Jaxed NMF NMF optimized through stochastic gradient descent with the option of adding l2 or l1 penalities on weight matrices. NMF minimizes: L(W, H) = ||X-WH||_2 subject to W>0, H>0