Basis embeddings should be sparse

[maedoc]

Sparse regression is a clear improvement for interpretability of embedding techniques, so adding the Lasso

ridge regression

or maybe just the elastic net 🐟

(from https://cims.nyu.edu/~cfgranda/pages/OBDA_spring16/material/sparse_regression.pdf) anyway of controlling the number of non zero weights is helpful.

[maedoc]

These penalties could be applied to the MLP flows as well.