The Goal: Learning Latent Trees

I Let x = (x1,..., xD)T. Model p(x) with the aid of latent variables I Latent class model (LCM) has a single latent variable I Latent tree (or hierarchical latent class, HLC) model has a tree structure, with visible variables as leaves I Tree-structured network allows linear time inference I Inspiration from parse-trees I Zhang (2004), Zhang and Kočka (2004) search over HLCs, starting from LCM: O(D3) I Our method: greedy, bottom-up determination of a tree/forest O(D2) I Aim is to be fast, cf Bayesian MCMC over tree structures I Note: Chow-Liu trees do not contain latent variables 2 / 19 aeroplane bicycle bird boat bottle bus car cat chair cow diningtable dog horse motorbike person pottedplant sheep sofa train tvmonito