IOutputDerivative<T>

Interfaces · AiDotNet.Interfaces

Interface for activation functions that can compute their derivative given the post-activation output value rather than the pre-activation input. This avoids the common bug where the derivative re-applies the activation (e.g., sigmoid(sigmoid(x)) instead of sigmoid(x)*(1-sigmoid(x))).

Known implementations: SigmoidActivation<T>, TanhActivation<T>

Methods

Method Summary
Tensor<T> DerivativeFromOutput(Tensor<T> output) Computes the activation derivative given the post-activation output value.