J-zin/shape-constraints-on-neural-network

Shape Constraints on Nerual Networks

This repo containts the shape constraints methods based on ordinary differential equation, which is first proposed in UMNN (Unconstrainted Monontonic Neural Network). In UMNN, the authors conduct monotonic constraint on neural network by parameterzing the function as

Here, we extend this framework into another shape constraints beyond monotone, i.e., increasing concave and general concave functions.

increasing concave:

general concave:

For more details, please refers to https://j-zin.github.io/files/shape_constraint_NN_slides.pdf. We demonstrate a toy experiment here of the regression task to show its effectiveness on shape constraints. The experimental setting follows that in UMNN and the code logic is based on https://github.com/AWehenkel/UMNN.