Re-think how we cross-validate

rodluger / flatironhack

Flatiron Institute Everest Hack Week

MIT License

0 stars 0 forks source link

Re-think how we cross-validate #6

Open rodluger opened 7 years ago

rodluger commented 7 years ago

Currently doing a grid search in each PLD order as a shortcut. Should optimize the full 3D function.

rodluger commented 7 years ago

The derivative of the RMS in the validation set with respect to Lambda is analytic. This can be used in a gradient descent / nonlinear optimizer in 3D. We may still have issues with the non-smoothness of the function, but this could save us a lot of time.

img_20171025_110656

rodluger commented 7 years ago

Currently thinking of optimizing full likelihood instead of CV.