Contrastive divergence training.

(A) Schematic of the training procedure depicting how the native state is stabilized relative to other states upon parameter updates. (B)-(E) In all plots, the blue curves indicate larger initial step-size training and the green plots indicate smaller step-size (fine-tuning). (B) The upper left plot shows the decline in minibatch-averaged RMSD over the course of the optimization. The remaining plots show (C) the convergence of the hydrogen bonding and side chain-side chain interaction parameters over the optimization for (D) Met-Met and (E) Val-Val potential. The larger step-size optimization of the side chain parameters exhibits large oscillations that inhibit convergence.