Figure 7
From: Using Memristors for Robust Local Learning of Hardware Restricted Boltzmann Machines
(a) Test error rate achieved by the Discriminative RBM as a function of cycle-to-cycle variability for every combination of the pulse width programming scheme (Cst, RProp) and number of parallel Gibbs chains used to evaluate Contrastive Divergence (# CD). (b) Optimal conductance increment-to-noise ratio as a function of cycle-to-cycle variability associated with each curve of the left panel. When using 20 Gibbs chains (blue curves), from σintra/(Gmax − Gmin) = 6.10−3 onwards (vertical gray dashed line) the conductance update overcomes the noise increase, accounting for the improved performance compared to the use of a single Gibbs chain (orange curves), regardless of the programming scheme. Each simulation was ran over 30 epochs with a mini-batch size of 100, 5 times per value of pulse width, error bars indicate median, first quartile and third quartile.
