---
OA_place: publisher
OA_type: gold
PlanS_conform: '1'
_id: '20934'
abstract:
- lang: eng
text: ' Supervised learning for causal discovery from observational data often achieves
competitive performance despite seemingly avoiding the explicit assumptions that
traditional methods require for identifiability. In this work, we analyze CSIvA
(Ke et al., 2023) on bivariate causal models, a transformer architecture for amortized
inference promising to train on synthetic data and transfer to real ones. First,
we bridge the gap with identifiability theory, showing that the training distribution
implicitly defines a prior on the causal model of the test observations: consistent
with classical approaches, good performance is achieved when we have a good prior
on the test data, and the underlying model is identifiable. Second, we find that
CSIvA can not generalize to classes of causal models unseen during training: to
overcome this limitation, we theoretically and empirically analyze \textit{when}
training CSIvA on datasets generated by multiple identifiable causal models with
different structural assumptions improves its generalization at test time. Overall,
we find that amortized causal discovery still adheres to identifiability theory,
violating the previous hypothesis from Lopez-Paz et al. (2015) that supervised
learning methods could overcome its restrictions.'
alternative_title:
- TMLR
article_processing_charge: No
article_type: original
arxiv: 1
author:
- first_name: Francesco
full_name: Montagna, Francesco
id: 353afc8e-19f4-11f0-9db9-811f1723c83f
last_name: Montagna
- first_name: Maximilian T
full_name: Cairney-Leeming, Maximilian T
id: 2214a80c-31f8-11ee-a48d-cf52cc58759b
last_name: Cairney-Leeming
- first_name: Dhanya
full_name: Sridhar, Dhanya
last_name: Sridhar
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: Montagna F, Cairney-Leeming MT, Sridhar D, Locatello F. Demystifying amortized
causal discovery with transformers. Transactions on Machine Learning Research.
2025.
apa: Montagna, F., Cairney-Leeming, M. T., Sridhar, D., & Locatello, F. (2025).
Demystifying amortized causal discovery with transformers. Transactions on
Machine Learning Research. ML Research Press.
chicago: Montagna, Francesco, Maximilian T Cairney-Leeming, Dhanya Sridhar, and
Francesco Locatello. “Demystifying Amortized Causal Discovery with Transformers.”
Transactions on Machine Learning Research. ML Research Press, 2025.
ieee: F. Montagna, M. T. Cairney-Leeming, D. Sridhar, and F. Locatello, “Demystifying
amortized causal discovery with transformers,” Transactions on Machine Learning
Research. ML Research Press, 2025.
ista: Montagna F, Cairney-Leeming MT, Sridhar D, Locatello F. 2025. Demystifying
amortized causal discovery with transformers. Transactions on Machine Learning
Research.
mla: Montagna, Francesco, et al. “Demystifying Amortized Causal Discovery with Transformers.”
Transactions on Machine Learning Research, ML Research Press, 2025.
short: F. Montagna, M.T. Cairney-Leeming, D. Sridhar, F. Locatello, Transactions
on Machine Learning Research (2025).
corr_author: '1'
date_created: 2026-01-04T23:01:35Z
date_published: 2025-12-18T00:00:00Z
date_updated: 2026-01-05T09:54:59Z
day: '18'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2405.16924'
file:
- access_level: open_access
checksum: 968c471bb1f682cf823b2d4cadea8a3f
content_type: application/pdf
creator: dernst
date_created: 2026-01-05T09:51:28Z
date_updated: 2026-01-05T09:51:28Z
file_id: '20939'
file_name: 2025_PMLR_Montagna.pdf
file_size: 1030280
relation: main_file
success: 1
file_date_updated: 2026-01-05T09:51:28Z
has_accepted_license: '1'
language:
- iso: eng
month: '12'
oa: 1
oa_version: Published Version
publication: Transactions on Machine Learning Research
publication_identifier:
eissn:
- 2835-8856
publication_status: published
publisher: ML Research Press
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/francescomontagna/learning-to-induce.git
scopus_import: '1'
status: public
title: Demystifying amortized causal discovery with transformers
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: journal_article
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2025'
...
---
OA_place: publisher
OA_type: gold
_id: '21049'
abstract:
- lang: eng
text: 'Post-hoc importance attribution methods are a popular tool for “explaining”
Deep Neural Networks (DNNs) and are inherently based on the assumption that the
explanations can be applied independently of how the models were trained. Contrarily,
in this work we bring forward empirical evidence that challenges this very notion.
Surprisingly, we discover a strong dependency on and demonstrate that the training
details of a pre-trained model’s classification layer (<10% of model parameters)
play a crucial role, much more than the pre-training scheme itself. This is of
high practical relevance: (1) as techniques for pre-training models are becoming
increasingly diverse, understanding the interplay between these techniques and
attribution methods is critical; (2) it sheds light on an important yet overlooked
assumption of post-hoc attribution methods which can drastically impact model
explanations and how they are interpreted eventually. With this finding we also
present simple yet effective adjustments to the classification layers, that can
significantly enhance the quality of model explanations. We validate our findings
across several visual pre-training frameworks (fully-supervised, self-supervised,
contrastive vision-language training) and analyse how they impact explanations
for a wide range of attribution methods on a diverse set of evaluation metrics.'
acknowledgement: "We sincerely thank Sukrut Rao and Yue Fan for their valuable feedback
on the paper and insightful discussions throughout the project. Additionally, we
appreciate Sukrut’s help\r\nwith some LATEX sorcery. This work was partially supported
by ELSA Mobility Program1\r\nas part of the ELLIS2 exchange program to the Institute
of Science and Technology Austria (ISTA), where a portion of this research was conducted."
article_processing_charge: No
arxiv: 1
author:
- first_name: Siddhartha
full_name: Gairola, Siddhartha
id: fb21489d-057c-11f1-b1b6-d68cd6ae64f5
last_name: Gairola
- first_name: Moritz
full_name: Böhle, Moritz
last_name: Böhle
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Bernt
full_name: Schiele, Bernt
last_name: Schiele
citation:
ama: 'Gairola S, Böhle M, Locatello F, Schiele B. How to probe: Simple yet effective
techniques for improving post-hoc explanations. In: 13th International Conference
on Learning Representations. ICLR; 2025.'
apa: 'Gairola, S., Böhle, M., Locatello, F., & Schiele, B. (2025). How to probe:
Simple yet effective techniques for improving post-hoc explanations. In 13th
International Conference on Learning Representations. Singapore: ICLR.'
chicago: 'Gairola, Siddhartha, Moritz Böhle, Francesco Locatello, and Bernt Schiele.
“How to Probe: Simple yet Effective Techniques for Improving Post-Hoc Explanations.”
In 13th International Conference on Learning Representations. ICLR, 2025.'
ieee: 'S. Gairola, M. Böhle, F. Locatello, and B. Schiele, “How to probe: Simple
yet effective techniques for improving post-hoc explanations,” in 13th International
Conference on Learning Representations, Singapore, 2025.'
ista: 'Gairola S, Böhle M, Locatello F, Schiele B. 2025. How to probe: Simple yet
effective techniques for improving post-hoc explanations. 13th International Conference
on Learning Representations. ICLR: International Conference on Learning Representations.'
mla: 'Gairola, Siddhartha, et al. “How to Probe: Simple yet Effective Techniques
for Improving Post-Hoc Explanations.” 13th International Conference on Learning
Representations, ICLR, 2025.'
short: S. Gairola, M. Böhle, F. Locatello, B. Schiele, in:, 13th International Conference
on Learning Representations, ICLR, 2025.
conference:
end_date: 2025-04-28
location: Singapore
name: 'ICLR: International Conference on Learning Representations'
start_date: 2025-04-24
corr_author: '1'
date_created: 2026-01-27T12:48:35Z
date_published: 2025-01-22T00:00:00Z
date_updated: 2026-02-09T06:11:17Z
day: '22'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2503.00641'
file:
- access_level: open_access
checksum: 6c8dfe4291c41d5a2c2fd838105e10b9
content_type: application/pdf
creator: dernst
date_created: 2026-02-09T06:06:14Z
date_updated: 2026-02-09T06:06:14Z
file_id: '21162'
file_name: 2025_ICLR_Gairola.pdf
file_size: 24386863
relation: main_file
success: 1
file_date_updated: 2026-02-09T06:06:14Z
has_accepted_license: '1'
language:
- iso: eng
month: '01'
oa: 1
oa_version: Published Version
publication: 13th International Conference on Learning Representations
publication_status: published
publisher: ICLR
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/sidgairo18/how-to-probe
status: public
title: 'How to probe: Simple yet effective techniques for improving post-hoc explanations'
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2025'
...
---
OA_place: publisher
OA_type: gold
_id: '21066'
abstract:
- lang: eng
text: "Causal discovery from observational data holds great promise, but existing
methods rely on strong assumptions about the underlying causal structure, often
requiring full observability of all relevant variables. We tackle these challenges
by leveraging the score function ∇logp(X)\r\n of observed variables for causal
discovery and propose the following contributions. First, we generalize the existing
results of identifiability with the score to additive noise models with minimal
requirements on the causal mechanisms. Second, we establish conditions for inferring
causal relations from the score even in the presence of hidden variables; this
result is two-faced: we demonstrate the score’s potential as an alternative to
conditional independence tests to infer the equivalence class of causal graphs
with hidden variables, and we provide the necessary conditions for identifying
direct causes in latent variable models. Building on these insights, we propose
a flexible algorithm for causal discovery across linear, nonlinear, and latent
variable models, which we empirically validate."
acknowledgement: "Philipp M. Faller was supported by a doctoral scholarship of the
Studienstiftung des deutschen\r\nVolkes (German Academic Scholarship Foundation).
This work has been supported by AFOSR,\r\ngrant n. FA8655-20-1-7035. FM is supported
by Programma Operativo Nazionale ricerca e innovazione 2014-2020. We thank Atalanti
A. Mastakouri, Kun Zhang and Haoyue Dai for the insightful discussions."
alternative_title:
- PMLR
article_processing_charge: No
arxiv: 1
author:
- first_name: Francesco
full_name: Montagna, Francesco
last_name: Montagna
- first_name: Philipp
full_name: Faller, Philipp
last_name: Faller
- first_name: Patrik
full_name: Blöbaum, Patrik
last_name: Blöbaum
- first_name: Elke
full_name: Kirschbaum, Elke
last_name: Kirschbaum
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Montagna F, Faller P, Blöbaum P, Kirschbaum E, Locatello F. Score matching
through the roof: Linear, nonlinear, and latent variables causal discovery. In:
Proceedings of the Fourth Conference on Causal Learning and Reasoning.
Vol 275. ML Research Press; 2025:552-605.'
apa: 'Montagna, F., Faller, P., Blöbaum, P., Kirschbaum, E., & Locatello, F.
(2025). Score matching through the roof: Linear, nonlinear, and latent variables
causal discovery. In Proceedings of the Fourth Conference on Causal Learning
and Reasoning (Vol. 275, pp. 552–605). Lausanne, Switzerland: ML Research
Press.'
chicago: 'Montagna, Francesco, Philipp Faller, Patrik Blöbaum, Elke Kirschbaum,
and Francesco Locatello. “Score Matching through the Roof: Linear, Nonlinear,
and Latent Variables Causal Discovery.” In Proceedings of the Fourth Conference
on Causal Learning and Reasoning, 275:552–605. ML Research Press, 2025.'
ieee: 'F. Montagna, P. Faller, P. Blöbaum, E. Kirschbaum, and F. Locatello, “Score
matching through the roof: Linear, nonlinear, and latent variables causal discovery,”
in Proceedings of the Fourth Conference on Causal Learning and Reasoning,
Lausanne, Switzerland, 2025, vol. 275, pp. 552–605.'
ista: 'Montagna F, Faller P, Blöbaum P, Kirschbaum E, Locatello F. 2025. Score matching
through the roof: Linear, nonlinear, and latent variables causal discovery. Proceedings
of the Fourth Conference on Causal Learning and Reasoning. CLeaR: Conference on
Causal Learning and Reasoning, PMLR, vol. 275, 552–605.'
mla: 'Montagna, Francesco, et al. “Score Matching through the Roof: Linear, Nonlinear,
and Latent Variables Causal Discovery.” Proceedings of the Fourth Conference
on Causal Learning and Reasoning, vol. 275, ML Research Press, 2025, pp. 552–605.'
short: F. Montagna, P. Faller, P. Blöbaum, E. Kirschbaum, F. Locatello, in:, Proceedings
of the Fourth Conference on Causal Learning and Reasoning, ML Research Press,
2025, pp. 552–605.
conference:
end_date: 2025-05-09
location: Lausanne, Switzerland
name: 'CLeaR: Conference on Causal Learning and Reasoning'
start_date: 2025-05-07
corr_author: '1'
date_created: 2026-01-29T14:19:09Z
date_published: 2025-05-01T00:00:00Z
date_updated: 2026-02-10T11:54:02Z
day: '01'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2407.18755'
file:
- access_level: open_access
checksum: f2bc44b2320667d4049b3518b1f2fe5d
content_type: application/pdf
creator: flocatel
date_created: 2026-01-29T14:17:48Z
date_updated: 2026-01-29T14:17:48Z
file_id: '21067'
file_name: montagna25a.pdf
file_size: 1739334
relation: main_file
success: 1
file_date_updated: 2026-01-29T14:17:48Z
has_accepted_license: '1'
intvolume: ' 275'
language:
- iso: eng
main_file_link:
- open_access: '1'
url: https://proceedings.mlr.press/v275/montagna25a.html
month: '05'
oa: 1
oa_version: Published Version
page: 552-605
publication: Proceedings of the Fourth Conference on Causal Learning and Reasoning
publication_identifier:
eissn:
- 2640-3498
publication_status: published
publisher: ML Research Press
quality_controlled: '1'
status: public
title: 'Score matching through the roof: Linear, nonlinear, and latent variables causal
discovery'
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 275
year: '2025'
...
---
OA_place: repository
OA_type: green
_id: '21068'
abstract:
- lang: eng
text: "Causal reasoning and discovery, two fundamental tasks of causal analysis,\r\noften
face challenges in applications due to the complexity, noisiness, and highdimensionality
of real-world data. Despite recent progress in identifying latent\r\ncausal structures
using causal representation learning (CRL), what makes learned\r\nrepresentations
useful for causal downstream tasks and how to evaluate them are\r\nstill not well
understood. In this paper, we reinterpret CRL using a measurement\r\nmodel framework,
where the learned representations are viewed as proxy measurements of the latent
causal variables. Our approach clarifies the conditions under\r\nwhich learned
representations support downstream causal reasoning and provides\r\na principled
basis for quantitatively assessing the quality of representations using\r\na new
Test-based Measurement EXclusivity (T-MEX) score. We validate T-MEX\r\nacross
diverse causal inference scenarios, including numerical simulations and\r\nreal-world
ecological video analysis, demonstrating that the proposed framework\r\nand corresponding
score effectively assess the identification of learned representations and their
usefulness for causal downstream tasks. Reproducible code can\r\nbe found at https://github.com/shimenghuang/a-measurement-perspective-of-crl."
acknowledgement: "This research was funded in whole or in part by the Austrian Science
Fund (FWF) 10.55776/COE12. For open access purposes, the author has applied a CC
BY public copyright license to any accepted manuscript version arising from this
submission.\r\n"
alternative_title:
- Advances in Neural Information Processing Systems
article_processing_charge: No
arxiv: 1
author:
- first_name: Dingling
full_name: Yao, Dingling
id: d3e02e50-48a8-11ee-8f62-c108061797fa
last_name: Yao
- first_name: Shimeng
full_name: Huang, Shimeng
id: 989c2a06-fb4e-11ef-a992-ab766442255b
last_name: Huang
orcid: 0000-0001-6919-821X
- first_name: Riccardo
full_name: Cadei, Riccardo
id: 0fa8b76f-72f0-11ef-b75a-a5da96e5ad6b
last_name: Cadei
- first_name: Kun
full_name: Zhang, Kun
last_name: Zhang
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Yao D, Huang S, Cadei R, Zhang K, Locatello F. The third pillar of causal
analysis? A measurement perspective on causal representations. In: 39th Annual
Conference on Neural Information Processing Systems. Vol 38. Neural Information
Processing Systems Foundation; 2025.'
apa: 'Yao, D., Huang, S., Cadei, R., Zhang, K., & Locatello, F. (2025). The
third pillar of causal analysis? A measurement perspective on causal representations.
In 39th Annual Conference on Neural Information Processing Systems (Vol.
38). San Diego, CA, United States: Neural Information Processing Systems Foundation.'
chicago: Yao, Dingling, Shimeng Huang, Riccardo Cadei, Kun Zhang, and Francesco
Locatello. “The Third Pillar of Causal Analysis? A Measurement Perspective on
Causal Representations.” In 39th Annual Conference on Neural Information Processing
Systems, Vol. 38. Neural Information Processing Systems Foundation, 2025.
ieee: D. Yao, S. Huang, R. Cadei, K. Zhang, and F. Locatello, “The third pillar
of causal analysis? A measurement perspective on causal representations,” in 39th
Annual Conference on Neural Information Processing Systems, San Diego, CA,
United States, 2025, vol. 38.
ista: 'Yao D, Huang S, Cadei R, Zhang K, Locatello F. 2025. The third pillar of
causal analysis? A measurement perspective on causal representations. 39th Annual
Conference on Neural Information Processing Systems. NeurIPS: Neural Information
Processing Systems, Advances in Neural Information Processing Systems, vol. 38.'
mla: Yao, Dingling, et al. “The Third Pillar of Causal Analysis? A Measurement Perspective
on Causal Representations.” 39th Annual Conference on Neural Information Processing
Systems, vol. 38, Neural Information Processing Systems Foundation, 2025.
short: D. Yao, S. Huang, R. Cadei, K. Zhang, F. Locatello, in:, 39th Annual Conference
on Neural Information Processing Systems, Neural Information Processing Systems
Foundation, 2025.
conference:
end_date: 2025-12-07
location: San Diego, CA, United States
name: 'NeurIPS: Neural Information Processing Systems'
start_date: 2025-12-02
corr_author: '1'
date_created: 2026-01-29T14:24:56Z
date_published: 2025-12-15T00:00:00Z
date_updated: 2026-02-10T12:08:52Z
day: '15'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2505.17708'
has_accepted_license: '1'
intvolume: ' 38'
language:
- iso: eng
main_file_link:
- open_access: '1'
url: https://doi.org/10.48550/arXiv.2505.17708
month: '12'
oa: 1
oa_version: Preprint
publication: 39th Annual Conference on Neural Information Processing Systems
publication_identifier:
issn:
- 1049-5258
publication_status: epub_ahead
publisher: Neural Information Processing Systems Foundation
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/shimenghuang/a-measurement-perspective-of-crl
status: public
title: The third pillar of causal analysis? A measurement perspective on causal representations
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 38
year: '2025'
...
---
OA_place: repository
OA_type: green
_id: '21070'
abstract:
- lang: eng
text: 'Deep learning systems deployed in real-world applications often encounter
data that is different from their in-distribution (ID). A reliable model should
ideally abstain from making decisions in this out-of-distribution (OOD) setting.
Existing state-of-the-art methods primarily focus on feature distances, such as
k-th nearest neighbors and distances to decision boundaries, either overlooking
or ineffectively using in-distribution statistics. In this work, we propose a
novel angle-based metric for OOD detection that is computed relative to the in-distribution
structure. We demonstrate that the angles between feature representations and
decision boundaries, viewed from the mean of in-distribution features, serve as
an effective discriminative factor between ID and OOD data. We evaluate our method
on nine ImageNet-pretrained models. Our approach achieves the lowest FPR in 5
out of 9 ImageNet models, obtains the best average FPR overall, and consistently
ranking among the top 3 across all evaluated models. Furthermore, we highlight
the benefits of contrastive representations by showing strong performance with
ResNet SCL and CLIP architectures. Finally, we demonstrate that the scale-invariant
nature of our score enables an ensemble strategy via simple score summation. '
acknowledgement: "This research was funded in whole or in part by the Austrian Science
Fund (FWF) 10.55776/COE12. For open access purposes, the author has applied a CC
BY public copyright license to any accepted manuscript version arising from this
submission.\r\n"
alternative_title:
- Advances in Neural Information Processing Systems
article_processing_charge: No
arxiv: 1
author:
- first_name: Berker
full_name: Demirel, Berker
id: 8b4bc47f-3200-11ee-973b-8f0e7be21a9f
last_name: Demirel
- first_name: 'Marco '
full_name: 'Fumero, Marco '
last_name: Fumero
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Demirel B, Fumero M, Locatello F. Out-of-Distribution detection with relative
angles. In: 39th Annual Conference on Neural Information Processing Systems.
Vol 38. Neural Information Processing Systems Foundation; 2025.'
apa: 'Demirel, B., Fumero, M., & Locatello, F. (2025). Out-of-Distribution detection
with relative angles. In 39th Annual Conference on Neural Information Processing
Systems (Vol. 38). San Diego, CA, United States: Neural Information Processing
Systems Foundation.'
chicago: Demirel, Berker, Marco Fumero, and Francesco Locatello. “Out-of-Distribution
Detection with Relative Angles.” In 39th Annual Conference on Neural Information
Processing Systems, Vol. 38. Neural Information Processing Systems Foundation,
2025.
ieee: B. Demirel, M. Fumero, and F. Locatello, “Out-of-Distribution detection with
relative angles,” in 39th Annual Conference on Neural Information Processing
Systems, San Diego, CA, United States, 2025, vol. 38.
ista: 'Demirel B, Fumero M, Locatello F. 2025. Out-of-Distribution detection with
relative angles. 39th Annual Conference on Neural Information Processing Systems.
NeurIPS: Neural Information Processing Systems, Advances in Neural Information
Processing Systems, vol. 38.'
mla: Demirel, Berker, et al. “Out-of-Distribution Detection with Relative Angles.”
39th Annual Conference on Neural Information Processing Systems, vol. 38,
Neural Information Processing Systems Foundation, 2025.
short: B. Demirel, M. Fumero, F. Locatello, in:, 39th Annual Conference on Neural
Information Processing Systems, Neural Information Processing Systems Foundation,
2025.
conference:
end_date: 2025-12-07
location: San Diego, CA, United States
name: 'NeurIPS: Neural Information Processing Systems'
start_date: 2025-12-02
corr_author: '1'
date_created: 2026-01-29T14:26:47Z
date_published: 2025-12-01T00:00:00Z
date_updated: 2026-02-16T11:38:25Z
day: '01'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2410.04525'
has_accepted_license: '1'
intvolume: ' 38'
language:
- iso: eng
main_file_link:
- open_access: '1'
url: https://doi.org/10.48550/arXiv.2410.04525
month: '12'
oa: 1
oa_version: Preprint
publication: 39th Annual Conference on Neural Information Processing Systems
publication_identifier:
issn:
- 1049-5258
publication_status: epub_ahead
publisher: Neural Information Processing Systems Foundation
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/berkerdemirel/ORA-OOD-Detection-with-Relative-Angles
status: public
title: Out-of-Distribution detection with relative angles
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 38
year: '2025'
...
---
OA_place: publisher
OA_type: gold
_id: '21072'
abstract:
- lang: eng
text: Language and vision-language models have shown impressive performance across
a wide range of tasks, but their internal mechanisms remain only partly understood.
In this work, we study how individual attention heads in text-generative models
specialize in specific semantic or visual attributes. Building on an established
interpretability method, we reinterpret the practice of probing intermediate activations
with the final decoding layer through the lens of signal processing. This lets
us analyze multiple samples in a principled way and rank attention heads based
on their relevance to target concepts. Our results show consistent patterns of
specialization at the head level across both unimodal and multimodal transformers.
Remarkably, we find that editing as few as 1% of the heads, selected using our
method, can reliably suppress or enhance targeted concepts in the model output.
We validate our approach on language tasks such as question answering and toxicity
mitigation, as well as vision-language tasks including image classification and
captioning. Our findings highlight an interpretable and controllable structure
within attention layers, offering simple tools for understanding and editing large-scale
generative models.
acknowledgement: 'The authors acknowledge the Area Science Park supercomputing platform
ORFEO made available for conducting the research reported in this paper, and the
technical support of the Laboratory of Data Engineering staff. LB, DD and AC were
supported by the project “Supporto alla diagnosi di malattie rare tramite l’intelligenza
artificiale" CUP: F53C22001770002 and “Valutazione automatica delle immagini diagnostiche
tramite l’intelligenza artificiale", CUP: F53C22001780002. LB was supported by the
European Union – NextGenerationEU within the project PNRR “Finanziamento di progetti
presentati da giovani ricercatori" - Mission 4 Component 2 Investment 1.2, CUP:
J93C25000440001. AC was supported by the European Union – NextGenerationEU within
the project PNRR “PRP@CERIC" IR0000028 - Mission 4 Component 2 Investment 3.1 Action
3.1.1. '
article_processing_charge: No
arxiv: 1
author:
- first_name: Lorenzo
full_name: Basile, Lorenzo
last_name: Basile
- first_name: Valentino
full_name: Maiorca, Valentino
last_name: Maiorca
- first_name: Diego
full_name: Doimo, Diego
last_name: Doimo
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Alberto
full_name: Cazzaniga, Alberto
last_name: Cazzaniga
citation:
ama: 'Basile L, Maiorca V, Doimo D, Locatello F, Cazzaniga A. Head pursuit: Probing
attention specialization in multimodal transformers. In: 39th Annual Conference
on Neural Information Processing Systems. Vol 38. Neural Information Processing
Systems Foundation; 2025.'
apa: 'Basile, L., Maiorca, V., Doimo, D., Locatello, F., & Cazzaniga, A. (2025).
Head pursuit: Probing attention specialization in multimodal transformers. In
39th Annual Conference on Neural Information Processing Systems (Vol. 38).
San Diego, CA, United States: Neural Information Processing Systems Foundation.'
chicago: 'Basile, Lorenzo, Valentino Maiorca, Diego Doimo, Francesco Locatello,
and Alberto Cazzaniga. “Head Pursuit: Probing Attention Specialization in Multimodal
Transformers.” In 39th Annual Conference on Neural Information Processing Systems,
Vol. 38. Neural Information Processing Systems Foundation, 2025.'
ieee: 'L. Basile, V. Maiorca, D. Doimo, F. Locatello, and A. Cazzaniga, “Head pursuit:
Probing attention specialization in multimodal transformers,” in 39th Annual
Conference on Neural Information Processing Systems, San Diego, CA, United
States, 2025, vol. 38.'
ista: 'Basile L, Maiorca V, Doimo D, Locatello F, Cazzaniga A. 2025. Head pursuit:
Probing attention specialization in multimodal transformers. 39th Annual Conference
on Neural Information Processing Systems. NeurIPS: Neural Information Processing
Systems vol. 38.'
mla: 'Basile, Lorenzo, et al. “Head Pursuit: Probing Attention Specialization in
Multimodal Transformers.” 39th Annual Conference on Neural Information Processing
Systems, vol. 38, Neural Information Processing Systems Foundation, 2025.'
short: L. Basile, V. Maiorca, D. Doimo, F. Locatello, A. Cazzaniga, in:, 39th Annual
Conference on Neural Information Processing Systems, Neural Information Processing
Systems Foundation, 2025.
conference:
end_date: 2025-12-07
location: San Diego, CA, United States
name: 'NeurIPS: Neural Information Processing Systems'
start_date: 2025-12-02
date_created: 2026-01-29T14:29:23Z
date_published: 2025-12-15T00:00:00Z
date_updated: 2026-02-11T08:55:36Z
day: '15'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2510.21518'
file:
- access_level: open_access
checksum: 85be3f98663e2595cf37001852b477cb
content_type: application/pdf
creator: flocatel
date_created: 2026-01-29T14:29:14Z
date_updated: 2026-01-29T14:29:14Z
file_id: '21073'
file_name: 2510.21518v2.pdf
file_size: 4271547
relation: main_file
success: 1
file_date_updated: 2026-01-29T14:29:14Z
has_accepted_license: '1'
intvolume: ' 38'
language:
- iso: eng
main_file_link:
- open_access: '1'
url: https://doi.org/10.48550/arXiv.2510.21518
month: '12'
oa: 1
oa_version: Preprint
publication: 39th Annual Conference on Neural Information Processing Systems
publication_identifier:
issn:
- 1049-5258
publication_status: epub_ahead
publisher: Neural Information Processing Systems Foundation
quality_controlled: '1'
status: public
title: 'Head pursuit: Probing attention specialization in multimodal transformers'
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 38
year: '2025'
...
---
OA_place: publisher
OA_type: gold
_id: '21074'
abstract:
- lang: eng
text: Neural models learn representations of high-dimensional data on low-dimensional
manifolds. Multiple factors, including stochasticities in the training process,
model architectures, and additional inductive biases, may induce different representations,
even when learning the same task on the same data. However, it has recently been
shown that when a latent structure is shared between distinct latent spaces, relative
distances between representations can be preserved, up to distortions. Building
on this idea, we demonstrate that exploiting the differential-geometric structure
of latent spaces of neural models, it is possible to capture precisely the transformations
between representational spaces trained on similar data distributions. Specifically,
we assume that distinct neural models parametrize approximately the same underlying
manifold, and introduce a representation based on the pullback metric that captures
the intrinsic structure of the latent space, while scaling efficiently to large
models. We validate experimentally our method on model stitching and retrieval
tasks, covering autoencoders and vision foundation discriminative models, across
diverse architectures, datasets, pretraining schemes and modalities. Code is available
at the following link.
acknowledgement: 'We thank Gregor Krzmanc, German Magai, Vital Fernandez for insightful
discussions in the early stages of the project. HY was supported by the Research
Council of Finland Flagship programme: Finnish Center for Artificial Intelligence
FCAI. HY wishes to acknowledge CSC - IT Center for Science, Finland, for computational
resources. GA was supported by the DFF Sapere Aude Starting Grant “GADL”. SH was
supported by a research grant (42062) from VILLUM FONDEN and partly funded by the
Novo Nordisk Foundation through the Center for Basic Research in Life Science (NNF20OC0062606).
SH received funding from the European Research Council (ERC) under the European
Union’s Horizon Programme (grant agreement 101125003). MF is supported by the MSCA
IST-Bridge fellowship which has received funding from the European Union’s Horizon
2020 research and innovation program under the Marie Skłodowska-Curie grant agreement
No 101034413.'
alternative_title:
- Advances in Neural Information Processing Systems
article_processing_charge: No
arxiv: 1
author:
- first_name: Hanlin
full_name: Yu, Hanlin
last_name: Yu
- first_name: Befrin
full_name: Inal, Befrin
last_name: Inal
- first_name: Georgios
full_name: Arvanitidis, Georgios
last_name: Arvanitidis
- first_name: Soren
full_name: Hauberg, Soren
last_name: Hauberg
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Marco
full_name: Fumero, Marco
id: 1c1593eb-393f-11ef-bb8e-ab4f1e979650
last_name: Fumero
citation:
ama: 'Yu H, Inal B, Arvanitidis G, Hauberg S, Locatello F, Fumero M. Connecting
neural models latent geometries with relative geodesic representations. In: 39th
Annual Conference on Neural Information Processing Systems. Vol 38. Neural
Information Processing Systems Foundation; 2025.'
apa: 'Yu, H., Inal, B., Arvanitidis, G., Hauberg, S., Locatello, F., & Fumero,
M. (2025). Connecting neural models latent geometries with relative geodesic representations.
In 39th Annual Conference on Neural Information Processing Systems (Vol.
38). San Diego, CA, United States: Neural Information Processing Systems Foundation.'
chicago: Yu, Hanlin, Befrin Inal, Georgios Arvanitidis, Soren Hauberg, Francesco
Locatello, and Marco Fumero. “Connecting Neural Models Latent Geometries with
Relative Geodesic Representations.” In 39th Annual Conference on Neural Information
Processing Systems, Vol. 38. Neural Information Processing Systems Foundation,
2025.
ieee: H. Yu, B. Inal, G. Arvanitidis, S. Hauberg, F. Locatello, and M. Fumero, “Connecting
neural models latent geometries with relative geodesic representations,” in 39th
Annual Conference on Neural Information Processing Systems, San Diego, CA,
United States, 2025, vol. 38.
ista: 'Yu H, Inal B, Arvanitidis G, Hauberg S, Locatello F, Fumero M. 2025. Connecting
neural models latent geometries with relative geodesic representations. 39th Annual
Conference on Neural Information Processing Systems. NeurIPS: Neural Information
Processing Systems, Advances in Neural Information Processing Systems, vol. 38.'
mla: Yu, Hanlin, et al. “Connecting Neural Models Latent Geometries with Relative
Geodesic Representations.” 39th Annual Conference on Neural Information Processing
Systems, vol. 38, Neural Information Processing Systems Foundation, 2025.
short: H. Yu, B. Inal, G. Arvanitidis, S. Hauberg, F. Locatello, M. Fumero, in:,
39th Annual Conference on Neural Information Processing Systems, Neural Information
Processing Systems Foundation, 2025.
conference:
end_date: 2025-12-07
location: San Diego, CA, United States
name: 'NeurIPS: Neural Information Processing Systems'
start_date: 2025-12-02
corr_author: '1'
date_created: 2026-01-29T14:31:52Z
date_published: 2025-12-15T00:00:00Z
date_updated: 2026-02-11T09:03:37Z
day: '15'
ddc:
- '000'
department:
- _id: FrLo
ec_funded: 1
external_id:
arxiv:
- '2506.01599'
file:
- access_level: open_access
checksum: b1a645418025f46394764cd16d0cb089
content_type: application/pdf
creator: flocatel
date_created: 2026-01-29T14:31:42Z
date_updated: 2026-01-29T14:31:42Z
file_id: '21075'
file_name: 2506.01599v2.pdf
file_size: 7749349
relation: main_file
success: 1
file_date_updated: 2026-01-29T14:31:42Z
has_accepted_license: '1'
intvolume: ' 38'
language:
- iso: eng
month: '12'
oa: 1
oa_version: Published Version
project:
- _id: fc2ed2f7-9c52-11eb-aca3-c01059dda49c
call_identifier: H2020
grant_number: '101034413'
name: 'IST-BRIDGE: International postdoctoral program'
publication: 39th Annual Conference on Neural Information Processing Systems
publication_identifier:
issn:
- 1049-5258
publication_status: epub_ahead
publisher: Neural Information Processing Systems Foundation
quality_controlled: '1'
status: public
title: Connecting neural models latent geometries with relative geodesic representations
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 38
year: '2025'
...
---
OA_place: publisher
OA_type: gold
_id: '21076'
abstract:
- lang: eng
text: In many scientific experiments, the data annotating cost constraints the pace
for testing novel hypotheses. Yet, modern machine learning pipelines offer a promising
solution—provided their predictions yield correct conclusions. We focus on Prediction-Powered
Causal Inferences (PPCI), i.e., estimating the treatment effect in an unlabeled
target experiment, relying on training data with the same outcome annotated but
potentially different treatment or effect modifiers. We first show that conditional
calibration guarantees valid PPCI at population level. Then, we introduce a sufficient
representation constraint transferring validity across experiments, which we propose
to enforce in practice in Deconfounded Empirical Risk Minimization, our new model-agnostic
training objective. We validate our method on synthetic and real-world scientific
data, solving impossible problem instances for Empirical Risk Minimization even
with standard invariance constraints. In particular, for the first time, we achieve
valid causal inference on a scientific experiment with complex recording and no
human annotations, fine-tuning a foundational model on our similar annotated experiment.
acknowledgement: We thank the Causal Learning and Artificial Intelligence group at
ISTA for the continuous feedback on the project and valuable discussions. We thank
the Social Immunity group at ISTA, particularly Jinook Oh, for the annotation program
and Michaela Hoenigsberger for supporting our ecological experiment. Riccardo Cadei
is supported by a Google Research Scholar Award and a Google Initiated Gift to Francesco
Locatello. This research was funded in part by the Austrian Science Fund (FWF) 10.55776/COE12).
It was further partially supported by the ISTA Interdisciplinary Project Committee
for the collaborative project “ALED” between Francesco Locatello and Sylvia Cremer.
For open access purposes, the author has applied a CC BY public copyright license
to any author accepted manuscript version arising from this submission.
alternative_title:
- Advances in Neural Information Processing Systems
article_processing_charge: No
author:
- first_name: Riccardo
full_name: Cadei, Riccardo
id: 0fa8b76f-72f0-11ef-b75a-a5da96e5ad6b
last_name: Cadei
- first_name: Ilker
full_name: Demirel, Ilker
last_name: Demirel
- first_name: Piersilvio
full_name: De Bartolomeis, Piersilvio
last_name: De Bartolomeis
- first_name: Lukas
full_name: Lindorfer, Lukas
id: 85f0e6d3-06b3-11ec-8982-8c5049fa4455
last_name: Lindorfer
- first_name: Sylvia
full_name: Cremer, Sylvia
id: 2F64EC8C-F248-11E8-B48F-1D18A9856A87
last_name: Cremer
orcid: 0000-0002-2193-3868
- first_name: Cordelia
full_name: Schmid, Cordelia
last_name: Schmid
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Cadei R, Demirel I, De Bartolomeis P, et al. Prediction-powered causal inferences.
In: 39th Annual Conference on Neural Information Processing Systems. Vol
38. Neural Information Processing Systems Foundation; 2025.'
apa: 'Cadei, R., Demirel, I., De Bartolomeis, P., Lindorfer, L., Cremer, S., Schmid,
C., & Locatello, F. (2025). Prediction-powered causal inferences. In 39th
Annual Conference on Neural Information Processing Systems (Vol. 38). San
Diego, CA, United States: Neural Information Processing Systems Foundation.'
chicago: Cadei, Riccardo, Ilker Demirel, Piersilvio De Bartolomeis, Lukas Lindorfer,
Sylvia Cremer, Cordelia Schmid, and Francesco Locatello. “Prediction-Powered Causal
Inferences.” In 39th Annual Conference on Neural Information Processing Systems,
Vol. 38. Neural Information Processing Systems Foundation, 2025.
ieee: R. Cadei et al., “Prediction-powered causal inferences,” in 39th
Annual Conference on Neural Information Processing Systems, San Diego, CA,
United States, 2025, vol. 38.
ista: 'Cadei R, Demirel I, De Bartolomeis P, Lindorfer L, Cremer S, Schmid C, Locatello
F. 2025. Prediction-powered causal inferences. 39th Annual Conference on Neural
Information Processing Systems. NeurIPS: Neural Information Processing Systems,
Advances in Neural Information Processing Systems, vol. 38.'
mla: Cadei, Riccardo, et al. “Prediction-Powered Causal Inferences.” 39th Annual
Conference on Neural Information Processing Systems, vol. 38, Neural Information
Processing Systems Foundation, 2025.
short: R. Cadei, I. Demirel, P. De Bartolomeis, L. Lindorfer, S. Cremer, C. Schmid,
F. Locatello, in:, 39th Annual Conference on Neural Information Processing Systems,
Neural Information Processing Systems Foundation, 2025.
conference:
end_date: 2025-12-07
location: San Diego, CA, United States
name: 'NeurIPS: Neural Information Processing Systems'
start_date: 2025-12-02
date_created: 2026-01-29T14:35:11Z
date_published: 2025-12-15T00:00:00Z
date_updated: 2026-02-16T11:39:33Z
day: '15'
ddc:
- '000'
department:
- _id: FrLo
- _id: SyCr
file:
- access_level: open_access
checksum: 92467fa566cd36671a6a3b9e71ae0f71
content_type: application/pdf
creator: flocatel
date_created: 2026-01-29T14:35:02Z
date_updated: 2026-01-29T14:35:02Z
file_id: '21077'
file_name: 17546_Prediction_Powered_Causa.pdf
file_size: 8489023
relation: main_file
success: 1
file_date_updated: 2026-01-29T14:35:02Z
has_accepted_license: '1'
intvolume: ' 38'
language:
- iso: eng
month: '12'
oa: 1
oa_version: Published Version
publication: 39th Annual Conference on Neural Information Processing Systems
publication_identifier:
issn:
- 1049-5258
publication_status: epub_ahead
publisher: Neural Information Processing Systems Foundation
quality_controlled: '1'
status: public
title: Prediction-powered causal inferences
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 38
year: '2025'
...
---
OA_place: publisher
OA_type: gold
_id: '19010'
abstract:
- lang: eng
text: Causal representation learning aims at recovering latent causal variables
from high-dimensional observations to solve causal downstream tasks, such as predicting
the effect of new interventions or more robust classification. A plethora of methods
have been developed, each tackling carefully crafted problem settings that lead
to different types of identifiability. The folklore is that these different settings
are important, as they are often linked to different rungs of Pearl's causal hierarchy,
although not all neatly fit. Our main contribution is to show that many existing
causal representation learning approaches methodologically align the representation
to known data symmetries. Identification of the variables is guided by equivalence
classes across different "data pockets" that are not necessarily causal. This
result suggests important implications, allowing us to unify many existing approaches
in a single method that can mix and match different assumptions, including non-causal
ones, based on the invariances relevant to our application. It also significantly
benefits applicability, which we demonstrate by improving treatment effect estimation
on real-world high-dimensional ecological data. Overall, this paper clarifies
the role of causality assumptions in the discovery of causal variables and shifts
the focus to preserving data symmetries.
acknowledgement: "We thank Jiaqi Zhang, Francesco Montagna, David Lopez-Paz, Kartik
Ahuja, Thomas Kipf, Sara\r\nMagliacane, Julius von Kügelgen, Kun Zhang, and Bernhard
Schölkopf for extremely helpful discussion. Riccardo Cadei was supported by a Google
Research Scholar Award to Francesco Locatello. We acknowledge the Third Bellairs
Workshop on Causal Representation Learning held at the Bellairs Research Institute,
February 9/16, 2024, and a debate on the difference between interventions and counterfactuals
in disentanglement and CRL that took place during Dhanya Sridhar’s lecture, which
motivated us to significantly broaden the scope of the paper. We thank Dhanya and
all participants of the workshop."
article_processing_charge: No
arxiv: 1
author:
- first_name: Dingling
full_name: Yao, Dingling
id: d3e02e50-48a8-11ee-8f62-c108061797fa
last_name: Yao
- first_name: Dario
full_name: Rancati, Dario
id: feb58f2e-72ef-11ef-b75a-8f0894539cd0
last_name: Rancati
- first_name: Riccardo
full_name: Cadei, Riccardo
id: 0fa8b76f-72f0-11ef-b75a-a5da96e5ad6b
last_name: Cadei
- first_name: Marco
full_name: Fumero, Marco
id: 1c1593eb-393f-11ef-bb8e-ab4f1e979650
last_name: Fumero
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Yao D, Rancati D, Cadei R, Fumero M, Locatello F. Unifying causal representation
learning with the invariance principle. In: 13th International Conference on
Learning Representations. ICLR; 2025.'
apa: 'Yao, D., Rancati, D., Cadei, R., Fumero, M., & Locatello, F. (2025). Unifying
causal representation learning with the invariance principle. In 13th International
Conference on Learning Representations. Singapore: ICLR.'
chicago: Yao, Dingling, Dario Rancati, Riccardo Cadei, Marco Fumero, and Francesco
Locatello. “Unifying Causal Representation Learning with the Invariance Principle.”
In 13th International Conference on Learning Representations. ICLR, 2025.
ieee: D. Yao, D. Rancati, R. Cadei, M. Fumero, and F. Locatello, “Unifying causal
representation learning with the invariance principle,” in 13th International
Conference on Learning Representations, Singapore, 2025.
ista: 'Yao D, Rancati D, Cadei R, Fumero M, Locatello F. 2025. Unifying causal representation
learning with the invariance principle. 13th International Conference on Learning
Representations. ICLR: International Conference on Learning Representations.'
mla: Yao, Dingling, et al. “Unifying Causal Representation Learning with the Invariance
Principle.” 13th International Conference on Learning Representations,
ICLR, 2025.
short: D. Yao, D. Rancati, R. Cadei, M. Fumero, F. Locatello, in:, 13th International
Conference on Learning Representations, ICLR, 2025.
conference:
end_date: 2025-04-28
location: Singapore
name: 'ICLR: International Conference on Learning Representations'
start_date: 2025-04-24
corr_author: '1'
date_created: 2025-02-05T09:23:25Z
date_published: 2025-01-22T00:00:00Z
date_updated: 2026-02-09T05:52:14Z
day: '22'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2409.02772'
file:
- access_level: open_access
checksum: c4b5a4a644228c6d1b0283e1368bce9e
content_type: application/pdf
creator: flocatel
date_created: 2026-01-27T12:43:25Z
date_updated: 2026-01-27T12:43:25Z
file_id: '21048'
file_name: 4356_Unifying_Causal_Represent (1).pdf
file_size: 877014
relation: main_file
success: 1
file_date_updated: 2026-01-27T12:43:25Z
has_accepted_license: '1'
language:
- iso: eng
month: '01'
oa: 1
oa_version: Published Version
publication: 13th International Conference on Learning Representations
publication_status: published
publisher: ICLR
quality_controlled: '1'
scopus_import: '1'
status: public
title: Unifying causal representation learning with the invariance principle
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2025'
...
---
OA_place: repository
OA_type: green
_id: '19674'
abstract:
- lang: eng
text: 'When examined through the lens of their residual streams, a puzzling property
emerges in transformer networks: residual contributions (e.g., attention heads)
sometimes specialize in specific tasks or input attributes. In this paper, we
analyze this phenomenon in vision transformers, focusing on the spectral geometry
of residuals, and explore its implications for modality alignment in vision-language
models. First, we link it to the intrinsically low-dimensional structure of visual
head representations, zooming into their principal components and showing that
they encode specialized roles across a wide variety of input data distributions.
Then, we analyze the effect of head specialization in multimodal models, focusing
on how improved alignment between text and specialized heads impacts zero-shot
classification performance. This specialization-performance link consistently
holds across diverse pre-training data, network sizes, and objectives, demonstrating
a powerful new mechanism for boosting zero-shot classification through targeted
alignment. Ultimately, we translate these insights into actionable terms by introducing
ResiDual, a technique for spectral alignment of the residual stream. Much like
panning for gold, it lets the noise from irrelevant unit principal components
(i.e., attributes) wash away to amplify task-relevant ones. Remarkably, this dual
perspective on modality alignment yields fine-tuning level performance on different
data distributions while modelling an extremely interpretable and parameter-efficient
transformation, as we extensively show on 70 pre-trained network-dataset combinations
(7 models, 10 datasets).'
acknowledgement: "The authors gratefully acknowledge Volkan Cevher for an insightful
discussion about sparse recovery algorithms, Alex Smola for valuable feedback on
the experiments, and Marco Baroni for an engaging conversation on the phenomenon
of head specialization in NLP.\r\n"
article_number: '2411.00246'
article_processing_charge: No
arxiv: 1
author:
- first_name: Lorenzo
full_name: Basile, Lorenzo
last_name: Basile
- first_name: Valentino
full_name: Maiorca, Valentino
last_name: Maiorca
- first_name: Luca
full_name: Bortolussi, Luca
last_name: Bortolussi
- first_name: Emanuele
full_name: Rodolà, Emanuele
last_name: Rodolà
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: Basile L, Maiorca V, Bortolussi L, Rodolà E, Locatello F. ResiDual transformer
alignment with spectral decomposition. arXiv. doi:10.48550/arXiv.2411.00246
apa: Basile, L., Maiorca, V., Bortolussi, L., Rodolà, E., & Locatello, F. (n.d.).
ResiDual transformer alignment with spectral decomposition. arXiv. https://doi.org/10.48550/arXiv.2411.00246
chicago: Basile, Lorenzo, Valentino Maiorca, Luca Bortolussi, Emanuele Rodolà, and
Francesco Locatello. “ResiDual Transformer Alignment with Spectral Decomposition.”
ArXiv, n.d. https://doi.org/10.48550/arXiv.2411.00246.
ieee: L. Basile, V. Maiorca, L. Bortolussi, E. Rodolà, and F. Locatello, “ResiDual
transformer alignment with spectral decomposition,” arXiv. .
ista: Basile L, Maiorca V, Bortolussi L, Rodolà E, Locatello F. ResiDual transformer
alignment with spectral decomposition. arXiv, 2411.00246.
mla: Basile, Lorenzo, et al. “ResiDual Transformer Alignment with Spectral Decomposition.”
ArXiv, 2411.00246, doi:10.48550/arXiv.2411.00246.
short: L. Basile, V. Maiorca, L. Bortolussi, E. Rodolà, F. Locatello, ArXiv (n.d.).
date_created: 2025-05-11T22:02:41Z
date_published: 2025-04-14T00:00:00Z
date_updated: 2025-05-19T07:03:16Z
day: '14'
department:
- _id: FrLo
doi: 10.48550/arXiv.2411.00246
external_id:
arxiv:
- '2411.00246'
language:
- iso: eng
main_file_link:
- open_access: '1'
url: https://doi.org/10.48550/arXiv.2411.00246
month: '04'
oa: 1
oa_version: Preprint
publication: arXiv
publication_status: submitted
status: public
title: ResiDual transformer alignment with spectral decomposition
type: preprint
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2025'
...
---
OA_place: publisher
OA_type: diamond
_id: '20032'
abstract:
- lang: eng
text: We propose Scalable Mechanistic Neural Network (S-MNN), an enhanced neural
network framework designed for scientific machine learning applications involving
long temporal sequences. By reformulating the original Mechanistic Neural Network
(MNN) (Pervez et al., 2024), we reduce the computational time and space complexities
from cubic and quadratic with respect to the sequence length, respectively, to
linear. This significant improvement enables efficient modeling of long-term dynamics
without sacrificing accuracy or interpretability. Extensive experiments demonstrate
that S-MNN matches the original MNN in precision while substantially reducing
computational resources. Consequently, S-MNN can drop-in replace the original
MNN in applications, providing a practical and efficient tool for integrating
mechanistic bottlenecks into neural network models of complex dynamical systems.
Source code is available at https://github.com/IST-DASLab/ScalableMNN.
article_processing_charge: No
arxiv: 1
author:
- first_name: Jiale
full_name: Chen, Jiale
id: 4d0a9064-1ff6-11ee-9fa6-ec046c604785
last_name: Chen
orcid: 0000-0001-5337-5875
- first_name: Dingling
full_name: Yao, Dingling
id: d3e02e50-48a8-11ee-8f62-c108061797fa
last_name: Yao
- first_name: Adeel A
full_name: Pervez, Adeel A
id: fca6d90c-d47f-11ee-bc87-93ff51604981
last_name: Pervez
- first_name: Dan-Adrian
full_name: Alistarh, Dan-Adrian
id: 4A899BFC-F248-11E8-B48F-1D18A9856A87
last_name: Alistarh
orcid: 0000-0003-3650-940X
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Chen J, Yao D, Pervez AA, Alistarh D-A, Locatello F. Scalable mechanistic
neural networks. In: 13th International Conference on Learning Representations.
ICLR; 2025:63716-63737.'
apa: 'Chen, J., Yao, D., Pervez, A. A., Alistarh, D.-A., & Locatello, F. (2025).
Scalable mechanistic neural networks. In 13th International Conference on Learning
Representations (pp. 63716–63737). Singapore, Singapore: ICLR.'
chicago: Chen, Jiale, Dingling Yao, Adeel A Pervez, Dan-Adrian Alistarh, and Francesco
Locatello. “Scalable Mechanistic Neural Networks.” In 13th International Conference
on Learning Representations, 63716–37. ICLR, 2025.
ieee: J. Chen, D. Yao, A. A. Pervez, D.-A. Alistarh, and F. Locatello, “Scalable
mechanistic neural networks,” in 13th International Conference on Learning
Representations, Singapore, Singapore, 2025, pp. 63716–63737.
ista: 'Chen J, Yao D, Pervez AA, Alistarh D-A, Locatello F. 2025. Scalable mechanistic
neural networks. 13th International Conference on Learning Representations. ICLR:
International Conference on Learning Representations, 63716–63737.'
mla: Chen, Jiale, et al. “Scalable Mechanistic Neural Networks.” 13th International
Conference on Learning Representations, ICLR, 2025, pp. 63716–37.
short: J. Chen, D. Yao, A.A. Pervez, D.-A. Alistarh, F. Locatello, in:, 13th International
Conference on Learning Representations, ICLR, 2025, pp. 63716–63737.
conference:
end_date: 2025-04-28
location: Singapore, Singapore
name: 'ICLR: International Conference on Learning Representations'
start_date: 2025-04-24
corr_author: '1'
date_created: 2025-07-20T22:02:01Z
date_published: 2025-04-01T00:00:00Z
date_updated: 2025-08-04T08:03:11Z
day: '01'
ddc:
- '000'
department:
- _id: DaAl
- _id: FrLo
external_id:
arxiv:
- '2410.06074'
file:
- access_level: open_access
checksum: 64cfdb12ae3e4e8ba57b1403e1066776
content_type: application/pdf
creator: dernst
date_created: 2025-07-22T07:58:22Z
date_updated: 2025-07-22T07:58:22Z
file_id: '20065'
file_name: 2025_ICLR_Chen.pdf
file_size: 732745
relation: main_file
success: 1
file_date_updated: 2025-07-22T07:58:22Z
has_accepted_license: '1'
language:
- iso: eng
month: '04'
oa: 1
oa_version: Published Version
page: 63716-63737
publication: 13th International Conference on Learning Representations
publication_identifier:
isbn:
- '9798331320850'
publication_status: published
publisher: ICLR
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/IST-DASLab/ScalableMNN
scopus_import: '1'
status: public
title: Scalable mechanistic neural networks
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2025'
...
---
OA_place: publisher
OA_type: diamond
_id: '20036'
abstract:
- lang: eng
text: 'We introduce NeCo: Patch Neighbor Consistency, a novel self-supervised training
loss that enforces patch-level nearest neighbor consistency across a student and
teacher model. Compared to contrastive approaches that only yield binary learning
signals, i.e. "attract" and "repel", this approach benefits from the more fine-grained
learning signal of sorting spatially dense features relative to reference patches.
Our method leverages differentiable sorting applied on top of pretrained representations,
such as DINOv2-registers to bootstrap the learning signal and further improve
upon them. This dense post-pretraining leads to superior performance across various
models and datasets, despite requiring only 19 hours on a single GPU. This method
generates high-quality dense feature encoders and establishes several new state-of-the-art
results such as +2.3 % and +4.2% for non-parametric in-context semantic segmentation
on ADE20k and Pascal VOC, +1.6% and +4.8% for linear segmentation evaluations
on COCO-Things and -Stuff and improvements in the 3D understanding of multi-view
consistency on SPair-71k, by more than 1.5%.'
article_processing_charge: No
arxiv: 1
author:
- first_name: Valentinos
full_name: Pariza, Valentinos
last_name: Pariza
- first_name: Mohammadreza
full_name: Salehi, Mohammadreza
last_name: Salehi
- first_name: Gertjan
full_name: Burghouts, Gertjan
last_name: Burghouts
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Yuki M.
full_name: Asano, Yuki M.
last_name: Asano
citation:
ama: 'Pariza V, Salehi M, Burghouts G, Locatello F, Asano YM. Near, far: Patch-ordering
enhances vision foundation models’ scene understanding. In: 13th International
Conference on Learning Representations. ICLR; 2025:72303-72330.'
apa: 'Pariza, V., Salehi, M., Burghouts, G., Locatello, F., & Asano, Y. M. (2025).
Near, far: Patch-ordering enhances vision foundation models’ scene understanding.
In 13th International Conference on Learning Representations (pp. 72303–72330).
Singapore, Singapore: ICLR.'
chicago: 'Pariza, Valentinos, Mohammadreza Salehi, Gertjan Burghouts, Francesco
Locatello, and Yuki M. Asano. “Near, Far: Patch-Ordering Enhances Vision Foundation
Models’ Scene Understanding.” In 13th International Conference on Learning
Representations, 72303–30. ICLR, 2025.'
ieee: 'V. Pariza, M. Salehi, G. Burghouts, F. Locatello, and Y. M. Asano, “Near,
far: Patch-ordering enhances vision foundation models’ scene understanding,” in
13th International Conference on Learning Representations, Singapore, Singapore,
2025, pp. 72303–72330.'
ista: 'Pariza V, Salehi M, Burghouts G, Locatello F, Asano YM. 2025. Near, far:
Patch-ordering enhances vision foundation models’ scene understanding. 13th International
Conference on Learning Representations. ICLR: International Conference on Learning
Representations, 72303–72330.'
mla: 'Pariza, Valentinos, et al. “Near, Far: Patch-Ordering Enhances Vision Foundation
Models’ Scene Understanding.” 13th International Conference on Learning Representations,
ICLR, 2025, pp. 72303–30.'
short: V. Pariza, M. Salehi, G. Burghouts, F. Locatello, Y.M. Asano, in:, 13th International
Conference on Learning Representations, ICLR, 2025, pp. 72303–72330.
conference:
end_date: 2025-04-28
location: Singapore, Singapore
name: 'ICLR: International Conference on Learning Representations'
start_date: 2025-04-24
date_created: 2025-07-20T22:02:03Z
date_published: 2025-04-01T00:00:00Z
date_updated: 2025-08-04T08:10:55Z
day: '01'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2408.11054'
file:
- access_level: open_access
checksum: ddbe981f3ad3f6cb6daf12c954822eb8
content_type: application/pdf
creator: dernst
date_created: 2025-08-04T08:09:43Z
date_updated: 2025-08-04T08:09:43Z
file_id: '20109'
file_name: 2025_ICLR_Pariza.pdf
file_size: 37788223
relation: main_file
success: 1
file_date_updated: 2025-08-04T08:09:43Z
has_accepted_license: '1'
language:
- iso: eng
month: '04'
oa: 1
oa_version: Published Version
page: 72303-72330
publication: 13th International Conference on Learning Representations
publication_identifier:
isbn:
- '9798331320850'
publication_status: published
publisher: ICLR
quality_controlled: '1'
scopus_import: '1'
status: public
title: 'Near, far: Patch-ordering enhances vision foundation models'' scene understanding'
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2025'
...
---
OA_place: publisher
OA_type: gold
_id: '20817'
abstract:
- lang: eng
text: We present Mechanistic PDE Networks -- a model for discovery of governing
partial differential equations from data. Mechanistic PDE Networks represent spatiotemporal
data as space-time dependent linear partial differential equations in neural network
hidden representations. The represented PDEs are then solved and decoded for specific
tasks. The learned PDE representations naturally express the spatiotemporal dynamics
in data in neural network hidden space, enabling increased modeling power. Solving
the PDE representations in a compute and memory-efficient way, however, is a significant
challenge. We develop a native, GPU-capable, parallel, sparse and differentiable
multigrid solver specialized for linear partial differential equations that acts
as a module in Mechanistic PDE Networks. Leveraging the PDE solver we propose
a discovery architecture that can discovers nonlinear PDEs in complex settings,
while being robust to noise. We validate PDE discovery on a number of PDEs including
reaction-diffusion and Navier-Stokes equations.
acknowledgement: "AP. This project has received funding from the European Union’s
Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie
Grant Agreement No. 101034413.\r\nFL. This research was funded in whole or in part
by the Austrian Science Fund (FWF) 10.55776/COE12. For open access purposes, the
author has applied a CC BY public\r\ncopyright license to any author accepted manuscript
version arising from this submission."
alternative_title:
- PMLR
article_processing_charge: No
arxiv: 1
author:
- first_name: Adeel A
full_name: Pervez, Adeel A
id: fca6d90c-d47f-11ee-bc87-93ff51604981
last_name: Pervez
- first_name: Efstratios
full_name: Gavves, Efstratios
last_name: Gavves
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Pervez AA, Gavves E, Locatello F. Mechanistic PDE networks for discovery of
governing equations. In: 42nd International Conference on Machine Learning.
Vol 267. ML Research Press; 2025:48962-48973.'
apa: 'Pervez, A. A., Gavves, E., & Locatello, F. (2025). Mechanistic PDE networks
for discovery of governing equations. In 42nd International Conference on Machine
Learning (Vol. 267, pp. 48962–48973). Vancouver, Canada: ML Research Press.'
chicago: Pervez, Adeel A, Efstratios Gavves, and Francesco Locatello. “Mechanistic
PDE Networks for Discovery of Governing Equations.” In 42nd International Conference
on Machine Learning, 267:48962–73. ML Research Press, 2025.
ieee: A. A. Pervez, E. Gavves, and F. Locatello, “Mechanistic PDE networks for discovery
of governing equations,” in 42nd International Conference on Machine Learning,
Vancouver, Canada, 2025, vol. 267, pp. 48962–48973.
ista: 'Pervez AA, Gavves E, Locatello F. 2025. Mechanistic PDE networks for discovery
of governing equations. 42nd International Conference on Machine Learning. ICML:
International Conference on Machine Learning, PMLR, vol. 267, 48962–48973.'
mla: Pervez, Adeel A., et al. “Mechanistic PDE Networks for Discovery of Governing
Equations.” 42nd International Conference on Machine Learning, vol. 267,
ML Research Press, 2025, pp. 48962–73.
short: A.A. Pervez, E. Gavves, F. Locatello, in:, 42nd International Conference
on Machine Learning, ML Research Press, 2025, pp. 48962–48973.
conference:
end_date: 2025-07-19
location: Vancouver, Canada
name: 'ICML: International Conference on Machine Learning'
start_date: 2025-07-13
corr_author: '1'
date_created: 2025-12-14T23:02:04Z
date_published: 2025-05-01T00:00:00Z
date_updated: 2025-12-16T12:24:55Z
day: '01'
ddc:
- '000'
department:
- _id: FrLo
ec_funded: 1
external_id:
arxiv:
- '2502.18377'
file:
- access_level: open_access
checksum: 933cb673fb41416f537278fb990df6c3
content_type: application/pdf
creator: dernst
date_created: 2025-12-16T12:21:49Z
date_updated: 2025-12-16T12:21:49Z
file_id: '20827'
file_name: 2025_ICML_Pervez.pdf
file_size: 993381
relation: main_file
success: 1
file_date_updated: 2025-12-16T12:21:49Z
has_accepted_license: '1'
intvolume: ' 267'
language:
- iso: eng
month: '05'
oa: 1
oa_version: Published Version
page: 48962-48973
project:
- _id: fc2ed2f7-9c52-11eb-aca3-c01059dda49c
call_identifier: H2020
grant_number: '101034413'
name: 'IST-BRIDGE: International postdoctoral program'
publication: 42nd International Conference on Machine Learning
publication_identifier:
eissn:
- 2640-3498
publication_status: published
publisher: ML Research Press
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/ alpz/mech-nn-discovery-pde
scopus_import: '1'
status: public
title: Mechanistic PDE networks for discovery of governing equations
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 267
year: '2025'
...
---
_id: '14213'
abstract:
- lang: eng
text: We introduce a method to segment the visual field into independently moving
regions, trained with no ground truth or supervision. It consists of an adversarial
conditional encoder-decoder architecture based on Slot Attention, modified to
use the image as context to decode optical flow without attempting to reconstruct
the image itself. In the resulting multi-modal representation, one modality (flow)
feeds the encoder to produce separate latent codes (slots), whereas the other
modality (image) conditions the decoder to generate the first (flow) from the
slots. This design frees the representation from having to encode complex nuisance
variability in the image due to, for instance, illumination and reflectance properties
of the scene. Since customary autoencoding based on minimizing the reconstruction
error does not preclude the entire flow from being encoded into a single slot,
we modify the loss to an adversarial criterion based on Contextual Information
Separation. The resulting min-max optimization fosters the separation of objects
and their assignment to different attention slots, leading to Divided Attention,
or DivA. DivA outperforms recent unsupervised multi-object motion segmentation
methods while tripling run-time speed up to 104FPS and reducing the performance
gap from supervised methods to 12% or less. DivA can handle different numbers
of objects and different image sizes at training and test time, is invariant to
permutation of object labels, and does not require explicit regularization.
article_processing_charge: No
arxiv: 1
author:
- first_name: Dong
full_name: Lao, Dong
last_name: Lao
- first_name: Zhengyang
full_name: Hu, Zhengyang
last_name: Hu
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Yanchao
full_name: Yang, Yanchao
last_name: Yang
- first_name: Stefano
full_name: Soatto, Stefano
last_name: Soatto
citation:
ama: 'Lao D, Hu Z, Locatello F, Yang Y, Soatto S. Divided attention: Unsupervised
multi-object discovery with contextually separated slots. In: 1st Conference
on Parsimony and Learning. ; 2024.'
apa: 'Lao, D., Hu, Z., Locatello, F., Yang, Y., & Soatto, S. (2024). Divided
attention: Unsupervised multi-object discovery with contextually separated slots.
In 1st Conference on Parsimony and Learning. Hong Kong, China.'
chicago: 'Lao, Dong, Zhengyang Hu, Francesco Locatello, Yanchao Yang, and Stefano
Soatto. “Divided Attention: Unsupervised Multi-Object Discovery with Contextually
Separated Slots.” In 1st Conference on Parsimony and Learning, 2024.'
ieee: 'D. Lao, Z. Hu, F. Locatello, Y. Yang, and S. Soatto, “Divided attention:
Unsupervised multi-object discovery with contextually separated slots,” in 1st
Conference on Parsimony and Learning, Hong Kong, China, 2024.'
ista: 'Lao D, Hu Z, Locatello F, Yang Y, Soatto S. 2024. Divided attention: Unsupervised
multi-object discovery with contextually separated slots. 1st Conference on Parsimony
and Learning. CPAL: Conference on Parsimony and Learning.'
mla: 'Lao, Dong, et al. “Divided Attention: Unsupervised Multi-Object Discovery
with Contextually Separated Slots.” 1st Conference on Parsimony and Learning,
2024.'
short: D. Lao, Z. Hu, F. Locatello, Y. Yang, S. Soatto, in:, 1st Conference on Parsimony
and Learning, 2024.
conference:
end_date: 2024-01-03
location: Hong Kong, China
name: 'CPAL: Conference on Parsimony and Learning'
start_date: 2024-01-03
date_created: 2023-08-22T14:19:59Z
date_published: 2024-01-03T00:00:00Z
date_updated: 2024-02-12T08:56:23Z
day: '03'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2304.01430'
file:
- access_level: open_access
checksum: 8fad894c34f1b3d5a14fb8ffb12f7277
content_type: application/pdf
creator: dernst
date_created: 2024-02-12T08:40:36Z
date_updated: 2024-02-12T08:40:36Z
file_id: '14978'
file_name: 2024_CPAL_Lao.pdf
file_size: 8038511
relation: main_file
success: 1
file_date_updated: 2024-02-12T08:40:36Z
has_accepted_license: '1'
language:
- iso: eng
month: '01'
oa: 1
oa_version: Published Version
publication: 1st Conference on Parsimony and Learning
publication_status: published
quality_controlled: '1'
status: public
title: 'Divided attention: Unsupervised multi-object discovery with contextually separated
slots'
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2024'
...
---
OA_place: repository
OA_type: green
_id: '14946'
abstract:
- lang: eng
text: We present a unified framework for studying the identifiability of representations
learned from simultaneously observed views, such as different data modalities.
We allow a partially observed setting in which each view constitutes a nonlinear
mixture of a subset of underlying latent variables, which can be causally related.
We prove that the information shared across all subsets of any number of views
can be learned up to a smooth bijection using contrastive learning and a single
encoder per view. We also provide graphical criteria indicating which latent variables
can be identified through a simple set of rules, which we refer to as identifiability
algebra. Our general framework and theoretical results unify and extend several
previous work on multi-view nonlinear ICA, disentanglement, and causal representation
learning. We experimentally validate our claims on numerical, image, and multi-modal
data sets. Further, we demonstrate that the performance of prior methods is recovered
in different special cases of our setup. Overall, we find that access to multiple
partial views offers unique opportunities for identifiable representation learning,
enabling the discovery of latent structures from purely observational data.
acknowledgement: 'This work was initiated at the Second Bellairs Workshop on Causality
held at the Bellairs Research Institute, January 6–13, 2022; we thank all workshop
participants for providing a stimulating research environment. Further, we thank
Cian Eastwood, Luigi Gresele, Stefano Soatto, Marco Bagatella and A. René Geist
for helpful discussion. GM is a member of the Machine Learning Cluster of Excellence,
EXC number 2064/1 – Project number 390727645. JvK and GM acknowledge support from
the German Federal Ministry of Education and Research (BMBF) through the Tübingen
AI Center (FKZ: 01IS18039B). The research of DX and SM was supported by the Air
Force Office of Scientific Research under award number FA8655-22-1-7155. Any opinions,
findings, and conclusions or recommendations expressed in this material are those
of the author(s) and do not necessarily reflect the views of the United States Air
Force. We also thank SURF for the support in using the Dutch National Supercomputer
Snellius. SL was supported by an IVADO excellence PhD scholarship and by Samsung
Electronics Co., Ldt. DY was supported by an Amazon fellowship, the International
Max Planck Research School for Intelligent Systems (IMPRS-IS) and the ISTA graduate
school. Work done outside of Amazon.'
article_processing_charge: No
arxiv: 1
author:
- first_name: Dingling
full_name: Yao, Dingling
id: d3e02e50-48a8-11ee-8f62-c108061797fa
last_name: Yao
- first_name: Danru
full_name: Xu, Danru
last_name: Xu
- first_name: Sébastien
full_name: Lachapelle, Sébastien
last_name: Lachapelle
- first_name: Sara
full_name: Magliacane, Sara
last_name: Magliacane
- first_name: Perouz
full_name: Taslakian, Perouz
last_name: Taslakian
- first_name: Georg
full_name: Martius, Georg
last_name: Martius
- first_name: Julius von
full_name: Kügelgen, Julius von
last_name: Kügelgen
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Yao D, Xu D, Lachapelle S, et al. Multi-view causal representation learning
with partial observability. In: 12th International Conference on Learning Representations.
Curran Associates; 2024.'
apa: 'Yao, D., Xu, D., Lachapelle, S., Magliacane, S., Taslakian, P., Martius, G.,
… Locatello, F. (2024). Multi-view causal representation learning with partial
observability. In 12th International Conference on Learning Representations.
Vienna, Austria: Curran Associates.'
chicago: Yao, Dingling, Danru Xu, Sébastien Lachapelle, Sara Magliacane, Perouz
Taslakian, Georg Martius, Julius von Kügelgen, and Francesco Locatello. “Multi-View
Causal Representation Learning with Partial Observability.” In 12th International
Conference on Learning Representations. Curran Associates, 2024.
ieee: D. Yao et al., “Multi-view causal representation learning with partial
observability,” in 12th International Conference on Learning Representations,
Vienna, Austria, 2024.
ista: 'Yao D, Xu D, Lachapelle S, Magliacane S, Taslakian P, Martius G, Kügelgen
J von, Locatello F. 2024. Multi-view causal representation learning with partial
observability. 12th International Conference on Learning Representations. ICLR:
International Conference on Learning Representations.'
mla: Yao, Dingling, et al. “Multi-View Causal Representation Learning with Partial
Observability.” 12th International Conference on Learning Representations,
Curran Associates, 2024.
short: D. Yao, D. Xu, S. Lachapelle, S. Magliacane, P. Taslakian, G. Martius, J.
von Kügelgen, F. Locatello, in:, 12th International Conference on Learning Representations,
Curran Associates, 2024.
conference:
end_date: 2024-05-07
location: Vienna, Austria
name: 'ICLR: International Conference on Learning Representations'
start_date: 2024-05-07
corr_author: '1'
date_created: 2024-02-07T14:28:34Z
date_published: 2024-11-07T00:00:00Z
date_updated: 2025-02-11T10:34:32Z
day: '07'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2311.04056'
file:
- access_level: open_access
checksum: 8ed3c34706eeec622c7e8968dc0f747a
content_type: application/pdf
creator: dernst
date_created: 2025-02-04T12:34:23Z
date_updated: 2025-02-04T12:34:23Z
file_id: '18995'
file_name: 2024_ICLR_Yao.pdf
file_size: 1713606
relation: main_file
success: 1
file_date_updated: 2025-02-04T12:34:23Z
has_accepted_license: '1'
language:
- iso: eng
month: '11'
oa: 1
oa_version: Published Version
publication: 12th International Conference on Learning Representations
publication_status: published
publisher: Curran Associates
quality_controlled: '1'
status: public
title: Multi-view causal representation learning with partial observability
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2024'
...
---
OA_place: publisher
OA_type: gold
_id: '18847'
abstract:
- lang: eng
text: "Machine Learning and AI have the potential to transform data-driven\r\nscientific
discovery, enabling accurate predictions for several scientific\r\nphenomena.
As many scientific questions are inherently causal, this paper looks\r\nat the
causal inference task of treatment effect estimation, where the outcome\r\nof
interest is recorded in high-dimensional observations in a Randomized\r\nControlled
Trial (RCT). Despite being the simplest possible causal setting and\r\na perfect
fit for deep learning, we theoretically find that many common choices\r\nin the
literature may lead to biased estimates. To test the practical impact of\r\nthese
considerations, we recorded ISTAnt, the first real-world benchmark for\r\ncausal
inference downstream tasks on high-dimensional observations as an RCT\r\nstudying
how garden ants (Lasius neglectus) respond to microparticles applied\r\nonto their
colony members by hygienic grooming. Comparing 6 480 models\r\nfine-tuned from
state-of-the-art visual backbones, we find that the sampling\r\nand modeling choices
significantly affect the accuracy of the causal estimate,\r\nand that classification
accuracy is not a proxy thereof. We further validated\r\nthe analysis, repeating
it on a synthetically generated visual data set\r\ncontrolling the causal model.
Our results suggest that future benchmarks should\r\ncarefully consider real downstream
scientific questions, especially causal\r\nones. Further, we highlight guidelines
for representation learning methods to\r\nhelp answer causal questions in the
sciences."
acknowledgement: We thank Piersilvio De Bartolomeis, and the full Causal Learning
and Artificial Intelligence (CLAI) group at ISTA for the extremely helpful discussions.
Riccardo Cadei was supported by a Google Research Scholar Award and a Google Initiated
Gift to Francesco Locatello. We thank the Social Immunity team at ISTA particularly
Michaela Hönigsberger and Wilfrid Jean Louis, for supporting the ecological experiment
and Farnaz Beikzadeh Abbasi, Luisa Fiebig and Martin Estermann for annotating ant
behavior in ISTAnt.
article_processing_charge: No
arxiv: 1
author:
- first_name: Riccardo
full_name: Cadei, Riccardo
id: 0fa8b76f-72f0-11ef-b75a-a5da96e5ad6b
last_name: Cadei
- first_name: Lukas
full_name: Lindorfer, Lukas
id: 85f0e6d3-06b3-11ec-8982-8c5049fa4455
last_name: Lindorfer
- first_name: Sylvia
full_name: Cremer, Sylvia
id: 2F64EC8C-F248-11E8-B48F-1D18A9856A87
last_name: Cremer
orcid: 0000-0002-2193-3868
- first_name: Cordelia
full_name: Schmid, Cordelia
last_name: Schmid
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
citation:
ama: 'Cadei R, Lindorfer L, Cremer S, Schmid C, Locatello F. Smoke and mirrors in
causal downstream tasks. In: ICML 2024 Workshop AI4Science. Vol 38. Curran
Associates; 2024.'
apa: Cadei, R., Lindorfer, L., Cremer, S., Schmid, C., & Locatello, F. (2024).
Smoke and mirrors in causal downstream tasks. In ICML 2024 Workshop AI4Science
(Vol. 38). Curran Associates.
chicago: Cadei, Riccardo, Lukas Lindorfer, Sylvia Cremer, Cordelia Schmid, and Francesco
Locatello. “Smoke and Mirrors in Causal Downstream Tasks.” In ICML 2024 Workshop
AI4Science, Vol. 38. Curran Associates, 2024.
ieee: R. Cadei, L. Lindorfer, S. Cremer, C. Schmid, and F. Locatello, “Smoke and
mirrors in causal downstream tasks,” in ICML 2024 Workshop AI4Science,
2024, vol. 38.
ista: 'Cadei R, Lindorfer L, Cremer S, Schmid C, Locatello F. 2024. Smoke and mirrors
in causal downstream tasks. ICML 2024 Workshop AI4Science. ICML: International
Conference on Machine Learning vol. 38.'
mla: Cadei, Riccardo, et al. “Smoke and Mirrors in Causal Downstream Tasks.” ICML
2024 Workshop AI4Science, vol. 38, Curran Associates, 2024.
short: R. Cadei, L. Lindorfer, S. Cremer, C. Schmid, F. Locatello, in:, ICML 2024
Workshop AI4Science, Curran Associates, 2024.
conference:
end_date: 2024-07-26
name: 'ICML: International Conference on Machine Learning'
start_date: 2024-07-26
corr_author: '1'
date_created: 2025-01-14T07:27:26Z
date_published: 2024-09-25T00:00:00Z
date_updated: 2025-07-10T11:51:50Z
day: '25'
ddc:
- '000'
- '570'
department:
- _id: SyCr
- _id: FrLo
- _id: GradSch
external_id:
arxiv:
- '2405.17151'
file:
- access_level: open_access
checksum: beedf05388bbdb7ddda81ec3d5ec7026
content_type: application/pdf
creator: dernst
date_created: 2025-01-27T11:42:24Z
date_updated: 2025-01-27T11:42:24Z
file_id: '18896'
file_name: 2024_ICML_Cadei.pdf
file_size: 4453014
relation: main_file
success: 1
file_date_updated: 2025-01-27T11:42:24Z
has_accepted_license: '1'
intvolume: ' 38'
language:
- iso: eng
month: '09'
oa: 1
oa_version: Published Version
publication: ICML 2024 Workshop AI4Science
publication_status: published
publisher: Curran Associates
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/CausalLearningAI/ISTAnt
record:
- id: '18895'
relation: research_data
status: public
- id: '19509'
relation: is_continued_by
status: for_moderation
scopus_import: '1'
status: public
title: Smoke and mirrors in causal downstream tasks
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 38
year: '2024'
...
---
OA_place: repository
OA_type: gold
_id: '18895'
abstract:
- lang: eng
text: 'ISTAnt is a new ecological dataset for social immunity and represents the
first real-world benchmark for causal inference downstream tasks on high-dimensional
observations. It analyzes grooming behavior in the ant Lasius neglectus in groups
of three worker ants. The workers for the experiment were obtained from their
laboratory stock colony, which had been collected from the field in 2022 in the
Botanical Garden Jena, Germany. Ant collection and all experimental work were
performed in compliance with international, national and institutional regulations
and ethical guidelines. For the experiment, the body surface of one of the three
ants was treated with a suspension of either of two microparticle types (diameter
~5 µm) to induce grooming by the two nestmates, which were individually color-coded
by application of a dot of blue or orange paint, respectively. The three ants
were housed in small plastic containers (diameter 28mm, height 30mm) with moistened,
plastered ground and the interior walls covered with PTFE (polytetrafluoroethane)
to hamper climbing by the ants. Filming occurred in a temperature- and humidity-controlled
room at 23°C within a custom-made filming box with controlled lighting and ventilation
conditions. We set up nine ant groups at a time (always containing both treatments)
and placed them randomly on positions 1-9 marked on the floor in a 3x3 grid, about
3mm from each other. The experiment was performed on two consecutive days. Videos
were acquired using a USB camera (FLIR blackfly S BFS-U3-120S4C, Teledyne FLIR)
with a high-performance lens (HP Series 25mm Focal Length, Edmund optics 86-572)
in OBS studio 29.0.0 \citep{bailey2017obs} at a framerate of 30 FPS and a resolution
of 2500x2500 pixels. From each original video (105x105 mm), we generated nine
individual videos .mkv (each ~32x32 mm, 770x770 pixels) by determining exact coordinates
per container from one frame in GIMP 2.10.36 and cropping of the videos with FFmpeg
6.1.1. Annotation was performed over two consecutive days by three observers who
had not been involved in the experimental setup or recording and were unaware
of the treatment assignments to ensure bias-free behavioral annotation. They annotated
the behavior of the ants during video observations, using custom-made software
that saves the start and end frames of behaviors marked in a .csv file (see ''annotations''
folder). In one of the videos, one of the nestmates'' legs got inadvertently stuck
to its body surface during the color-coding, interfering with its behavior, so
the video was discarded. This left 44 videos from 5 independent setups (n=24 of
treatment 1 and n=20 of treatment 2) of 10 minutes each for a total of 792 000
annotated frames (see ''video'' folder). For each video, we provide the following
information: the number of the set to which it belongs (1-5); the number of the
position within the set reflecting the position of the ant group under the camera
(1-9), for which we also provide ‘coordinates’ in the 3x3 grid (taking values
-1/0/1 for both X and Y axis); treatment (1 or 2); the hour of the day when the
recording was started (in 24h CEST); experimental day (A or B); the top left coordinate
of the cropping square from the original video (CropX/CropY); the person annotating
the video (given as A, B, C); the date of annotation (1: first day, 2: second
day) and in which order the videos were annotated by each person, both reflecting
a possible training effect of the person (see ''experiments_settings.csv'' file).'
article_processing_charge: No
author:
- first_name: Riccardo
full_name: Cadei, Riccardo
id: 0fa8b76f-72f0-11ef-b75a-a5da96e5ad6b
last_name: Cadei
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Sylvia M
full_name: Cremer, Sylvia M
id: 2F64EC8C-F248-11E8-B48F-1D18A9856A87
last_name: Cremer
orcid: 0000-0002-2193-3868
- first_name: Lukas
full_name: Lindorfer, Lukas
id: 85f0e6d3-06b3-11ec-8982-8c5049fa4455
last_name: Lindorfer
- first_name: Cordelia
full_name: Schmid, Cordelia
last_name: Schmid
citation:
ama: Cadei R, Locatello F, Cremer S, Lindorfer L, Schmid C. ISTAnt. 2024. doi:10.6084/M9.FIGSHARE.26484934.V2
apa: Cadei, R., Locatello, F., Cremer, S., Lindorfer, L., & Schmid, C. (2024).
ISTAnt. Institute of Science and Technology Austria. https://doi.org/10.6084/M9.FIGSHARE.26484934.V2
chicago: Cadei, Riccardo, Francesco Locatello, Sylvia Cremer, Lukas Lindorfer, and
Cordelia Schmid. “ISTAnt.” Institute of Science and Technology Austria, 2024.
https://doi.org/10.6084/M9.FIGSHARE.26484934.V2.
ieee: R. Cadei, F. Locatello, S. Cremer, L. Lindorfer, and C. Schmid, “ISTAnt.”
Institute of Science and Technology Austria, 2024.
ista: Cadei R, Locatello F, Cremer S, Lindorfer L, Schmid C. 2024. ISTAnt, Institute
of Science and Technology Austria, 10.6084/M9.FIGSHARE.26484934.V2.
mla: Cadei, Riccardo, et al. ISTAnt. Institute of Science and Technology
Austria, 2024, doi:10.6084/M9.FIGSHARE.26484934.V2.
short: R. Cadei, F. Locatello, S. Cremer, L. Lindorfer, C. Schmid, (2024).
corr_author: '1'
date_created: 2025-01-27T11:45:43Z
date_published: 2024-10-23T00:00:00Z
date_updated: 2025-01-27T11:58:38Z
day: '23'
ddc:
- '570'
department:
- _id: SyCr
- _id: FrLo
- _id: GradSch
doi: 10.6084/M9.FIGSHARE.26484934.V2
main_file_link:
- open_access: '1'
url: https://10.6084/M9.FIGSHARE.26484934.V2
month: '10'
oa: 1
oa_version: Published Version
publisher: Institute of Science and Technology Austria
related_material:
record:
- id: '18847'
relation: used_in_publication
status: public
status: public
title: ISTAnt
type: research_data_reference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
year: '2024'
...
---
OA_place: repository
OA_type: green
_id: '18964'
abstract:
- lang: eng
text: Object-centric learning (OCL) extracts the representation of objects with
slots, offering an exceptional blend of flexibility and interpretability for abstracting
low-level perceptual features. A widely adopted method within OCL is slot attention,
which utilizes attention mechanisms to iteratively refine slot representations.
However, a major draw-back of most object-centric models, including slot attention,
is their reliance on predefining the number of slots. This not only necessitates
prior knowledge of the dataset but also overlooks the inherent variability in
the number of objects present in each instance. To overcome this fundamental limitation,
we present a novel complexity-aware object auto-encoder framework. Within this
framework, we introduce an adaptive slot attention (AdaSlot) mecha-nism that dynamically
determines the optimal number of slots based on the content of the data. This
is achieved by proposing a discrete slot sampling module that is responsible for
selecting an appropriate number of slots from a candidate list. Furthermore, we
introduce a masked slot decoder that suppresses unselected slots during the decoding
process. Our framework, tested extensively on object discovery tasks with various
datasets, shows performance matching or exceeding top fixed-slot models. Moreover,
our analysis substantiates that our method exhibits the capability to dynamically
adapt the slot number according to each instance's complexity, offering the potential
for further exploration in slot attention research. Project will be available
at https://kfan21.github.io/AdaSlot/
acknowledgement: Yanwei Fu is the corresponding authour. Yanwei Fu is with School
of Data Science, Fudan University, Shanghai Key Lab of Intelligent Information Processing,
Fudan University, and Fudan ISTBI-ZJNU Algorithm Centre for Brain-inspired Intelligence,
Zhejiang Normal University, Jinhua, China.
article_processing_charge: No
arxiv: 1
author:
- first_name: Ke
full_name: Fan, Ke
last_name: Fan
- first_name: Zechen
full_name: Bai, Zechen
last_name: Bai
- first_name: Tianjun
full_name: Xiao, Tianjun
last_name: Xiao
- first_name: Tong
full_name: He, Tong
last_name: He
- first_name: Max
full_name: Horn, Max
last_name: Horn
- first_name: Yanwei
full_name: Fu, Yanwei
last_name: Fu
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Zheng
full_name: Zhang, Zheng
last_name: Zhang
citation:
ama: 'Fan K, Bai Z, Xiao T, et al. Adaptive slot attention: Object discovery with
dynamic slot number. In: 2024 IEEE/CVF Conference on Computer Vision and Pattern
Recognition. IEEE; 2024. doi:10.1109/cvpr52733.2024.02176'
apa: 'Fan, K., Bai, Z., Xiao, T., He, T., Horn, M., Fu, Y., … Zhang, Z. (2024).
Adaptive slot attention: Object discovery with dynamic slot number. In 2024
IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle, WA,
United States: IEEE. https://doi.org/10.1109/cvpr52733.2024.02176'
chicago: 'Fan, Ke, Zechen Bai, Tianjun Xiao, Tong He, Max Horn, Yanwei Fu, Francesco
Locatello, and Zheng Zhang. “Adaptive Slot Attention: Object Discovery with Dynamic
Slot Number.” In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition.
IEEE, 2024. https://doi.org/10.1109/cvpr52733.2024.02176.'
ieee: 'K. Fan et al., “Adaptive slot attention: Object discovery with dynamic
slot number,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition,
Seattle, WA, United States, 2024.'
ista: 'Fan K, Bai Z, Xiao T, He T, Horn M, Fu Y, Locatello F, Zhang Z. 2024. Adaptive
slot attention: Object discovery with dynamic slot number. 2024 IEEE/CVF Conference
on Computer Vision and Pattern Recognition. CVPR: Conference on Computer Vision
and Pattern Recognition.'
mla: 'Fan, Ke, et al. “Adaptive Slot Attention: Object Discovery with Dynamic Slot
Number.” 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition,
IEEE, 2024, doi:10.1109/cvpr52733.2024.02176.'
short: K. Fan, Z. Bai, T. Xiao, T. He, M. Horn, Y. Fu, F. Locatello, Z. Zhang, in:,
2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition, IEEE, 2024.
conference:
end_date: 2024-06-22
location: Seattle, WA, United States
name: 'CVPR: Conference on Computer Vision and Pattern Recognition'
start_date: 2024-06-16
date_created: 2025-01-29T14:27:39Z
date_published: 2024-06-15T00:00:00Z
date_updated: 2025-09-09T12:15:17Z
day: '15'
department:
- _id: FrLo
doi: 10.1109/cvpr52733.2024.02176
external_id:
arxiv:
- '2406.09196'
isi:
- '001342515506043'
isi: 1
language:
- iso: eng
main_file_link:
- open_access: '1'
url: https://doi.org/10.48550/arXiv.2406.09196
month: '06'
oa: 1
oa_version: Preprint
publication: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition
publication_identifier:
eisbn:
- '9798350353006'
publication_status: published
publisher: IEEE
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://kfan21.github.io/AdaSlot/
status: public
title: 'Adaptive slot attention: Object discovery with dynamic slot number'
type: conference
user_id: 317138e5-6ab7-11ef-aa6d-ffef3953e345
year: '2024'
...
---
OA_place: repository
OA_type: green
_id: '18971'
abstract:
- lang: eng
text: 'Models prone to spurious correlations in training data often produce brittle
predictions and introduce unintended biases. Addressing this challenge typically
involves methods relying on prior knowledge and group annotation to remove spurious
correlations, which may not be readily available in many applications. In this
paper, we establish a novel connection between unsupervised object-centric learning
and mitigation of spurious correlations. Instead of directly inferring subgroups
with varying correlations with labels, our approach focuses on discovering concepts:
discrete ideas that are shared across input samples. Leveraging existing object-centric
representation learning, we introduce CoBalT: a concept balancing technique that
effectively mitigates spurious correlations without requiring human labeling of
subgroups. Evaluation across the benchmark datasets for sub-population shifts
demonstrate superior or competitive performance compared state-of-the-art baselines,
without the need for group annotation. Code is available at https://github.com/rarefin/CoBalT'
acknowledgement: "We acknowledge the support of the Canada CIFAR AI Chair Program
and IVADO. We thank Mila and Compute Canada for providing computational resources.\r\n"
alternative_title:
- PMLR
article_processing_charge: No
arxiv: 1
author:
- first_name: Rifat
full_name: Arefin, Rifat
last_name: Arefin
- first_name: Yan
full_name: Zhang, Yan
last_name: Zhang
- first_name: Aristide
full_name: Baratin, Aristide
last_name: Baratin
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Irina
full_name: Rish, Irina
last_name: Rish
- first_name: Dianbo
full_name: Liu, Dianbo
last_name: Liu
- first_name: Kenji
full_name: Kawaguchi, Kenji
last_name: Kawaguchi
citation:
ama: 'Arefin R, Zhang Y, Baratin A, et al. Unsupervised concept discovery mitigates
spurious correlations. In: Proceedings of the 41st International Conference
on Machine Learning. Vol 235. ML Research Press; 2024:1672-1688.'
apa: 'Arefin, R., Zhang, Y., Baratin, A., Locatello, F., Rish, I., Liu, D., &
Kawaguchi, K. (2024). Unsupervised concept discovery mitigates spurious correlations.
In Proceedings of the 41st International Conference on Machine Learning
(Vol. 235, pp. 1672–1688). Vienna, Austria: ML Research Press.'
chicago: Arefin, Rifat, Yan Zhang, Aristide Baratin, Francesco Locatello, Irina
Rish, Dianbo Liu, and Kenji Kawaguchi. “Unsupervised Concept Discovery Mitigates
Spurious Correlations.” In Proceedings of the 41st International Conference
on Machine Learning, 235:1672–88. ML Research Press, 2024.
ieee: R. Arefin et al., “Unsupervised concept discovery mitigates spurious
correlations,” in Proceedings of the 41st International Conference on Machine
Learning, Vienna, Austria, 2024, vol. 235, pp. 1672–1688.
ista: 'Arefin R, Zhang Y, Baratin A, Locatello F, Rish I, Liu D, Kawaguchi K. 2024.
Unsupervised concept discovery mitigates spurious correlations. Proceedings of
the 41st International Conference on Machine Learning. ICML: International Conference
on Machine Learning, PMLR, vol. 235, 1672–1688.'
mla: Arefin, Rifat, et al. “Unsupervised Concept Discovery Mitigates Spurious Correlations.”
Proceedings of the 41st International Conference on Machine Learning, vol.
235, ML Research Press, 2024, pp. 1672–88.
short: R. Arefin, Y. Zhang, A. Baratin, F. Locatello, I. Rish, D. Liu, K. Kawaguchi,
in:, Proceedings of the 41st International Conference on Machine Learning, ML
Research Press, 2024, pp. 1672–1688.
conference:
end_date: 2024-07-27
location: Vienna, Austria
name: 'ICML: International Conference on Machine Learning'
start_date: 2024-07-21
date_created: 2025-01-30T07:21:57Z
date_published: 2024-07-30T00:00:00Z
date_updated: 2025-01-30T07:23:10Z
day: '30'
department:
- _id: FrLo
external_id:
arxiv:
- '2402.13368'
intvolume: ' 235'
language:
- iso: eng
main_file_link:
- open_access: '1'
url: https://doi.org/10.48550/arXiv.2402.13368
month: '07'
oa: 1
oa_version: Preprint
page: 1672-1688
publication: Proceedings of the 41st International Conference on Machine Learning
publication_identifier:
eissn:
- 2640-3498
publication_status: published
publisher: ML Research Press
quality_controlled: '1'
related_material:
link:
- relation: software
url: https://github.com/rarefin/CoBalT
scopus_import: '1'
status: public
title: Unsupervised concept discovery mitigates spurious correlations
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 235
year: '2024'
...
---
OA_place: repository
OA_type: green
_id: '18996'
abstract:
- lang: eng
text: 'We consider the linear causal representation learning setting where we observe
a linear mixing of d unknown latent factors, which follow a linear structural
causal model. Recent work has shown that it is possible to recover the latent
factors as well as the underlying structural causal model over them, up to permutation
and scaling, provided that we have at least d environments, each of which corresponds
to perfect interventions on a single latent node (factor). After this powerful
result, a key open problem faced by the community has been to relax these conditions:
allow for coarser than perfect single-node interventions, and allow for fewer
than d of them, since the number of latent factors d could be very large. In this
work, we consider precisely such a setting, where we allow a smaller than d number
of environments, and also allow for very coarse interventions that can very coarsely
\textit{change the entire causal graph over the latent factors}. On the flip side,
we relax what we wish to extract to simply the \textit{list of nodes that have
shifted between one or more environments}. We provide a surprising identifiability
result that it is indeed possible, under some very mild standard assumptions,
to identify the set of shifted nodes. Our identifiability proof moreover is a
constructive one: we explicitly provide necessary and sufficient conditions for
a node to be a shifted node, and show that we can check these conditions given
observed data. Our algorithm lends itself very naturally to the sample setting
where instead of just interventional distributions, we are provided datasets of
samples from each of these distributions. We corroborate our results on both synthetic
experiments as well as an interesting psychometric dataset. The code can be found
at https://github.com/TianyuCodings/iLCS.'
alternative_title:
- Advances in Neural Information Processing Systems
article_processing_charge: No
arxiv: 1
author:
- first_name: Tianyu
full_name: Chen, Tianyu
last_name: Chen
- first_name: Kevin
full_name: Bello, Kevin
last_name: Bello
- first_name: Francesco
full_name: Locatello, Francesco
id: 26cfd52f-2483-11ee-8040-88983bcc06d4
last_name: Locatello
orcid: 0000-0002-4850-0683
- first_name: Bryon
full_name: Aragam, Bryon
last_name: Aragam
- first_name: Pradeep Kumar
full_name: Ravikumar, Pradeep Kumar
last_name: Ravikumar
citation:
ama: 'Chen T, Bello K, Locatello F, Aragam B, Ravikumar PK. Identifying general
mechanism shifts in linear causal representations. In: 38th Conference on Neural
Information Processing Systems. Vol 37. Neural Information Processing Systems
Foundation; 2024.'
apa: 'Chen, T., Bello, K., Locatello, F., Aragam, B., & Ravikumar, P. K. (2024).
Identifying general mechanism shifts in linear causal representations. In 38th
Conference on Neural Information Processing Systems (Vol. 37). Vancouver,
Canada: Neural Information Processing Systems Foundation.'
chicago: Chen, Tianyu, Kevin Bello, Francesco Locatello, Bryon Aragam, and Pradeep
Kumar Ravikumar. “Identifying General Mechanism Shifts in Linear Causal Representations.”
In 38th Conference on Neural Information Processing Systems, Vol. 37. Neural
Information Processing Systems Foundation, 2024.
ieee: T. Chen, K. Bello, F. Locatello, B. Aragam, and P. K. Ravikumar, “Identifying
general mechanism shifts in linear causal representations,” in 38th Conference
on Neural Information Processing Systems, Vancouver, Canada, 2024, vol. 37.
ista: 'Chen T, Bello K, Locatello F, Aragam B, Ravikumar PK. 2024. Identifying general
mechanism shifts in linear causal representations. 38th Conference on Neural Information
Processing Systems. NeurIPS: Neural Information Processing Systems, Advances in
Neural Information Processing Systems, vol. 37.'
mla: Chen, Tianyu, et al. “Identifying General Mechanism Shifts in Linear Causal
Representations.” 38th Conference on Neural Information Processing Systems,
vol. 37, Neural Information Processing Systems Foundation, 2024.
short: T. Chen, K. Bello, F. Locatello, B. Aragam, P.K. Ravikumar, in:, 38th Conference
on Neural Information Processing Systems, Neural Information Processing Systems
Foundation, 2024.
conference:
end_date: 2024-12-16
location: Vancouver, Canada
name: 'NeurIPS: Neural Information Processing Systems'
start_date: 2024-12-16
date_created: 2025-02-04T13:09:34Z
date_published: 2024-09-25T00:00:00Z
date_updated: 2025-07-07T13:23:49Z
day: '25'
ddc:
- '000'
department:
- _id: FrLo
external_id:
arxiv:
- '2410.24059'
file:
- access_level: open_access
checksum: 75c3091e70bd2916cd94afbf40a0c425
content_type: application/pdf
creator: dernst
date_created: 2025-02-04T13:09:08Z
date_updated: 2025-02-04T13:09:08Z
file_id: '18997'
file_name: 2024_NeurIPS_Chen.pdf
file_size: 5659119
relation: main_file
success: 1
file_date_updated: 2025-02-04T13:09:08Z
has_accepted_license: '1'
intvolume: ' 37'
language:
- iso: eng
month: '09'
oa: 1
oa_version: Published Version
publication: 38th Conference on Neural Information Processing Systems
publication_identifier:
eissn:
- 1049-5258
publication_status: published
publisher: Neural Information Processing Systems Foundation
quality_controlled: '1'
scopus_import: '1'
status: public
title: Identifying general mechanism shifts in linear causal representations
tmp:
image: /images/cc_by.png
legal_code_url: https://creativecommons.org/licenses/by/4.0/legalcode
name: Creative Commons Attribution 4.0 International Public License (CC-BY 4.0)
short: CC BY (4.0)
type: conference
user_id: 2DF688A6-F248-11E8-B48F-1D18A9856A87
volume: 37
year: '2024'
...