 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| • |
The weights in
an RBM define p(h|v) and p(v|h) in a very
|
|
straightforward
way (lets ignore biases for
now)
|
|
|
|
– |
To
sample from p(v|h), sample the binary state of
|
|
|
each
visible unit from a logistic that depends on its
|
|
|
weight
vector times h.
|
|
|
|
– |
To
sample from p(h|v), sample the binary state of
|
|
|
each
hidden unit from a logistic that depends on its
|
|
|
weight
vector times v.
|
|
|
| • |
If we use these
two conditional distributions to do
|
|
|
alternating
Gibbs sampling for a long time, we can get
|
|
|
p(v) or p(h)
|
|
|
|
– |
i.e.
we can sample from the model’s distribution over
|
|
|
the
visible or hidden units.
|
|