The simulation visualizes a small RBM's bipartite visible-hidden layer as an interactive energy landscape, showing how weight updates from contrastive divergence reshape the energy surface so that valleys align with training patterns while alternating Gibbs sampling steps hop between visible and hidden states.
Choose a training pattern set, press play to run alternating Gibbs sampling steps between the visible and hidden layers, and watch the energy landscape lower around the data as contrastive divergence updates the weights; use the CD-k slider to add more Gibbs steps per update and the rebuild control to reset with fresh random weights.
Training pattern select, CD-k Gibbs steps slider, play/pause, rebuild
Geoffrey Hinton's 2002 contrastive divergence paper and the subsequent 2006 Deep Belief Network work, which stacked RBMs to pretrain deep networks layer by layer, are widely credited with helping reignite mainstream interest in deep learning after decades of relative neglect.
The simulation visualizes a small RBM's bipartite visible-hidden layer as an interactive energy landscape, showing how weight updates from contrastive divergence reshape the energy surface so that valleys align with training patterns while alternating Gibbs sampling steps hop between visible and hidden states.
The simulation visualizes a small RBM's bipartite visible-hidden layer as an interactive energy landscape, showing how weight updates from contrastive divergence reshape the energy surface so that valleys align with training patterns while alternating Gibbs sampling steps hop between visible and hidden states.
Choose a training pattern set, press play to run alternating Gibbs sampling steps between the visible and hidden layers, and watch the energy landscape lower around the data as contrastive divergence updates the weights; use the CD-k slider to add more Gibbs steps per update and the rebuild control to reset with fresh random weights.
Geoffrey Hinton's 2002 contrastive divergence paper and the subsequent 2006 Deep Belief Network work, which stacked RBMs to pretrain deep networks layer by layer, are widely credited with helping reignite mainstream interest in deep learning after decades of relative neglect.