Printable Version of this PageHome PageRecent ChangesSearchSign In

Swaroop Vattam

I am a technical staff member at Lincoln Lab at MIT. Previously, I was a research scientist at Georgia Tech and a NAS fellow at NRL. I got my PhD in Computer Science from Georgia Tech in 2012. My research is broadly focused on applications of machine reasoning and machine learning to natural language processing problems. I am currently investigating data-driven model discovery systems in an effort to tackle the problem of automated machine learning.



<December 2017>
MonTueWedThuFriSatSun
    123
45678910
11121314151617
18192021222324
25262728293031
braindump.png

Brain dump

:7 November 2016

Sublinear time algorithms. Even as recently as early 2000's, we would be happy to settle for a polynomial time algorithm for any non trivial algorithmic problem. It would had been hard to imagine doing better than linear time; after all, we expect algorithms to consider all of its input in order to make a decision. How the times have changed! For the kinds of data sets that I have been dealing with these days, linear is not good enough. The prevalence of extremely large data sets in a wide variety of settings is growing rapidly and beating linear time has become a matter of practical concern as opposed to mere academic curiosity. One solution often touted to this problem is the MapReduce programming model - processing data sets with parallel, distributed algorithms on large clusters. Although this works well in some cases, it's utility has been blown out of proportion in my humble opinion. In many cases, this paradigm is not applicable or does not yield any significant advantage. This is especially true in many big graph data sets.

If linear time is not good enough, it is natural to wonder the existence of sublinear time algorithms for some of these problems. In fact, there has been a lot of recent interest in this direction. Sublinear time is a crazy concept - it allows algorithms to process only a small fraction of the input and yet demands a reasonable solution. There are a handful of problems for which deterministic sublinear time algorithms are known. But in a majority of cases, the algorithm must use randomization and give approximate answers. There is a growing body of work in this area and recent results have shown that there are classical optimization problems whose values can be approximated in sublinear time. Also, property testing has been applied to give sublinear algorithms for a variety of problems. Further, several cool techniques like low rank approximation of matrices have emerged for designing sublinear algorithms. This is all good news! Still, the scope of sublinear algorithms remains minuscule compared to the problems out there with existing solutions considered efficient a decade ago but are no longer acceptable today.

I am grappling with one such graph matching problem presently, namely, maximum weighted matching (MWM) problem. MWM is a classic graph problem for which exact polytime solutions have been known for some time. But for massive graphs this is a non-starter. There is some (relatively) recent work where linear-time approximate algorithms have been worked out for the unweighted case (Drake & Hougardy, 2003; Pettie & Sanders, 2004). I am trying these, but strongly suspect that they may also not fit the bill. Meanwhile, I don't think there are any known sublinear time solutions for this problem. Hope theoreticians are listening!
  • Drake, D. E., & Hougardy, S. (2003). Improved linear time approximation algorithms for weighted matchings. In Approximation, Randomization, and Combinatorial Optimization.. Algorithms and Techniques (pp. 14-23). Springer Berlin Heidelberg.
  • Pettie, S., & Sanders, P. (2004). A simpler linear time 2/3− ε approximation for maximum weight matching. Information Processing Letters, 91(6), 271-276.

:2 November 2016

Results for the 2016 NIST Speaker Recognition Evaluation was announced today. I was part of the MIT team which was placed second in a pool of more than 45 serious research teams competing from around the world. My contribution included developing a backend for calibrating and fusing several different speaker recognition models developed by other researchers on my team. Calibration is a serious issue when there are several different models and we have to combine their individual predictions in order to make an overall prediction. This issue arises because different models have their own quirks - for example, some models tend to predict probabilities conservatively (meaning closer to mid-range), some toward extremes, and others none at all (true conditional probabilities are unknown). If your metric cares about exact probabilities (e.g., logarithmic loss), we need to calibrate the models before fusing their predictions to get an average estimate. This involves a post-processing step where you learn the characteristics of the model behavior. There are two popular methods for calibration: Platt's scaling and Isotonic regression. Platt's scaling amounts to training a logistic regression model on the estimator outputs. In Isotonic regression, the idea is to fit a piecewise-constant monotonically increasing function (e.g., stair shaped function) instead of logistic regression. Alas, I cannot disclose the secret sauce that went into our backend calibrator yet.

All posts ...



rss.jpg

Research feeds

AMS Feature Column: AMS Feature Column - RSS Feed
How to Differentiate with a Computer:
Solar Daze:
The Early History of Calculus Problems, II:
Ranking, Grading, and Genealogy:
Untangling Your Square Dance:
The Joy of Barycentric Subdivision:
Surface Topology in Bach Canons, II: The Torus:
Remembering Bill Thurston (1946-2012):
Patterns in Permutations:
Assembly Required:

Mathematical Moments from the American Mathematical Society: The American Mathematical Societys Mathematical Moments program promotes appreciation and understanding of the role mathematics plays in science, nature, technology, and human culture. Listen to researchers talk about how they use math: from presenting realistic animation to beating cancer.
Generating Patterns Part 2: Researcher: Michel C. Molinkovitch, University of Geneva Description: Michel C. Milinkovitch used math, physics, and biology for an amazing discovery about the patterns on a lizard's skin.
Generating Patterns Part 1: Researcher: Michel C. Molinkovitch, University of Geneva Description: Michel C. Milinkovitch used math, physics, and biology for an amazing discovery about the patterns on a lizard's skin.
Hunting for Planets: Researcher: Konstantin Batygin, Caltech
Description: Konstantin Batygin talks about using math to investigate the existence of Planet Nine.
Designing Better Bicycles: Researcher: Jim Papadopoulos, Northeastern University
Description: Jim Papadopoulos talks about his years of research analyzing bicycles.
Farming Better: Researchers: Eleanor Jenkins, Clemson University and Kathleen (Fowler) Kavanagh, Clarkson University. Lea Jenkins and Katie Kavanagh talk about their work making farming more efficient while using water wisely.

Communications of the ACM: Artificial Intelligence: The latest news, opinion and research in artificial intelligence, from Communications online.
If an AI Doesn't Take Your Job, It Will Design Your Office:

Arranging employees in an office is like creating a 13-dimensional matrix that triangulates human wants, corporate needs, and the cold hard laws of physics: Joe needs to be near Jane but Jane needs natural light, and Jim is sensitive to smells and can't be near the kitchen but also needs to work with the product ideation and customer happiness team—oh, and Jane hates fans.

CSAIL Launches Artificial Intelligence Initiative With Industry:

The SystemsThatLearn@CSAIL industry collaboration aims to use machine learning to create functional human-like systems for data science and other fields.

Machine Learning Predicts the Look of Stem Cells:

No two stem cells are identical, even if they are genetic clones.

Stanford Researchers Create Deep Learning Algorithm That Could Boost Drug Development:

A new type of deep learning, known as one-shot learning, could be used to help drug development because it only requires a small number of data points.

Building an AI Chip Saved Google From Building a Dozen New Data Centers:

Google operates what is surely the largest computer network on Earth, a system that comprises custom-built, warehouse-sized data centers spanning 15 locations in four continents.






Last modified 17 November 2016 at 1:22 pm by svattam