Enabling Uncertainty Estimation in Iterative Neural Networks

Authors: Nikita Durasov, Doruk Oner, Jonathan Donier, Hieu Le, Pascal Fua

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate its practical value by embedding it in two application domains: road detection in aerial images and the estimation of aerodynamic properties of 2D and 3D shapes. Our approach applies to both classification and regression. To demonstrate the first, we use it for road delineation purposes, that is, classifying pixels in aerial images as belonging to roads or not. To demonstrate the second, we use it to assess the reliability of performance numbers -drag for cars and lift-to-drag for airfoils predicted by networks given 2D and 3D shapes as input. We then use these reliability estimates to implement a Bayesian optimization scheme that enables us to refine the shapes for improved performance. We report accuracy and uncertainty results in Tabs. 1 and 2.
Researcher Affiliation Collaboration 1Computer Vision Laboratory, Ecole Polytechnique F ed erale de Lausanne, Lausanne, Switzerland 2Neural Concept SA, Lausanne, Switzerland.
Pseudocode No The paper describes steps for its methods (e.g., in Section 3.2 or Appendix A.2) but does not provide any structured pseudocode or algorithm blocks. These descriptions are presented as narrative text rather than formatted code.
Open Source Code Yes The code will be made publicly available.
Open Datasets Yes Datasets. We experimented on two publicly available datasets. Road Tracer. It comprises high-resolution satellite images covering urban areas of forty cities in six different countries (Bastani et al., 2018). Fifteen cities are set aside for validation purposes. The ground truth was generated using Open Street Map. Massachusetts. The Massachusetts dataset features both urban and rural neighborhoods, with many different kinds of roads ranging from small paths to highways. We used the same splits as in (Hu et al., 2019).
Dataset Splits Yes Fifteen cities are set aside for validation purposes. We used the same splits as in (Hu et al., 2019).
Hardware Specification Yes For our experiments, we used single Tesla V100 GPU with 32Gb of memory.
Software Dependencies No The paper mentions software frameworks like "Pytorch (Paszke et al., 2017) and Pytorch Geometrics (Fey & Lenssen, 2019)" but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes For training, we use Adam optimizer (Kingma & Ba, 2015) and perform 200 epochs with 128 batch size and 0.001 learning rate. Both for lift-to-drag and pressure, we use mean squared error (MSE) loss and combine them into final loss with weights 1 for scalar and 100 for pressure. For cars dataset... For training, we use Adam optimizer and perform 6 epochs with 8 batch size and 0.001 learning rate. Both for lift-to-drag and pressure, we use mean squared error (MSE) loss and combine them into final loss with weights 1 for scalar and 1/200 for pressure.