Topic: In this Master thesis we address the problem of how to robustly train a ConvNet for regression, or deep robust regression [1,2]. Traditionally, deep regression employs the L2 loss function , known to be sensitive to outliers, i.e. samples that either lie at an abnormal distance away from the majority of the training samples, or that correspond to wrongly annotated targets. This means that, during back-propagation, outliers may bias the training process due to the high magnitude of their gradient. In this Master thesis, we will investigate the use of the heavy-tailed Student-T distribution in combination with deep architectures. We expect the candidate be happy with both mathematical derivations and deep architecture implementation and training. Most probably, the hybrid probabilistic-deep model will be optimised with an EM-like algorithm combined with stochastic gradient descent. Various data sets for training and evaluation are available for tasks such as age estimation, full body pose estimation, depth estimation or landmark detection, but other suggestions are more than welcome.
Environment: This project will be carried out in the Perception Team, at Inria Grenoble Rhône-Alpes in collaboration with the Multimedia Team at Telecom Paris. The research progress will be closely supervised by Dr. Xavier Alameda-Pineda, Dr. Stéphane Lathuilière, and Dr. Radu Horaud, head of the Perception Team. At the perception team we have the necessary computational resources (GPU & CPU) to carry on the proposed research.
 Lathuilière S, Mesejo P, Alameda-Pineda X, Horaud R. DeepGUM: Learning Deep Robust Regression with a Gaussian-Uniform Mixture Model. ECCV (2018).
 Belagiannis, V., Rupprecht, C., Carneiro, G., Navab, N. Robust optimization for deep regression. ICCV (2015)
 Lathuilière, S., Mesejo, P., Alameda-Pineda, X., Horaud, R. A comprehensive analysis of
deep regression. IEEE TPAMI (2019).