The paper notes that its theoretical validation-loss approximation holds for SGD, whereas its practical multidimensional models are trained with Adam; successful controlled NTK-based amplitude tuning should not be treated as established for those Adam tasks. · CiteArk
The paper notes that its theoretical validation-loss approximation holds for SGD, whereas its practical multidimensional models are trained with Adam; successful controlled NTK-based amplitude tuning should not be treated as established for those Adam tasks.
