NEW COURSE

CS/CNS/EE 156 B

Learning Systems
Spring 2000

http://www.vision.caltech.edu/welling/class/LearningSystemsB.html


Instructor: Max Welling     welling@vision.caltech.edu
TA: Silvio Savarese           savarese@vision.caltech.edu
Place and time: 070 Moore, Wednesday/Friday, 9.00-10.30 AM
prerequisites: Ma2 or equivalent
Units: 9 (3-0-6)



Description:

This class will introduce some widely used latent variable models in engineering. Among others we will discuss
PCA, ICA, Factor Analysis, K-means, Mixture of Gaussians, Generative Topographic Mapping, Cluster Weighted Models,  Mixture of Experts, Kalman Filter,  HMM, Helmholtz Machine, Boltzman Machine. It will be shown that most of the learning schemes for these models can be understood as versions of the Expectation Maximization (EM) algorithm, thus providing a unified view.

Students are expected to implement a rudimentary MATLAB version for some of the models and discuss this in class.
The final project will consist of a study of a recent paper and the implementation a more involved model.



Homeworks and Policy:

- 4 homework problem sets (15% each)
- One final project (40%)
- Homeworks will be published on this web page as soon as available.
- No midterm, no final
- Two or more students can collaborate on any homework, but every student is expected to hand in his/her own answers.



Downloads & Links  (last update March 31 - 4pm)

Syllabus  (last update March 31 - 4pm)         - maximum likelihood estimation (  demo1.m ginput2.mplotGauss.m )
        - Baysian learning
        -  complexity measures: MLD.         - EM update rules
        - variational EM
        -  GEM
        - example: learning gaussian parameters from incomplete data         - principal component analysis (PCA)
        - probabilistic PCA
        - factor analysis (FA)
        -  independent component analysis (ICA)         - vector quantization through K-means
        - mixture of Gaussians
        -  mixture of factor analysers         - generative topographic mapping
        - cluster-weighted modelling
        - mixture of experts         - hidden markov models (HMM)
        - Kalman filter         - Laplace approximation
        - variationa approximation
        -  sampling methods: importance sampling, rejection sampling, condensation,
          Markov chain Monte Carlo (MCMC), Gibbs sampling.         - Helmholtz machine
        - Boltzman machine
        -  belief propagation         - leftovers & summary


 Literature:

California Institute of Technology, Pasadena, CA 91125.