Bayesian Metric Learning for Uncertainty Quantification in Image Retrieval

Authors: Frederik Warburg, Marco Miani, Silas Brack, Søren Hauberg

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, we show that our Laplacian Metric Learner (LAM) yields well-calibrated uncertainties, reliably detects out-of-distribution examples, and has state-of-the-art predictive performance.
Researcher Affiliation Academia Frederik Warburg Technical University of Denmark frwa@dtu.dk Marco Miani Technical University of Denmark mmia@dtu.dk Silas Brack Technical University of Denmark silasbrack@gmail.com Søren Hauberg Technical University of Denmark sohau@dtu.dk
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes We extend Stoch Man [11] with the Hessian backpropagation for the contrastive loss, and the Py Torch [34] code is publicly available2. 2See https://github.com/Frederik Warburg/bayesian-metric-learning
Open Datasets Yes We evaluate OOD performance on the commonly used benchmarks [31], where we use (1) Fashion MNIST [49] as ID and MNIST [26] as OOD, and (2) CIFAR10 [21] as ID and SVHN [32] as OOD. ... We first test with CUB200 [45] as ID and CAR196 [20] as OOD ... and second, test with LFW [17] as ID and CUB200 as OOD. We evaluate on MSLS [46]
Dataset Splits Yes We use the standard train/test split, training on 24 cities and testing on six other cities. We follow the zero-shot train/test split [30].
Hardware Specification Yes The last-layer LAM training time is 3 hours for online LAM vs 2.3 hours for deterministic contrastive loss on LFW, and 30 minutes vs 15 minutes loss on CUB200 on an NVIDIA RTX A5000.
Software Dependencies No The paper mentions 'Py Torch [34]' but does not specify a version number for PyTorch or any other software dependency.
Experiment Setup Yes Table 4: Ablation on LFW shows parameters like 'Margin', 'Number of pairs per batch', 'Latent dimension', 'Memory factor α', 'Post-hoc tempering β' with specific values used in the experiments.