Store shapes, corrupt one, and watch the network repair it
1. Click shapes to store them (W = (1/d) Σ x xᵀ, d = 256 neurons)
2. start from
noise: pixels flipped
↻ New corruption
▶ Recall
One sweep
or click / drag on the state to flip pixels yourself
State
Energy E(s) = −½ sᵀWs, after every neuron update
Overlap with each stored shape, m = x · s / d
How many patterns fit? Random patterns against the 0.138·d limit
try it: store N random patterns in d = 256 neurons
Update all neurons at once, or one at a time?
Synchronous: all 256 at once
Asynchronous: one at a time
▶ Play
↻ Another example
synchronous
asynchronous (per sweep of 256 updates)
neurons that changed in the last synchronous step