The term project is a significant component of the term's effort, and is really your opportunity to learn one particular portion of the course in greater depth. The range of possible project topics is extremely broad. This year I am proposing some specific project ideas. Each student will can one of the following areas for their project, or propose something else, related to the course and of interest to them or relevant to their research:

- Any study of the current state-of-the-art in pattern recognition
(classification, extrapolation, clustering, estimation)
- A study or implementation of recent work in classifier combining,
voting schemes etc.
- Classifier Boosting - Implementation or Recent developments.

A search on "boosting" will yield many references. Duda, Hart, & Stork has some material to get you started. - Quad-tree and other domain-decomposition or tree-based methods for
classification or clustering
- Support vector machines, a survey or discussion of the state of the
art in learning / training methods.
- Application of bootstrapping methods to some area of pattern rec

(Bootstrapping involves effective use of small data sets)

Ref: Introduction to the bootstrap, QA276.8.E3745

Ref: Statistical Data Analysis ..., Science, Vol 253, p. 390, 1991

Ref: The UW library has 17 entries under Subject: Bootstrap (Statistics) - A survey of robust statistics

(R-S involves the rejection of outlying data points)

Ref: The UW library has 41 entries under Subject: Robust Statistics

Ref: Robust statistics, QA276.R618 - Overview paper on Dempster-Shafer theory

(D-S is a non-Bayesian theory of probability, based on "beliefs")

Ref: A mathematical theory of evidence, QA273.S48

Ref: A mathematical theory of hints : ... the Dempster-Shafer theory of evidence, QA273.K597 - Numerical application of Dempster-Shafer theory

(i.e., implement a technique for D-S non-Bayesian data fusion)

Refs: See previous topic. - Classification using Logistic Regression

(L-R is an alternative to linear regression for cardinal variables)

Ref: Here is an excellent WWW reference. - A survey of current methods in unlabeled clustering

(e.g., what are current alternatives to K-means or hierarchical clustering?)

Ref: SD372 Course Notes, Chapter 7 - Higher order statistics

(i.e., non-Gaussian statistics; corresponding estimators are non-linear) - Compare fuzzy logic with probability theory

(Most axioms of F-L can be written probabilistically; what is the real difference?) - Blake's "Condensation" algorithm

(a highly effective real-time multiple-hypothesis method for computer vision)

Ref: Contour tracking by stochastic propagation ..., European Conference on Computer Vision, V. 1, p. 343

Ref: http://www.robots.ox.ac.uk/~misard/condensation.html - Perceptual / Multidimensional scaling

(how to infer the feature extraction of the human brain)

Ref: The UW library has 46 entries under Subject: Multidimensional Scaling

Ref: The UW library has 682 entries under Subject: Psychometrics.

Ref: Introduction to multidimensional scaling ..., BF39.S33

- You select a few journal papers in a particular research area and write a tutorial / summary paper. You paper should describe the work that has been done, perhaps assess or critique what you have found, and discuss similarities or differences between different researcher's approaches.
- Find an algorithm which has been proposed or described in the literature, implement it, and then test it. Discuss your observations from these tests; e.g., under what sort of circumstances is the algorithm applicable and how well does it perform?
- Choose a pattern recognition problem of modest difficulty and develop your own algorithm, based on the material discussed in class or based on ideas from journal papers.

- IEEE Trans. on Pattern Analysis and Machine Intelligence
- IEEE Trans. on Image Processing
- IEEE Trans. on Geoscience and Remote Sensing
- IEEE Trans. on Medical Imaging
- Pattern Recognition Journal

- Automated unlabeled clustering algorithms
- Density estimation from very few data points
- Pattern recognition in very high-dimensional spaces
- Computer simulation of error bounds (the error bounds in hypothesis testing or in parameter estimation)
- Context sensitive grammars and computational complexity
- Quantum computing / quantum state machines
- Do a quantitative assessment (training time, accuracy etc.) of neural networks compared with ``exact'' classical methods.
- Non-Bayesian classification
- Non-Gaussian statistical models

- Image segmentation
- Optical character recognition / handwriting recognition
- Security / speaker identification
- Speech recognition
- English language understanding and synthesis
- Face or gesture recognition
- Land-use classification and mapping
- Automated inspection / analysis methods in manufacturing

- Analysis of High School grades for starting UW students
- Comparison of Fuzzy and Bayesian theories
- Handwritten Music Recognition
- Fuzzy c-Means Clustering
- Parallel classifiers
- Real-time object tracking
- Approaches to tracking formant frequencies of human speech
- Texture classification using co-occurrence matrices
- Stock market prediction
- 3D surface classification and estimation using structured lighting
- Artificial neural-network speech recognition
- Analysis of single neurons as eigenfilters
- Fast training algorithms for multiplayer perceptrons
- Detecting word-pair correlations in a Chinese language corpus
- Eigenfaces for face recognition
- Human visual pattern recognition
- Stochastic model-based image segmentation
- Estimation of motor-unit firing statistics
- Stock selection system
- Bootstrapping for low-data problems
- Inference of covariances from small data sets

Home | Teaching | Research | Students | Software | Papers |

(Page last updated September 13, 2016, [an error occurred while processing this directive])