next up previous
Next: Results Up: The Bars Problem Previous: The Bars Problem

Learning Procedure

The network was initialised in stages shown in Figure 4. The first layer was added after 20 sweeps and the second after 100 sweeps. Each sweep corresponds to updating each latent variable node once. To prevent falling to local minima, 1) automatic pruning was discouraged initially, 2) new sources were generated and pruned sources removed from time to time and 3) the activations of the sources were reset a few times.



Harri Valpola 2001-10-01