Energy-based learning, visible
Train a full Boltzmann machine or an RBM and inspect stochastic states, weights, energy, and reconstruction error.
Network state
idleInference and reconstruction
Inference pauses training and ranks reconstructed visible states for the current clamped input without updating weights.
Hidden posterior P(H | input)
Visible reconstruction
No inference result.
Most likely reconstructed states
input-conditioned probabilityNo reconstructed states yet.
Learning functions
These functions execute in this tab. Invalid code is rejected without replacing the active functions.
Original functions active.
Reading the experiment
RBM: positive associations come from data and hidden probabilities; negative associations come from a CD-k reconstruction.
Full BM: the positive phase clamps visible data while the negative phase lets every unit run freely. More Gibbs steps generally reduce sampling bias but cost more time.
Temperature: lower values make states more deterministic. Edge width shows magnitude; green and red show weight sign.
Model behavior follows Ackley, Hinton, and Sejnowski (1985), with RBM contrastive-divergence controls based on Hinton's practical guide. Source links are in the README.