Skip to Main content Skip to Navigation

Coping with the Computational and Statistical Bipolar Nature of Machine Learning

Abstract : Machine Learning is known to have its roots in a broad spectrum of fields including Artificial Intelligence, Pattern Recognition, Statistics or Optimisation. From the earliest stages of Machine Learning, both computational issues and generalisation properties have been identified as central to the field. While the former address the question of computability, complexity (from a fundamental perspective) or computational efficiency (on a more practical standpoint) of learning systems, the latter aim at understanding and characterising how well the solutions they provide perform on new, unseen data. Those last years, the emergence of large-scale datasets in Machine Learning has been deeply reshaping the principles of Learning Theory. Taking into account possible constraints on the training time, one has to deal with more complex trade-offs than the ones classically addressed by Statistics. As a direct consequence, designing new efficient algorithms (both in theory and practice), able to handle large-scale datasets, imposes to jointly deal with the statistical and computational aspects of Learning. The present thesis aims at unravelling, analysing and exploiting some of the connections that naturally exist between the statistical and computational aspects of Learning. More precisely, in a first part, we extend the stability analysis, which relates some algorithmic properties to the generalisation abilities of learning algorithms, to a novel (and fine-grain) performance measure, namely the confusion matrix. In a second part, we present a novel approach to learn a kernel-based regression function, that serves the learning task at hand and exploits the structure of the problem so that the optimisation procedure is made inexpensive. Finally, we investigate the trade-off between convergence rate and computational cost when minimising a composite functional with inexact proximal-gradient methods. In that setting, we identify optimisation strategies that provably are computationally optimal.
Document type :
Complete list of metadata
Contributor : Pierre Machart <>
Submitted on : Wednesday, January 9, 2013 - 11:37:31 AM
Last modification on : Monday, March 30, 2020 - 8:50:46 AM


  • HAL Id : tel-00771718, version 1


Pierre Machart. Coping with the Computational and Statistical Bipolar Nature of Machine Learning. Machine Learning [cs.LG]. Aix-Marseille Université, 2012. English. ⟨tel-00771718⟩



Record views


Files downloads