dschonholtz / kan-optimization Goto Github PK

A huge portion of what makes KANs valuable is their interpretability.
As part of the effort to make the KAN's efficient this was effectively ripped out.

It looks like, when reading @Blealtan 's notes, and code, that it should be possible to keep the interpretability components associated with the network while reducing the training time.

I am a bit tired, and details will have to come later. If you are interested. Ping me, and we can chat

Currently, the interface for the FFTKan, convKan and efficientKan models all have very different methods of being instantiated, and the data must be reshaped for each of them.

Ideally, each is a class that has an init function that can be called, and a method on the class or something similar that will reshape the data so we can switch between any of the models without tweaking small parts of the pipeline for every run.