Research Projects
Active Evaluation for Large Scale Multi-class Problems (Ongoing B.Tech Thesis)
In this project, we are developing methods for actively assessing the performance of the classifier by helping an analyst explore regions of the data with very few labeled points and then seeking labels for these regions.
Eventually, the user should be able to see the accuracy of the data stratified across different regions and thereby get a more consistent view of the performance of the classifier.
Work submitted at ACM SIGKDD 2012 Abstract
Sparse Paths in Multiple Kernel Learning (Ongoing R&D Project)
Current methods in Multiple Kernel Learning employ various sparsity inducing norms for combining several expert kernels/feature-sets in the context of large-margin learning models.
However the question of which norm to select for the particular problem at hand is unanswered. In this work, we attempt to answer this important question.
Ranking Comments on Yahoo! News (Summer Internship, May-July 2011)
Dr.Srinivasan Sengamedu, Dr.Subhajit Sanyal (Yahoo! Research, Bengaluru, India)
Objective: Characterize, with a set of features, the subjective notion of a comment being “engaging” for the reader. Design a ranking function that promotes “engaging” comments but demotes racist & abusive content.
Designed an Abusive content classifier that was tested to have 79% cross-validation accuracy, on small manually built corpus. Accuracy improved with larger corpus of hate & non-hate comments
Came up with features that could be correlated with (our characterization of) “good” & “bad” comments. Also identified properties the ranking function must possess based on above features
Our work, with improvements, has been implemented and is supposed to be released on the Yahoo! News page for bucket testing in December. Based on my contributions, I was recommended for a Research Engineer position at Yahoo! Research, despite being an undergraduate.
Performance Measurement & Regression Framework for DTM Engine (Summer Internship, May-July 2010)
Dinesh Rathi, Development Manager (Informatica Labs, Bengaluru, India)
Objective: Provide profiler-like functionality to Mercury, Informatica's commercial Data Integration software, by generating various relevant statistics to aid performance analysis & optimization
Evaluated metrics like CPU time, Memory usage, CPU usage etc. of the DTM Engine of Mercury
Designed a generic framework to measure performance of DB operations like join, union etc
Psychometrics (Winter Internship, December 2009)
Mr. Udayan Kanade, CEO, Oneirix Labs, Pune
Worked on quantifying “abilities” of students and “difficulties” of questions given the response-set of students (of unknown abilities) to a subset of a pool of questions (of unknown difficulties) administered in a Computer Adaptive Test
The modelling was done using Item Response Theory (A 3-parameter Rasch Model was used)
Studied the Expectation Maximization algorithm.
Echo Canceller (Summer Internship, May-July 2009)
Mr. Udayan Kanade, CEO, Oneirix Labs, Pune.
Implemented an adaptive-filter based Echo Canceller for real-time audio transmission over the internet at the end of my freshman year
As a part of this intern, attended lectures on Advanced Digital Signal Processing and Statistical And Optimized Signal Processing by Mr. Udayan.
Also since audio transmission was to be done over the network, learnt concepts in Computer Networks and Sound Programming with ALSA, all done in C
Others
Proximity Tracker (Yahoo! HackU 2010 hack)
Designed a J2ME based mobile application that reminds the user of a task based on proximity of the user to the task-completion location as input by him
Selected for an internship at Yahoo! based on the hack & my performance during the event
Pothole-Filling Robot (Embedded Systems Lab Project, Spring 2011)
Prof. Kavi Arya, IIT Bombay
Built an ATmega-2560 microcontroller based four-wheeled robot that automatically detects & fills potholes in a road
Modular design that had an independent sharp distance sensor based detection mechanism & a stepper motor based hopper mechanism to fill the detected potholes
Course Registration Interface (Database Systems Course Project, Autumn 2010)
Prof. S. Sudarshan, IIT Bombay
Huffman-Encoding based File Compression (Program Abstractions and Paradigms Course Project, Spring 2009)
Prof. Amitabh Sanyal, CSE Dept., IIT Bombay
Cows and Bulls (Introductory Programming Course Project, Autumn 2008)
Prof. Abhiram Ranade, CSE Dept., IIT Bombay
|