I, along with prof. S. Dasgupta, have developed a first local search algorithm for heirarchical clustering that works for generalised family of distances known as Bregman distances. Bregman distance include euclidean distances as well as many other infomation-theoretic distances such as KL-distance. I am currently working on the paper for this heirarchical clustering algorithm that will be submitted to ICML 2005. The algorithm starts with randomly intialized clustering and iteratively reduces a chosen cost of the heirarchical clustering. As opposed to populat method such as Ward's method, our local search algorithm finds a local mimima in the entire space of the heirarchical clustering. Bregman distances allows us to cluster data more effectively that may have more complex distribution, since each Bregman distance corresponds to an exponential family distribution.

For empirical results, we have clustered english spoken phonemes using different metrics. The KL-divergence clustering of the phonemes clusters the phoneme based on its probability distributions given the phoneme that preceded or followed it, whereas the normal euclidean clustering is based on the distance between 39 dimensional feature vector for each phoneme.

My masters thesis will involve research that will extend some common machine learning algorithms to these information-theoretic distances.

Classifying EEG signals using Neural Networks and Independent Component Analysis (ICA).

My earlier project in the machine learning was geared towards supervised learning methods. The project involved classifying various sleep stages of a human by analyzing the brain waves i.e. EEG traces. I used neural networks, different clustering methods and pre-processing tools like ICA and PCA etc.

I am also working with Prof. Joel Watson in economics department doing research in game theory and mechanism design. I am interested in exploring real world situations using unsupervised learning methods and game theoretic principles. One of the problems I am currently working on is modeling constraints for different mechanisms for a repeated and enforced game with partial verifiability.

I have taught a robotics class in a UCSD summer program for high school students. The class taught introduction to computer programming using Legos mindstorm robotic kits. Please visit the class webpage for more details.

• Robotics 2003
• Robotics 2004
• Robotics 2005 (Coming soon)

I have also served as TA for the following undergraduate classes at UCSD

• CSE 150 - Introduction to Artificial Intelligence (1 quarter)
• CSE 140(L) - Digital Design (3 quarters)
• CSE 141(L) - Computer Architecture (3 quarters)
• CSE 20 - Discreet Mathematics (1 quarter)
• CSE 21 - Mathematics for Algorithm and Systems (1 quarter)
• Undergraduate tutor for math, economics and statistics courses (4 years)

I worked with Prof. Rik Belew in developing an introductory programming class for non-computer science majors. I also took part in a teaching computation seminar that discussed various pedagogical techniques for teaching programming and basic of computation. It was partly influenced by recent ACM’s computing curricula and Larry Snyder’s text book ‘Fluency.’ Here is a webpage that I created for the seminar.