Supervised learning for computer vision: theory and algorithms
The objective of the tutorial is to give a machine learning perspective to the supervised learning algorithms that are often used in various areas of computer vision. Putting most algorithms into a single framework allows to compare the advantages and disadvantages of different learning techniques, and to choose the most appropriate one. The course will be divided in two parts: a theoretical part where relevant results from statistical machine learning theory will be presented, in particular concerning nearest-neighbor algorithms, boosting algorithms and support vector machines. In a second part, the tutorial will focus on more practical aspects of supervised machine learning, in particular, the minimization of convex functionals (e.g., support vector machines and logistic regression), their links with kernel methods and sparsity-inducing norms such as the ℓ1-norm. Throughout the tutorial, examples of successful applications of these supervised learning techniques to computer vision will be presented.