| An early application (Sutskever) |
|
|
|
|
|
|
|
|
|
|
| • | We first tried
CRBM’s for modeling images of |
||||||||
| two balls
bouncing inside a box. |
|||||||||
| • | There are 400
logistic pixels. The net is not told |
||||||||
| about objects or
coordinates. |
|||||||||
| • | It has to learn
perceptual physics. |
||||||||
| • | It works better
if we add “lateral” connections |
||||||||
| between the
visible units. This does not mess |
|||||||||
| up Contrastive
Divergence learning. |
|||||||||