Figure 6.

Generation mode of the event-driven DBN. (A) Screen capture of the network while generating samples of input activations corresponding to class “4.” The neuron corresponding to label “4” was stimulated in the Label Layer (left), and activity propagated through the whole network. The snapshot shows a single example of activity in the Visual Input Layer (right) that is sampled from the generative model encoded in the weights of the DBN. Through short-term depressing synapses (see Figure 3) the network starts to drift through the space of all potential compatible input activations. (B) Raster plot of the DBN in Generation mode. The Label Layer (bottom) is initially stimulated for 2 s (shaded region) to fire in sequence for digits 1, 2, and 3. Afterwards, the network freely samples from the encoded generative model. Although activity in the Label Layer jumps between digits, activity in the Visual Input Layer transitions smoothly.