Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval
This addresses uncertainty quantification for image retrieval systems, which is incremental as it builds on prior metric learning methods.
The authors tackled the problem of uncertainty quantification in image retrieval by proposing the first Bayesian encoder for metric learning, which estimates well-calibrated uncertainties and achieves state-of-the-art predictive performance.
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.