This week we will have a tea talk by Pascal Vincent. Hope to see you all there.
When: Thursday May 10th, 15h00 Where: LISA Lab.
Title: Biasing estimation towards Manifold prior hypothesis
Abstract: I will present the fruits of my reflections on the manifold hypothesis, starting form the mere geometric intuition, and showing how it may be translated into specific regularization schemes when fitting energy based models or auto-encoders. I will contrast the resulting regularized parameter estimation schemes with what is implied by maximum likelihood, score matching, denoising auto-encoder and contractive auto-encoder criteria. The alternative regularization terms I propose should aim at modeling actual low dimensional manifolds, irrespective of model parametrization and contrary to our current approaches.
Cheers, Aaron