IFeedbackLearningRule<T>

Interfaces · AiDotNet.Interfaces

Optional extension of ICreditRule for rules whose feedback state is learned across training steps (e.g. Kolen-Pollack and Direct Kolen-Pollack, whose feedback matrices are updated each step to align with the forward weights). The training engine calls OnParametersUpdated once per gradient computation, after the step’s teaching signals and gradients have been produced, while the per-layer forward activations are still available on the context. Rules with fixed (non-learned) feedback do not implement this interface, so existing rules are entirely unaffected.

Implements: ICreditRule<T>

Known implementations: CreditRuleBase<T>

Methods

Method Summary
void OnParametersUpdated(ICreditAssignmentContext<T> context) Updates the rule’s learned feedback state for this training step, using the same context (forward activations, output error and teaching signals) that produced the step’s gradients.