Section: Scientific Foundations

Machine learning

Machine learning for computer vision.

A large portion of research in computer vision involves increasingly more refined machine learning techniques. Significant success has been obtained by the direct use of off-the-shelf techniques, such as kernel methods (support vector machines for example) and probabilistic graphical models. However, in order to achieve the level of performance that we aim for, a more careful integration of machine learning and computer vision algorithmic and theoretical frameworks is needed. A major part of our machine learning effort is dedicated to this integration, through: (a) applying the transductive learning framework to exploit the simultaneous availability of training and test data in semi-interactive segmentation and image retrieval tasks, (b) using specialized unsupervised matrix factorization algorithms for image representation and image denoising, and (c) developing efficient approximate inference algorithms for graphical models with geometric constraints, allowing a more faithful probabilistic model for scene analysis.

Algorithms and Learning theory.

We aim at providing a better understanding of the fundamental ideas underlying efficient learning algorithms. To understand well popular methods is often a key step in order to refine and generalize these methods, and also to design new learning algorithms. Apart from the computational complexity mentioned before, the common features encountered when using learning techniques in computer vision are (i) high dimensionality and (ii) complexity of the modelization. To avoid the curse of dimensionality, we intend to search for sparse representations of the prediction function. Sparsity inducing norms are raising increased interest in the statistics and learning theory communities; regularizing learning problems using such norms leads to both sparse predictors and good generalization performances. We are currently exploring structured sparse methods, where the idea is to introduce some prior knowledge into a sparse inference problem, for computational reasons or to improve interpretability and predictive performance. Moreover, all these methods lead to interesting practical hyperparameter selection, which can be tackled by theoretically grounded data-driven calibration procedures.