Sparse Probabilistic Projections
C´dric Archambeau, University College London (CSML) e Francis Bach, Ecole Normale Sup´rieure (INRIA ­ Willow project) e Why generative models for probabilistic projections (e.g. PPCA)? Direct access to the leading directions Suitable for constructing mixture models Extension to robust and/or Bayesian versions

W47

Aims of this work: General generative framework: (multiview, etc.) xp = Wp y0 + Vp yp + µp + p, y  N (0, -1 )

Induce sparsity by suitable choice of prior on the weight matrices Improve interpretability and/or use as feature extraction tool Preserve advantages of original approaches