Prof. Koppel’s Lab

Prof. Moshe Koppel of the Department of Computer Science conducts research on a variety of machine learning applications including text categorization, image processing, speaker recognition and automated game playing.

He is best known for his contributions to the branch of text categorization concerned with authorship attribution. More recently, he has begun researching fundamental problems in social choice theory.

The most basic and prevalent authorship attribution challenge – when a limited and closed set of candidates is given – is now fairly well understood, and can be solved with high accuracy using conventional machine learning methods.

In addition to making a number of contributions to the study of this basic problem, Koppel and his team have developed profiling techniques that successfully determine an author’s gender, age, native language, and personality type, using nothing but statistical properties of the author’s written work.

Koppel and his team have also made fundamental advances on more difficult authorship problems in which the candidate set might be very large (possibly including as many as tens of thousands of candidates) and/or open (so that there is no guarantee that the true author is in the candidate set).

Such problems pose enormous practical challenges, especially for security forces worldwide. Koppel’s team has developed several highly successful unsupervised learning techniques, including the well-known “unmasking” method, for solving this challenge.

In recent years, Koppel and colleagues have published several papers on social choice theory, offering new formal definitions of a number of basic concepts, including disproportionality and voting power.

In other work in social choice theory, Koppel and colleagues have shown how the “wisdom of crowds” could be optimally exploited. Using a variation of the EM algorithm, Koppel and his team showed that, given a set of voters voting on multiple independent issues, with each voter having fixed but unknown competence, one can determine both the voters’ competence and the true answer for each issue with maximum likelihood.

Recently, Koppel and colleagues have provided a new axiomatization of Raiffa’s solution to Nash’s cooperative bargaining problem.

In addition to their work on text characterization, Koppel’s group has worked on a variety of machine learning applications. These include shadow removal in image processing, enhanced fusion techniques for speaker recognition and genetic algorithms for automated game playing.

Along with Prof. Nathan Netanyahu and Dr. Omid David, both of the Department of Computer Science, Koppel has shown that using nothing but records of games played by grandmasters, a chess program could be trained essentially from scratch to play at grandmaster level.