Description
1. (30 points) Written Problems
(a) (10 points) Bishop 6.2
(b) (10 points) Bishop 7.3
(c) (10 points) Bishop 7.4
2. (70 points) Programming
Download and save the attached data set, images of handwritten digits: MNISTdata.mat,
prog2. The data represents a subsampling of the full MNIST data set available in
http://yann.lecun.com/exdb/mnist/
1. Please submit both report and Matlab codes
2. Your report must provide an analysis of each methods performance and reasoning
behind the analysis.
3. Compare in your report the relative strengths and weaknesses of the methods based
on the experimental results and your understanding of each algorithm.
(a) (35 points)
Develop code for training and testing an SVM classifier with nonlinear kernel. You
are welcome to use either formulation described in the textbook (chapter 7). You
cannot use an SVM library to complete this assignment. You can use quadratic
programming library if you like. Using your implementation of the SVM classifier,
compare multi-class classification performance of two different voting schemes:
i. one versus the rest
ii. one versus one
Be sure to specify your voting scheme using a method described in the book . To
analyze accuracy, you will find it helpful to produce and analyze the multiclass confusion matrix (http://en.wikipedia.org/wiki/Confusion matrix), in addition to
examining the overall error rate.
4-1
(b) (25 points) Use the same one versus one classifiers from the previous problem in a
DAGSVM approach. A paper describing the approach, DAGSVM.pdf, is attached.
Compare multi- class classification performance with the other two voting schemes.
(c) (c) (10 points) A baseline implementation of the DAGSVM with 6th degree polynomial kernels achieves 95% accuracy on the test set. See if you can do better than this
baseline, using the DAGSVM approach. baseline-CM.pdf contains the confusion
matrix of the baseline implementation:
2
