Week 1: Statistical learning
February 1 – 5, 2016
This week will be devoted to statistical learning from both the theoretical and applied perspectives. Statistical learning theory has been developed in the 70’s and brought a great revival in statistics. On the one hand, the development of computer science and tools allowed massive data collection, and implementation of powerful algorithms which are often memory and computation time consuming. On the other hand, classical asymptotic theory used to prove the efficiency of estimation methods in modeling and prediction was limited by dimensionality problems.
The approaches developed in statistical learning helped to face some new challenges such as the curse of dimensionality, small sample size, and now massive datasets. Data mining, feature selection, and more recently Big data appeared as specialized approaches for massive datasets modeling and analysis.
In addition to the theoretical developments (non asymptotic theory), many algorithms emerged in statistics and computer science to meet these new needs which arised in many areas such as bioinformatics, social sciences, medicine or telecommunications.
This week aims to bring together specialists from statistical learning working on advanced techniques and coming from different fields, mainly statistics, but also computer science, social science and bioinformatics.
The main topics of interest include:
Large-scale machine learning and convex optimization
Apprentissage et données massives
Is adaptive early stopping possible in statistical inverse problems?
Entropy, geometry, and a CLT for Wishart matrices
The power of heterogeneous large-scale data for high-dimensional causal inference
Mixed integer programming for sparse and non convex machine learning
A Lagrangian viewpoint on Robust PCA
Block-diagonal covariance selection for high-dimensional Gaussian
Statistical learning with Hawkes processes and new matrix concentration inequalities
Random forests variable importances: towards a better understanding and large-scale feature selection
About the Goldenshluger-Lepski methodology for bandwidth selection
Sub-Gaussian mean estimators
Reconstruction simpliciale de variétés via l’estimation des plans tangents
Quantization, Learning and Games with OPAC
Understanding (or not) Deep Neural Networks
Eigenvalue-free risk bounds for PCA projectors
Estimation of local independence graphs via Hawkes processes and link with the functional neuronal connectivity
A novel multi resolution framework for the statistical analysis of ranking data
Robust sequential learning with applications to the forecasting of electricity consumption and of exchange rates
Oracle inequalities for network models and sparse graphon estimation