Skip to Main content Skip to Navigation
Theses

Towards Adaptive Learning with Invariant Representations

Abstract : Although learning from data (Machine Learning) has dramatically improved Artificial Intelligence systems, these algorithms are not infallible ; they are sensitive to data shift, a ubiquitous situation in the industry. The Adaptation of machine learning models has been the subject of fruitful research, with an influential line of study that learns Invariant Representations, i.e. insensitive to changes in data. In this thesis, we show that learning invariant representations exposes to the risk of destroying their adaptability, a quantity that we, unfortunately, cannot control. We propose a theoretical analysis introducing a new error term, called hypothesis class reduction error, which captures the adaptability of a representation. Secondly, this thesis unifies two research fields for Adaptation, Importance Sampling and Invariant Representations, under the same theoretical framework. In particular, we show the need for inductive bias for adaptive learning, putting human expertise back at the centre of Machine Learning. Finally, we question a fundamental assumption when learning invariant representation ; the access to a large sample of unlabeled data of the new distribution. Indeed, this assumption is rarely met in practice, where we would ideally like to adapt with a few examples. This thesis contributes to this new problem by formalizing it and providing the community with a codebase for a reproducible search. Moreover, we offer a solid baseline based on Optimal Transport for this task.
Document type :
Theses
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03663398
Contributor : ABES STAR :  Contact
Submitted on : Tuesday, May 10, 2022 - 10:35:22 AM
Last modification on : Wednesday, May 11, 2022 - 4:41:08 AM

File

2021UPAST141_BOUVIER_archivage...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03663398, version 1

Citation

Victor Bouvier. Towards Adaptive Learning with Invariant Representations. Machine Learning [cs.LG]. Université Paris-Saclay, 2021. English. ⟨NNT : 2021UPAST141⟩. ⟨tel-03663398⟩

Share

Metrics

Record views

8

Files downloads

0