Core Concepts
Iterative neural networks offer a practical approach to uncertainty estimation, providing state-of-the-art estimates at a lower computational cost.
Abstract
Abstract: Proposes using convergence rate as a proxy for uncertainty estimation.
Introduction: Discusses the benefits of recursive refinement in deep networks.
Data Extraction:
"Turning pass-through network architectures into iterative ones is a well-known approach for boosting performance."
"Convergence rate of successive outputs correlates with the accuracy of the value they converge to."
Experiments:
Demonstrates effectiveness in road detection and aerodynamic properties estimation.
Outperforms Deep Ensembles and MC-Dropout in both classification and regression tasks.
Method:
Utilizes variance in outputs from iterative models for uncertainty estimation.
Related Work:
Compares Uncertainty Estimation methods like Deep Ensembles, MC-Dropout, and Bayesian Networks.
Aerodynamics Prediction:
Implements Bayesian optimization for shape optimization with aerodynamic properties.
Stats
"Turning pass-through network architectures into iterative ones is a well-known approach for boosting performance."
"Convergence rate of successive outputs correlates with the accuracy of the value they converge to."