Skip to content

MinSR algorithm

Vicentini Filippo requested to merge github/fork/chrisrothUT/NeuralTangentKernel into master

Created by: chrisrothUT

This is a draft of the MinSR algorithm relegated to netket.experimental for the time being.

The MinSR algorithm involves contracting the Jacobian with itself over the parameter dimension to compute the neural tangent kernel (NTK). There is a memory cost N(Y + P^l) where N is the number of samples, Y is the size of the total hidden state across all layers, and P^l is the parameters per layer. This can be reduced by computing the NTK matrix in chunks using the argument jacobian_chunk_size.

Merge request reports