I'm a Ph.D student in the Department of Computer Science, Courant Institute of Mathematical Sciences at New York University, with Prof. Yann LeCun.

I'm interested in the area of Machine Learning and Pattern Recognition. My research focuses on trying to answer interesting questions by exploiting the rich contextual information associated with most data sets in a variety of real world domains. To this end, I have focused on the design and application of efficient learning and inference algorithms for models that are capable of handling the uncertainties and interdependencies among samples associated with large scale data sets. I have worked on models of very diverse nature, such as Energy Based Models, Graphical Models, Deep Learning Architectures, and Relational Graphical Models. Furthermore, these models have been applied to real world problems in a variety of domains, such as in economics for predicting the prices of real estate properties, in computer vision for object recognition, face verification, distance metric learning, and data visualization and in robot navigation.

Rich contextual information Importance of contextual information has been well acknowledged in the fields of human and computer vision. Indeed, when a human observes a scene, or studies a photograph, he usually has at his disposal a wealth of information that is not captured by the image alone. Or, while detecting a car in an image it is beneficial to use the knowledge that a car is usually parked on the road and sky is always above the ground. Contextual information is equally essential for reasoning in other domains as well. For example, the box office revenue of a movie, while relying on its own unique features, will also depend on factors such as what other movies are currently running in theaters, and how did other movies by the same actor/director combination fared in the past. Similarly, in predicting the price of a house, the spatial context (influence of neighboring houses) and the temporal context (inflation) play important roles. In these scenarios, the data samples interact with one another in highly complex ways, while often exhibiting an underlying logical structure. I aim to capture this notion of context by modeling these interdependencies, to enable us to robustly reason for a diverse set of complex real world phenomenon.

Interesting questions Using the contextual information mentioned above, I am more interested in solving problems which go beyond the simple "yes/no" questions (straight forward classification), and which go beyond just predicting the value of a function using only the individual features of an input sample (straight forward regression). For example, in addition to knowing whether a car is present in an image, wouldn't it be interesting to infer all the other objects and their respective positions? Similarly, instead of just predicting the price of a house, it would be far more useful to understand how the economic properties of the entire neighborhood will change over time.