criterion is used in the least squares method, it is well suited for data fitting.

MeanAbsoluteError
The mean absolute error criterion is used in the least absolute deviations method.

NegativeLogLikelihood
The Negative Log Likelihood criterion is used in classification task.