A sweep of JSAE training with jacobian penalty from 1e-3 to 1e-2, trained for 20k steps. (added zero sparsity for baseline)