User profiles for J. Von Kügelgen
Julius von KügelgenETH Zürich Verified email at tuebingen.mpg.de Cited by 1264 |
Self-supervised learning with data augmentations provably isolates content from style
Self-supervised representation learning has shown remarkable success in a number of
domains. A common practice is to perform data augmentation via hand-crafted transformations …
domains. A common practice is to perform data augmentation via hand-crafted transformations …
Nonparametric identifiability of causal representations from unknown interventions
J von Kügelgen, M Besserve… - Advances in …, 2024 - proceedings.neurips.cc
We study causal representation learning, the task of inferring latent causal variables and
their causal relations from high-dimensional functions (“mixtures”) of the variables. Prior work …
their causal relations from high-dimensional functions (“mixtures”) of the variables. Prior work …
Algorithmic recourse under imperfect causal knowledge: a probabilistic approach
AH Karimi, J Von Kügelgen… - Advances in neural …, 2020 - proceedings.neurips.cc
Recent work has discussed the limitations of counterfactual explanations to recommend actions
for algorithmic recourse, and argued for the need of taking causal relationships between …
for algorithmic recourse, and argued for the need of taking causal relationships between …
Provably learning object-centric representations
…, B Schölkopf, J Von Kügelgen… - International …, 2023 - proceedings.mlr.press
… Without further constraints on f, the pixel subsets Ik(z) and Ij(z) can overlap for any k ̸= j
such that latent slots k, j can affect the same pixels and thus contribute to generating the same …
such that latent slots k, j can affect the same pixels and thus contribute to generating the same …
Probable domain generalization via quantile risk minimization
Abstract Domain generalization (DG) seeks predictors which perform well on unseen test
distributions by leveraging data drawn from multiple related training distributions or domains. …
distributions by leveraging data drawn from multiple related training distributions or domains. …
Causal component analysis
… Further, we say that Pj is a perfect intervention if the dependence of the j-th variable from its
parents is removed (pak(j) = ∅), corresponding to deleting all arrows pointing to i, sometimes …
parents is removed (pak(j) = ∅), corresponding to deleting all arrows pointing to i, sometimes …
Independent mechanism analysis, a new concept?
Independent component analysis provides a principled framework for unsupervised
representation learning, with solid theory on the identifiability of the latent code that generated the …
representation learning, with solid theory on the identifiability of the latent code that generated the …
Causal discovery in heterogeneous environments under the sparse mechanism shift hypothesis
R Perry, J Von Kügelgen… - Advances in Neural …, 2022 - proceedings.neurips.cc
… Xj for j 6= i. As discussed by Huang et al. [28], since E can be a common cause of variables
in X, causal … Each variable j in environment e has a randomly sampled mechanism …
in X, causal … Each variable j in environment e has a randomly sampled mechanism …
Visual representation learning does not generalize strongly within the same domain
An important component for generalization in machine learning is to uncover underlying latent
factors of variation as well as the mechanism through which each factor acts in the world. …
factors of variation as well as the mechanism through which each factor acts in the world. …
Spuriosity didn't kill the classifier: Using invariant predictions to harness spurious features
…, J von Kügelgen… - Advances in …, 2024 - proceedings.neurips.cc
To avoid failures on out-of-distribution data, recent works have sought to extract features that
have an invariant or stable relationship with the label across domains, discarding" spurious…
have an invariant or stable relationship with the label across domains, discarding" spurious…