C.40:5.7 - A discovered learning component depends on how it was trained

A neural-network developer can vary the formula used during training while keeping the receiving question about the resulting model. In TaylorGLO, a vector specifies a polynomial training loss; each trial trains a model with that loss, and validation accuracy guides the outer search. The vector’s own loss value is not the outcome compared across candidates. Short training makes initial trials cheaper, while later full training examines the selected result. The polynomial representation supplies a particular smooth search space, not every possible learning rule.

PANGAEA makes a different coupled choice: evolution changes an activation function’s computational form, while gradient training adjusts parameters within that form along with the network. Returning only the form and omitting its parameter initialization and training changes the proposed application. Its source comparisons include fixed functions and parameterized functions, with transfers that succeed on some architectures and fail on others. This makes fitting a complete candidate and checking its receiving architecture consequential operations.

EPBT instead continues partially trained models while changing settings and training loss. Its selected final model carries earlier training history. Treat that as a continued-development result; replaying only the final loss from a fresh initialization is a new test. AQuaSurF selects expensive activation trials with a fitted predictor based on function behavior and network-derived features, then updates from evaluated outcomes. Constructing those features also costs computation. These alternative realizations of §4.7 differ in what they vary, what they retain and what their evaluation costs. The linked primary sources in §11 supply their neural mechanisms; the scheduling construction shows what can be retained without those mechanisms.