On-Line Learning: Where are we so Far? - Archive ouverte HAL Accéder directement au contenu
Chapitre D'ouvrage Année : 2010

On-Line Learning: Where are we so Far?

Résumé

After a long period of neglect, on-line learning is re-emerging as an important topic in machine learning. On one hand, this is due to new applications involving data flows, the detection of, or adaption to, changing conditions and long-life learning. On the other hand, it is now apparent that the current statistical theory of learning, based on the independent and stationary distribution assumption, has reached its limits and must be completed or superseded to accunt for sequencing effects, and more generally, for the information carried by the evolution of the data generation process.This chapter first presents the current, still predominant paradigm. It then underlines the deviations to this framework introduced by new on-line learning settings, and the associated challenges that they raise both for devising novel algorithms and for developing a satisfactory new theory of learning. It concludes with a brief description of a new learning concept, called tracking, which may hint as to what could come off as algorithms and theretical questions from looking anew to this allpervading situation: never to stop learning.
Fichier non déposé

Dates et versions

hal-01197557 , version 1 (11-09-2015)

Identifiants

  • HAL Id : hal-01197557 , version 1
  • PRODINRA : 50835

Citer

Antoine Cornuéjols. On-Line Learning: Where are we so Far?. Ubiquitous Knowledge Discovery Challenges, Techniques, Applications, Springer - Verlag, 2010, 978-3-642-16391-3. ⟨hal-01197557⟩
23 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More