In Memory of Partha Niyogi
I first met Partha during his job interview with the Computer Science department and was immediately struck by his thoughtfulness, this quiet, soft-spoken, understated in-depth knowledge in so many fields. Not your typical flashy MIT AI character. I was very excited about recruiting him and delighted when he agreed to come to the CS department. Several months after his arrival it became clear to me that offering him a position in the Statistics Department, as well, would help create important bridges between Machine Learning--his field of research and more traditional Statistics. In the Statistics Department we all came to appreciate his special qualities, his deep insights and measured judgments, especially regarding hiring and promotion.
In our building, Eckhart Hall, the office doors are closed and locked for security reasons. In Statistics we have a curious system whereby full time faculty all carry a master key, thus saving each other the effort to get up and open the door. So when someone knocks you know it isn't faculty and its most likely a student. But students knock timidly. So whenever I heard a loud double rap--I knew it was Partha. He'd walk right in with his question of the day...
Perhaps the greatest thing about Partha was his ability to ask questions, and what is more important in scientific inquiry than coming up with good questions? Partha questioned everything, including his own work. He posed hard pointed questions. He really wanted to get to the root of things. For example in his early work in Speech Recognition he made use of Support Vector Machines, a classification method that became popular during the 90's that consisted of a modification of penalized regression with a data term more suitable for classification. But I remember Partha repeatedly wondering if this modification was really needed. Did it really make a difference if one used a quadratic loss or a hinge loss? And in much of his subsequent beautiful work on manifold learning he consistently compared the two, and I suspect his skepticism was justified. Partha also questioned the dominant HMM paradigm in speech recognition, he questioned whether we really need language level models to disambiguate the acoustic signal, or have we all simply been using the wrong processing of this signal. Partha's thinking on problems in AI was always motivated by observations on human perception. He would often dwell on how much background noise and clutter there is in a typical environment in which we listen to speech, and how sensitive existing speech recognition algorithms are to such noise. He would emphasize how adaptable and invariant human recognition is compared to machine algorithms. More than once he told me that he himself had never talked on the phone before his teenage years. That in no way affected his ability to understand speech on the phone once he did pick up the receiver. These conversations eventually led us to a joint project aimed at developing a more robust speech recognition algorithm. And true to his nature, Partha always questioned whether we had really achieved any progress?
So when the double rap came I knew I was in for some tough questions, and, beyond that, a very long conversation covering a myriad of topics. Because, you see, the question of robustness of human perception, the question of how much high level knowledge is required to understand speech, is closely connected to the questions Chomsky asked with respect to language. And as someone who had studied language and linguistics, and had done his Ph.D. at MIT, Partha took Chomsky's theories very seriously. We spent quite a few meetings with Partha introducing me to Chomsky's theory of a universal grammar. And after the first conversation touching on Chomsky we discovered that we both shared a great deal of respect for him as a political and social thinker. So you can imagine a conversation that started with a pointed question on say mixture models for the speech signal, and evolved into a higher level conversation on robust speech recognition, could very well end up with a discussion on Chomsky's positions on the Middle East conflict or US foreign policy. Now keep in mind that whereas Partha was more of an afternoon person, I am an early morning person, and that double rap at 3PM meant that I may not get home until quite late. But I really couldn't resist. I couldn't resist the richness and breadth of the conversations Partha triggered and engaged me in. He was a real renaissance person, both in his scientific interests, contributing in his research to a very broad range of topics, all somehow connected to artificial or human intelligence. and beyond that in his cultural and social interests, from classical Indian music, to Indian history, to economic theory, and on and on. Partha's passing is a great loss for his family, for his friends, for his colleagues, for his students, for the Computer Science and Statistics Departments, and for the University as a whole. But his inquisitive spirit will stay on with us, the questions he asked will continue to intrigue us, and the methods he developed will surely help us in our attempts to find some answers.