On the inductive biases of deep domain adaptation - Archive ouverte HAL Accéder directement au contenu
Pré-Publication, Document De Travail Année : 2021

On the inductive biases of deep domain adaptation

Rodrigue Siry
  • Fonction : Auteur
  • PersonId : 1105723
Louis Hémadou
  • Fonction : Auteur
  • PersonId : 1105724
Loïc Simon
Frédéric Jurie

Résumé

Domain alignment is currently the most prevalent solution to unsupervised domainadaptation tasks and are often being presented as minimizers of some theoretical upper-bounds on risk in the target domain. However, further works revealed severe inadequacies between theory and practice: we consolidate this analysis and confirm that imposing domain invariance on features is neither necessary nor sufficient to obtain low target risk. We instead argue that successful deep domain adaptation rely largely on hidden inductive biases found in the common practice, such as model pre-training or design of encoder architecture. We perform various ablation experiments on popular benchmarks and our own synthetic transfers to illustrate their role in prototypical situations. To conclude our analysis, we propose to metalearn parametric inductive biases to solve specific transfers and show their superior performance over handcrafted heuristics.
Fichier principal
Vignette du fichier
neurips_2021.pdf (221.26 Ko) Télécharger le fichier

Dates et versions

hal-03290701 , version 1 (19-07-2021)
hal-03290701 , version 2 (25-04-2023)

Identifiants

  • HAL Id : hal-03290701 , version 1

Citer

Rodrigue Siry, Louis Hémadou, Loïc Simon, Frédéric Jurie. On the inductive biases of deep domain adaptation. 2021. ⟨hal-03290701v1⟩
73 Consultations
211 Téléchargements

Partager

Gmail Facebook X LinkedIn More