The inversion task
	• We are given: – A generator that converts codes into images. – A large dataset of images. • We want a fast way of converting images into the codes that would generate them. – If we can do this inversion, we can use “analysis by synthesis” which is obviously the right way to do recognition (e.g. figure out how a digit was drawn) • We are not given: – The codes for the images in the training set. – A prior over codes.