Purely Optional/Extra-Credit Assignment

*No groups*– individual assignments ony. Due Dec. 12 by noon EST.

Take the existing HW4 data-set (or alternately a labeled ladar data-set you prefer) you have used for classification and explore two things with it:

1 ) Implement exponentiated gradient descent or an L1 regularized method (or both simultaneously)  on a loss function you have already implemented. (Log loss, hinge loss, squared loss etc…)

Are the exponential gradient algorithms good performers? Take the current feature set and:

- Add a large number of random features

- Add a large number of features that are noisy, corrupted versions of the features already in the

data-set.

How do the various methods perform in these situations? Compare and contrast with l2 methods.

2) Implement a technique for “contextual classification”. Possible options:

a) Implement the graph cut method in http://www.ri.cmu.edu/publication_view.html?pub_id=6297

b) Implement the multiple k-means clustering of data-points/voting scheme pioneered in http://www.cs.uiuc.edu/homes/dhoiem/publications/Hoiem_Geometric.pdf  and use features generated from that. (For Discussed of ladar points see  http://www.ri.cmu.edu/publication_view.html?pub_id=6297)

c) Put a continuous valued random field using, e.g., the l1 Total Variation norm between data-points. Using optimization to find the optimal assignment for each ladar point.  Feel free to do 2 classes only here.

d) Propose some other method to use multiple related/nearby labels to improve the structured prediction.

Can you get improvements on this data-set? Why or why not?