Olga Veksler The goal of the final project is to apply one or several techniques we learned to real world data. There are two approaches that you can take. First you may select the data you wish to work with and then to choose one or several methods which you think will work well with that data. Or conversely, you may want to work with a certain method and will have to think about the data which will work well with this method.

By November 1 you should hand in a one paragraph proposal. Proposal should include the data you will work with (including where you got that data), and the method(s) you will apply to this data. In the course of working with data, if may switch to a different method.

Examples of Final Projects:

(a) Face detection: Write an algorithm which will recognize if there is a face in an image, and output an approximate location of face(s). Training data should contain face images and non face images. First you develop a classifer which given a rectangular patch of pixels classifies this patch as a face or non-face, then apply this classifier to all patches in the input image. Example of methods to try is PCA, Fischer Discriminant, Nearest Neighbors, etc.

(b) Face Recognition: Similar to face detection only now there are several classes, each one representing a particular person. Classifier should label an input patch with the label of the person, if the input patch comes from some person in the database, or, if the input patch comes from unknown person or non person, label that patch "reject". Apply the classifier to every patch in the input image. Example of methods to try is PCA, Fischer Discriminant, Nearest Neighbors, etc.

Good places to get face data:

(c) Character recognition: The same as (a) and (b) only the training data comes from handwritten characters. Example of methods to try is PCA, Fischer Discriminant, Nearest Neighbors, etc.

Good places to get character recognition data: MNIST database

Good places to get various types of data:

UCI Machine Learning Repository The students will also write up a project report, 2 to 5 pages. The report will NOT be judged for its length. Rather it must include the 3 essential components. First you must state the problem you will try to solve, then the approach you are going to take to solve it, and last your results and what you have learned from your results. The results can be negative (that is my approach did not work so well), if they are negative, you should try to explain possible reasons for failure. For people interested in face recognition/detection, here are links to some important papers in the field. These papers are intended mostly for the graduate students who can afford to spend more time on the project.

Data for the "competition" project You can split the data into training/testing sets as you wish, or you can use cross-validation. I will test your classifier on the data which I'm not distributing.