Graphical Models and Machine LearningEE 639 Advanced Topics in Signal Processing and CommunicationSpring 2008 |
.
[Home] [Lectures] [Homework] [Related Links]
|
Documents |
Info |
Due |
Solution |
|
|
|||
|
|
2/19/2008 |
|
|
|
3/20/2008 |
|
||
|
4/18/2008 |
|
Final Project (one or two-person team) Timeline:
1. Proposal due 3/6, 5pm EST - email with title and scope
2. Progress report (literature, software, data) due 4/10, 5pm EST – one-page
3. Poster Session on 4/24
4. Final Report due 5/2, 5pm EST – in IEEE Transaction format (http://www.ieee.org/web/publications/authors/transjnl/index.html)
Some suggested topics:
1. Probabilistic representation of visual processing - computational models of visual cortex, and in particular those based on sparse coding, have recently enjoyed much attention. The basic assumption behind sparse coding is that natural scenes are composed of structural primitives, and although there are a potentially large number of these primitives, typically only a few are active in a single natural scene. Various models have been proposed to describe exactly which few are active and to model how our brain makes such selection. This project will be a survey of recent works. Here are a few to get you started.
a. M. Bethge. Factorial coding of natural images: How effective are linear model in removing higher-order dependencies? J. Opt. Soc. Am. A, 23(6): 1253-1268, June 2006.
b. M. Bethge and P. Berens. Near-Maximum Entropy Models for Binary Neural Representations of Natural Images. NIPS 2007
c. P. Berkes, R. Turner and M. Sahani. On Sparsity and Overcompleteness in Image Models. NIPS 2007.
2. Understanding Human Behavior – (From Henry Kautz) Build a system that recognizes human behavior from sensor data using a HMM and/or other probabilistic model. Alternatively, design and implement a route-finding algorithm that automatically learns user preferences, both about specific places (e.g., never drive down a particular street) and general rules (e.g. avoid highways unless using one saves a large amount of time.) For details and dataset, see http://www.cs.rochester.edu/~kautz/Courses/577autumn2007/index.html
3. Voice Conversion using Gaussian Process – Linear models have enjoyed a great deal of success in audio processing from linear prediction to speaker classification. In the voice conversion application, the vocal-tract parameters of a speaker are mapped to those of another speaker. The goal is to transform the original voice to a target voice by combining the original excitation with the target speaker’s vocal tract. The traditional method is to use linear regression on GMM models of line spectrum frequencies features. That is little reason why a linear model should work in the space of LSF. It makes more sense, for example, to use a much richer non-parametric methods such as Gaussian Process Regression. You can get started using the dataset and free software of voice conversion from http://www.tc-star.org. Theoretical knowledge and software on Gaussian Process Regression can be obtained from the excellent free book by Rasmussen and Williams: http://www.gaussianprocess.org/gpml/ .
4. Motion segmentation based on pixel linkage -- Use NetKit (http://www.research.rutgers.edu/~sofmac/NetKit-desc.html#BIN) to study the video segmentation problem in which pixels are linked based on the similarity of different features: color, edges, motion and so forth.
5. Privacy protected belief propagation – to protect privacy of individual nodes or cliques, the local potential function and/or evidence may be encrypted. The goal of this project is to develop a belief propagation / junction tree routine that allows encrypted potential functions to be used. This toolkit for example can be used as a privacy-protected probabilistic search engine viewing the evidence as private query. Homomorphic encryption can be used for this project.
a. C. Orlandi, A. Piva, and M. Barni. Oblivious neural network Computing via Homomorphic Encryption. EURASIP Journal on Information Security. Volume 2007, Article ID 37342, 2007.
6. Random sparse graphs in network coding – The use of belief propagation on a spare random graph as a view to achieve maximum likelihood error correction has obtained enormous success in recent years. Network coding is a new theory for multicast and broadcast showing that better communication bound is feasible by setting partial data to different branch of the network and to allow algebraic operations to be done in the network. Combining to the two to tackle realistic noisy network is a very interesting direction. This project is aimed at simulating some of the latest results in this area. Relevant literature includes:
a. A. Montanari and R. Urbank. Coding for Network Coding
b. R. Kotter and F. R. Kschischang. Coding for error and erasures in random network coding.
c. R. Young et al. Tutorial on Network Coding Theory - http://iest2.ie.cuhk.edu.hk/~whyeung/publications/tutorial.pdf
d. D. MacKay. Information Theory, Inference, and Learning Algorithm – online book at http://www.cs.toronto.edu/~mackay/itprnn/book.html