public override backwardTraining(context as LayerPass<T>, upstream as NDArray<T>) as LayerGradients<T>
Reuses saved probabilities rather than evaluating exponentials again.
| Parameter | Description |
|---|---|
context | Context returned by forwardTraining. |
upstream | Derivative with respect to the probabilities. |
Input derivatives without modifying saved outputs.
dx = activation.backwardTraining(context, upstream).inputs