|
Abstract:
For tasks such as speech and face recognition, a rich source of prior
knowledge about the domain may come in the form of a generative black box,
such as a speech synthesizer, or a graphics program that generates facial
images. We consider the problem of learning the inverse of such a
generative mapping from data. For example, given a set of faces and a
graphics program, train a neural network to infer from a face the graphics
inputs that would accurately reconstruct it. The problem is difficult
because we typically have only a small number of labelled training cases,
and the generative mapping is a black box in the sense that there is no
analytic expression for its gradient. This results in a nonlinear inverse
problem where the forward function is given and can be evaluated as many
times as we like.
We describe a way of training a network that starts with a small amount of
labelled training data and uses the generative black box to produce more
training data. As learning proceeds, the training set evolves and the
labels that the network assigns to unlabelled training data converge to
their correct values. We demonstrate our approach by learning to invert a
2D morphable model for faces.