APPLIED MATH SEMINAR

Speaker: Prof. Lawrence Carin, Department of Electrical and Computer
Engineering, Duke university

Title: Nonparametric Statistical Embeddings and Dynamic Motion Modeling

When/where: Tuesday, February 23rd, 4:15 PM, AKW 200

Abstract: Nonparametric Bayesian (NPB) methods are employed to learn a
statistical embedding, in which high-dimensional data are embedded
in a low-dimensional latent space. The model also allows synthesis of
high-dimensional data, from vectors in the latent space. The proposed method
integrates the Dirichlet process with a spike-slab prior, yielding nonlinear
factor analysis (FA), with the dimensionality of the latent space inferred from
the data. A novel one-step method is proposed for aligning the FA latent
features. Further, NPB methods are developed to learn a nonlinear
dynamic model in the latent space, based on motion-capture data. Example results
are presented for statistical embedding, synthesis of high-dimensional data
based on latent features, and motion-capture analysis and synthesis.