Maximization Bias

Maximization bias, the tendency for predicted maximum values to overestimate true maxima, is a significant problem across diverse machine learning applications. Current research focuses on understanding and mitigating this bias in various contexts, including image generation (using novel training paradigms based on gradient descent properties), classification (through algorithms designed to achieve exponential margin maximization), and reinforcement learning (employing two-sample testing-based estimators to control overestimation). Addressing maximization bias is crucial for improving the accuracy and reliability of machine learning models in areas such as advertising recommendation systems and robotics, where accurate estimations of maximum values are critical for optimal performance.

Papers

December 5, 2023

Generator Born from Classifier
Runpeng Yu, Xinchao Wang
Gradient Descent Simple Classifier Synchronous Generator Image Generator Image Generation Task Pre Trained Classifier Maximization Bias

November 24, 2023

Achieving Margin Maximization Exponentially Fast via Progressive Norm Rescaling
Mingze Wang, Zeping Min, Lei Wu
Gradient Descent Margin Maximization Gradient Based Algorithm Maximization Bias

October 7, 2022

The Asymmetric Maximum Margin Bias of Quasi-Homogeneous Neural Networks
Daniel Kunin, Atsushi Yamamura, Chao Ma, Surya Ganguli
Gradient Flow Homogeneous Neural Network Advanced Activation Mechanism Gradient Dynamic Maximization Bias

May 19, 2022

Calibration Matters: Tackling Maximization Bias in Large-scale Advertising Recommendation Systems
Yewen Fan, Nian Si, Kun Zhang
Recommendation System Optimal Calibration Sample Specific Debiasing Maximization Bias

January 20, 2022

Addressing Maximization Bias in Reinforcement Learning with Two-Sample Testing
Martin Waltz, Ostap Okhrin
Reinforcement Learning Mean Estimation Q$ Learning Estimation Bias Two Sample Value Based Reinforcement Learning Overestimation Bias Maximization Bias

Maximization Bias

Papers

Generator Born from Classifier

Achieving Margin Maximization Exponentially Fast via Progressive Norm Rescaling

The Asymmetric Maximum Margin Bias of Quasi-Homogeneous Neural Networks

Calibration Matters: Tackling Maximization Bias in Large-scale Advertising Recommendation Systems

Addressing Maximization Bias in Reinforcement Learning with Two-Sample Testing