What my research is about


I am a PhD student in Machine Learning/Object Recognition, in the Computer Vision and Robotics Group at Cambridge University. We are part of the Engineering Department's Machine Intelligence Lab. My supervisors are Prof Christopher Bishop and Prof Roberto Cipolla, and I am financed by the Microsoft Research European PhD Scholarship.

Probabilistic models in Object Recognition are classified into two categories: generative models, that try to describe how images as we see them are generated, by optimising the joint likelihood of the data and their labels p(x,c). And dicriminative models, that try to come up with the best boundary possible between different classes, by optimising the posterior distribution of the labels p(c|x).

The two approaches have different advantages. Generative models are powerful because they model what classes look like. The model of each class can be trained independently of the other classes, whereas discriminative models have to be retrained every time a new class is added. Plus, because they try to model the distribution of the data, generative models can handle missing data. On the other hand, dicriminative models only focus on the boundary between classes, on what is ambiguous, which makes them more efficient and better at classifying.

My main focus is on building hybrid models that combine generative and dicriminative approaches.