SFU MOCAD Seminar: Andrew Warren
Topic
Unsupervised learning of 1d branching structures
Speakers
Details
Suppose we have unlabeled data where we believe there is an unknown, latent branching (or tree-like) structure. Can we infer that structure? This type of unsupervised learning problem arises in a wide range of biological applications, including in evolutionary and developmental settings.
In this talk, I will present a variational approach to this problem, whereby the latent branching structure can be estimated by way of a discretization of the "average-distance problem" of Buttazzo, Oudet, and Stepanov. The resulting estimator is shown to be consistent in the zero-noise limit, and can be cheaply approximated numerically by a Lloyd- or EM-type algorithm. This work is joint with Anton Afanassiev, Forest Kobayashi, and Geoff Schiebinger.