On the Computational Cost of Stochastic Security

We investigate whether long-run persistent chain Monte Carlo simulation of Langevin dynamics improves the quality of the representations achieved by energy-based models (EBM). We consider a scheme wherein Monte Carlo simulation of a diffusion process using a trained EBM is used to improve the adversarial robustness and the calibration score of an independent classifier network. Our results show that increasing the computational budget of Gibbs sampling in persistent contrastive divergence improves the calibration and adversarial robustness of the model, elucidating the practical merit of realizing new quantum and classical hardware and software for efficient Gibbs sampling from continuous energy potentials.