 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| • |
Look inside the
generator to see how it works (Williams et. al)
|
|
|
|
– |
Too
tedious. Not possible for the real motor system.
|
|
|
| • |
Define
a prior over codes and generate
lots of (code, image) pairs.
|
|
Then train a
function approximator that does image code.
|
|
|
|
– |
What
about ambiguous images? The average code is bad.
|
|
|
| • |
Define
a prior over codes and train a
neural net to model code
|
|
|
image.
Then backpropagate image
residuals to iteratively find a
|
|
|
locally optimal
code for each test image.
|
|
|
|
– |
May
be better than using a fixed linear mapping from image
|
|
|
residuals
to code corrections (Cootes et al)
|
|
|
| • |
We need
the prior over codes to avoid learning to invert the
|
|
|
generator
in irrelevant parts of image space.
|
|