Projects

Locally Linear Embedding: unsupervised learning of nonlinear data manifolds
Products of Experts: modeling distributions using renormalized products of simpler learned distributions
Helmholtz machines: Unsupervised learning using bottom-up recognition models
Learning in Bayesian Networks: Graphical models relating random variables
Expectation Conjugate-Gradient: Improving the Speed of EM for Learning Latent Variable Models
Multiple-Cause Vector Quantization: Learning parts-based models of data.
Combining Discriminative Features To Infer Complex Trajectories: a conditional model for time-series regression.

Ensemble learning: Fitting weight distributions without Monte Carlo
Bayesian inference: Making predictions using all likely networks, not just one
Monte Carlo methods: Solving hard Bayesian inference problems stochastically

Video Processing: using Bayesian Networks to learn the structure of video sequences
Phase Unwrapping: using variational inference for 2D signal processing
Elastic models: Using deformable models to recognize hand-written digits
Glove-Talk: A neural network that converts gestures into real-time speech