|
|
|
| A three stage training procedure |
||
| (Taylor, Hinton and Roweis) |
||
|
|
|
|
|
|
|
|
|
|
| • | First learn a
static model of pairs or triples of |
||||||||
| time frames
ignoring the directed temporal |
|||||||||
| connections
between hidden units. |
|||||||||
| • | Then use the
inferred hidden states to train a |
||||||||
| “fully observed”
sigmoid belief net that captures |
|||||||||
| the temporal
structure of the hidden states. |
|||||||||
| • | Finally, use the
conditional RBM model to fine |
||||||||
| tune all of the
weights. |
|||||||||