Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval

TL;DR

This paper introduces a Bayesian encoder for metric learning using Laplace Approximation, enabling uncertainty quantification, out-of-distribution detection, and improved image retrieval performance.

Contribution

It presents the first Bayesian approach for metric learning with a novel Laplace Approximation method, enhancing uncertainty estimation and out-of-distribution detection.

Findings

Accurately estimates calibrated uncertainties

Effectively detects out-of-distribution examples

Achieves state-of-the-art predictive performance

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 well-calibrated uncertainties, reliably detects out-of-distribution examples, and yields state-of-the-art predictive performance.