Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval

Frederik Warburg; Marco Miani; Silas Brack; Søren Hauberg

Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval

Frederik Warburg
, Marco Miani
, Silas Brack
, Søren Hauberg

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

160 Downloads (Orbit)

Abstract

We propose the first Bayesian encoder for metric learning. Rather than relying on neural amortization as done in prior works, we learn a distribution over the network weights with the Laplace Approximation. We actualize this by first proving that the contrastive loss is a valid log-posterior. We then propose three methods that ensure a positive definite Hessian. Lastly, we present a novel decomposition of the Generalized Gauss-Newton approximation. Empirically, we show that our Laplacian Metric Learner (LAM) estimates wellcalibrated uncertainties, reliably detects out-ofdistribution examples, and yields state-of-the-art
predictive performance.

Original language	English
Title of host publication	Proceedings of the 37^th Conference on Neural Information Processing Systems
Number of pages	26
Volume	36
Publisher	Neural Information Processing Systems Foundation
Publication date	2023
Pages	69178-69190
Publication status	Published - 2023
Event	37^th Annual Conference on Neural Information Processing Systems - Ernest N. Morial Convention Center, New Orleans, United States Duration: 10 Dec 2023 → 16 Dec 2023 Conference number: 37

Conference

Conference	37^th Annual Conference on Neural Information Processing Systems
Number	37
Location	Ernest N. Morial Convention Center
Country/Territory	United States
City	New Orleans
Period	10/12/2023 → 16/12/2023

Access to Document

FulltextAccepted author manuscript, 13.2 MB

https://proceedings.neurips.cc/paper_files/paper/2023/file/da7ce04b3683b173691ecbb801f2690f-Paper-Conference.pdf

OpenUrl availability

Check availability in DTU Findit

Cite this

Warburg, F., Miani, M., Brack, S., & Hauberg, S. (2023). Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval. In Proceedings of the 37^th Conference on Neural Information Processing Systems (Vol. 36, pp. 69178-69190). Neural Information Processing Systems Foundation. https://proceedings.neurips.cc/paper_files/paper/2023/file/da7ce04b3683b173691ecbb801f2690f-Paper-Conference.pdf

@inproceedings{86ed7d070c3648529e520e98cc2f4d7d,

title = "Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval",

abstract = "We propose the first Bayesian encoder for metric learning. Rather than relying on neural amortization as done in prior works, we learn a distribution over the network weights with the Laplace Approximation. We actualize this by first proving that the contrastive loss is a valid log-posterior. We then propose three methods that ensure a positive definite Hessian. Lastly, we present a novel decomposition of the Generalized Gauss-Newton approximation. Empirically, we show that our Laplacian Metric Learner (LAM) estimates wellcalibrated uncertainties, reliably detects out-ofdistribution examples, and yields state-of-the-art predictive performance.",

author = "Frederik Warburg and Marco Miani and Silas Brack and S{\o}ren Hauberg",

year = "2023",

language = "English",

volume = "36",

pages = "69178--69190",

booktitle = "Proceedings of the 37th Conference on Neural Information Processing Systems",

publisher = "Neural Information Processing Systems Foundation",

note = "37<sup>th</sup> Annual Conference on Neural Information Processing Systems, NeurIPS 2023 ; Conference date: 10-12-2023 Through 16-12-2023",

}

Warburg, F, Miani, M, Brack, S & Hauberg, S 2023, Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval. in Proceedings of the 37^th Conference on Neural Information Processing Systems. vol. 36, Neural Information Processing Systems Foundation, pp. 69178-69190, 37^th Annual Conference on Neural Information Processing Systems, New Orleans, Louisiana, United States, 10/12/2023. <https://proceedings.neurips.cc/paper_files/paper/2023/file/da7ce04b3683b173691ecbb801f2690f-Paper-Conference.pdf>

Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval. / Warburg, Frederik; Miani, Marco; Brack, Silas et al.
Proceedings of the 37^th Conference on Neural Information Processing Systems. Vol. 36 Neural Information Processing Systems Foundation, 2023. p. 69178-69190.

Research output: Chapter in Book/Report/Conference proceeding › Article in proceedings › Research › peer-review

TY - GEN

T1 - Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval

AU - Warburg, Frederik

AU - Miani, Marco

AU - Brack, Silas

AU - Hauberg, Søren

N1 - Conference code: 37

PY - 2023

Y1 - 2023

N2 - We propose the first Bayesian encoder for metric learning. Rather than relying on neural amortization as done in prior works, we learn a distribution over the network weights with the Laplace Approximation. We actualize this by first proving that the contrastive loss is a valid log-posterior. We then propose three methods that ensure a positive definite Hessian. Lastly, we present a novel decomposition of the Generalized Gauss-Newton approximation. Empirically, we show that our Laplacian Metric Learner (LAM) estimates wellcalibrated uncertainties, reliably detects out-ofdistribution examples, and yields state-of-the-art predictive performance.

AB - We propose the first Bayesian encoder for metric learning. Rather than relying on neural amortization as done in prior works, we learn a distribution over the network weights with the Laplace Approximation. We actualize this by first proving that the contrastive loss is a valid log-posterior. We then propose three methods that ensure a positive definite Hessian. Lastly, we present a novel decomposition of the Generalized Gauss-Newton approximation. Empirically, we show that our Laplacian Metric Learner (LAM) estimates wellcalibrated uncertainties, reliably detects out-ofdistribution examples, and yields state-of-the-art predictive performance.

M3 - Article in proceedings

VL - 36

SP - 69178

EP - 69190

BT - Proceedings of the 37th Conference on Neural Information Processing Systems

PB - Neural Information Processing Systems Foundation

T2 - 37<sup>th</sup> Annual Conference on Neural Information Processing Systems

Y2 - 10 December 2023 through 16 December 2023

ER -

Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval

Abstract

Conference

Access to Document

OpenUrl availability

Fingerprint

Cite this