Quadratic Test Problems

DeepOBS includes a stochastic quadratic problem with an eigenspectrum similar to what has been reported for neural networks.

Other stochastic quadratic problems (of different dimensionality or with a different Hessian structure) can be created easily using the net_quadratic_deep class.

class deepobs.pytorch.testproblems.testproblems_modules.net_quadratic_deep(dim, Hessian)[source]

This arhcitecture creates an output which corresponds to a loss functions of the form

\(0.5* (\theta - x)^T * Q * (\theta - x)\)

with Hessian Q and "data" x coming from the quadratic data set, i.e., zero-mean normal. The parameters are initialized to 1.

Test Problems