Working vocabulary
| Term | Meaning |
|---|---|
| Derivative tower | Values of an activation and successive ordinary derivatives |
| Directional jet | Taylor coefficients along a selected direction; row k contains derivative divided by k! |
| Mixed jet | Coefficients D^alpha f / alpha! indexed by multi-index |
| Contraction | A requested aggregate such as the Hessian trace, computed without storing every tensor entry |
| Bias collapse | A normalized, weighted shift limit as delta tends to zero, yielding a derivative |
| Temperature hardening | A smooth decision becoming sharper as inverse temperature increases |
| Parameter autodiff | Differentiation of the training loss with respect to model parameters |
| Closed form | An analytic expression, whose floating-point evaluation can still have numerical error |
| Certificate | A checked statement about a specified quantity, domain, model, and assumptions |
| Release candidate | A prerelease version requiring explicit compatibility checks |
See numerical scope and the learning paths.