Core Concepts
Active learning strategy using Wasserstein distance and GroupSort neural networks for regression problems.
Abstract
The paper introduces a new active learning strategy for regression using the Wasserstein distance and GroupSort neural networks. It focuses on distribution matching, uncertainty-based sampling, and diversity to improve estimation accuracy. The study compares this method with other classical and recent solutions, showing its effectiveness in achieving precise estimations faster.
Introduction
Challenges in data collection for machine learning.
Importance of active learning in reducing labeling costs.
Active Learning Framework
Estimating unknown functions with labeled and unlabeled data subsets.
Utilizing an estimator belonging to a class of neural networks.
Wasserstein Distance
Definition of the Wasserstein distance for probability measures on metric spaces.
GroupSort Neural Networks
Introduction to GroupSort activation function and neural network architecture.
Theoretical Foundations
Assumptions about mathematical background for the approach.
Training the Estimator
Minimizing error risk with Lipschitz functions and loss function minimization.
Minimizing Uncertainty and Query Procedure
Construction of score function based on Wasserstein distance and uncertainty-based method.
Numerical Experiments
Comparison of different models on various datasets using RMSE metrics.
Stats
25%のデータがラベル付けされたときのRMSEは、WARがBostonで3.63、Airfoilで8.67、Energy1で2.65、Yachtで2.71、Concreteで6.15です。