Gauge conditions on nonlinear/interaction module

[wsdewitt]

How should we deal with gauge fixing when we have a pre-latent nonlinear module? We still want WT seq at zero in latent space, and that seems easy enough: we can just have no bias, and mask the WT entries on input. But we also want this interaction module to have zero response to single mutants (since we have our usual beta coefs for that). I’m less sure how best to handle that.

[wsdewitt]

LassoNet may be appropriate for this https://arxiv.org/abs/1907.12207