Fig. 1: Maximal conditional probability during training.
From: Mode-assisted unsupervised learning of restricted Boltzmann machines
We plot, in red, the maximal conditional probability of the hidden layer, r(v+), where v+ is the mode of the marginal distribution, p(v), as a function of contrastive divergence training iterations (k = 1). The results are averaged over an ensemble of 200 randomly initialized 15 × 10 restricted Boltzmann machines trained on a shifting bar dataset, with ±1σ defining the shaded regions, where σ is the standard deviation. The same calculation conditioned on a random visible configuration, vr, is plot as a baseline for comparison in yellow.
