CreditRule
Enums · AiDotNet.Enums
Selects the built-in credit-assignment (“learning”) rule used to produce per-parameter updates
during neural-network training via AiModelBuilder.ConfigureCreditRule(...). The rule decides how
the error signal is routed back to each layer; the forward pass, optimizer, batching and scheduler are
unchanged. Use the ICreditRule<T> overload of ConfigureCreditRule to supply a custom rule.
For Beginners
when a network makes a mistake, it must figure out how much each weight was to
blame so it knows how to change it. Standard Backprop sends the blame backwards through
the exact transpose of every weight. The alternative rules here send it through a fixed random
(or sign-only) shortcut instead — a family of “biologically plausible” methods that surprisingly still
learn because the forward weights rotate to align with the random feedback over training.
Fields
| Field | Summary |
|---|---|
CreditRule Backprop |
Standard reverse-mode back-propagation (the default). |
CreditRule DFANormalized |
Normalized Direct Feedback Alignment — Launay et al., 2020 style. |
CreditRule DRTP |
Direct Random Target Projection (DRTP) — Frenkel, Lefebvre & Bol, 2021 (“Learning without Feedback”). |
CreditRule DifferenceTargetPropagation |
Difference Target Propagation — Lee, Zhang, Fischer & Bengio, 2015. |
CreditRule DirectDifferenceTargetPropagation |
Direct Difference Target Propagation (DDTP-linear) — Meulemans et al., 2020. |
CreditRule DirectFeedbackAlignment |
Direct Feedback Alignment (DFA) — Nøkland, 2016. |
CreditRule DirectKolenPollack |
Direct Kolen-Pollack (DKP) — Kolen-Pollack learning applied in the Direct Feedback Alignment topology: the global output error is projected directly onto every hidden layer through per-layer feedback matrices, but those matrices are *le… |
CreditRule FeedbackAlignment |
Feedback Alignment (FA) — Lillicrap et al., 2016. |
CreditRule KolenPollack |
Kolen-Pollack (KP) — Kolen & Pollack, 1994; Akrout et al., 2019 (“Deep Learning without Weight Transport”). |
CreditRule LocalErrorSignal |
Local Error Signals — Nøkland & Eidnes, 2019 (“Training Neural Networks with Local Error Signals”). |
CreditRule SignSymmetric |
Sign-Symmetric feedback — Liao et al., 2016 / Xiao et al., 2018. |
CreditRule TiedReadoutLocalError |
Tied-readout local error — Local Error Signals for sequence models: each hidden layer is trained, with its input detached, by the network’s own loss applied to readout(h_k), where the readout is the network’s trailing norm + output he… |
int value__ |