On October 1, Partha Niyogi died of brain cancer at the age of 43. He was Louis Block Professor in Computer Science and Statistics at the University of Chicago, and on his web page, he tells us that
My research interests are generally in the field of artificial intelligence and specifically in pattern recognition and machine learning problems that arise in the computational study of human speech and language.
There's a thoughtful and well-written obituary on the University of Chicago's web site, which surveys some of his areas of work and presents some some personal tributes. Some other personal reminiscences can be found in the comments on Lance Fortnow's post at Computational Complexity, for example these:
He had the curiosity and enthusiasm of a child which a lot of us unfortunately lose after our PhD.
He had a special gift in approaching problems out of the box always.
He was a gifted, inspiring professor and kind man.
And this one, which is worth reproducing at greater length:
As one of his advisees, it is difficult to put in words how much and in how many different ways, he advised my thinking and approach to research in general. He was one of those very few people I met who always approached all problems with an open mind and stressed on the importance of simple solutions that could be easily explained. (Un)Surprisingly, he could more often than not, come up with such brilliant solutions. And, boy did he love talking! It was great fun to discuss and argue with him for hours on topics as diverse as cricket, IIT, food, city-life, neuroscience and the like.
Everyone with the good fortune to have known Partha will recognize that portrait.
I'll close with a personal note about one of my recent interactions with him.
Back in May of 2005, I gave a talk at Chicago under the title "The Invisible Academy: Nonlinear effects of linear learning" (slides here). This talk addressed the question of how a speech community might converge on a shared vocabulary, and specifically focused on the question of the development of shared pronunciations. I've given versions of this talk in a few other places, starting with a presentation at Penn in 2000, but I never really knew what to do with it, as it's well outside my normal areas of research.
This work started with an attempt to prove myself wrong. In my undergraduate intro lectures, I'd been asserting for years that the development of community norms in language is a mystery. It finally occurred to me to wonder whether it might not, instead, be an inevitability, given some simple and plausible assumptions about linguistic behavior and language learning. I had some leisure time while dog-sitting for Rich and Sally Thomason in rural Montana, and so I ran a few simple simulations and convinced myself that my hunch was on the right track. But these simulations were neither a proof nor a real-world discovery nor even a model that could be fit to easily-obtained experimental observations. So I presented the results in a few talks, in a frame intended to persuade people that perhaps linguists ought to be more interested in mathematical models of social interaction, and left it at that.
After my talk, Partha and I had one of those wide-ranging and entertaining conversations that were typical of interactions with him. I brought up Estes' attempts to distinguish among prototype theories, examplar theories and feature-frequency theories of category learning; Partha mentioned graph Laplacians and the theory of coalescent processes; and a good time was had by all.
Then, three years later, Partha sent me a draft of a paper: Hariharan Narayanan and Partha Niyogi, "Language Evolution, Coalescent Processes and the Consensus Problem on a Social Network". The opening paragraph:
In recent times, there has been an increased interest in theories of language evolution that have an applicability to the study of dialect formation, linguistic change, creolization, the origin of language, and animal and robot communication systems in general. One particular question that has attracted some interest has the following general form: how might a group of linguistic agents arrive at a shared communication system purely through local patterns of interaction and without any global agency enforcing uniformity? The linguistic agents in question might be humans, animals, or machines in a multi-agent society. For an example of interesting simulations that suggest how a shared vocabulary might emerge in a population , see Liberman (2005) (other simulations are also provided by [27, 8, 2, 1, 28] among others). In this paper, we consider a generalization of Liberman’s model, prove several theoretical properties, and establish connections to related phenomena in biology, social sciences, and physics.
This reminded me of something John McCarthy once said about the experience of interacting with Morris Halle. It's like you're playing football, John said, and you hand the ball off to him, and he jukes through the line, evades a couple of tackles in the secondary, sprints across the goal line, and tosses the ball triumphantly in the air — but it's not a football anymore, it's a bouquet of flowers.
Unfortunately, my attempts to master the mathematics of this particular bouquet coincided with Partha's declining health. I'll miss him.