Long-tailed Recognition by Routing Diverse Distribution-Aware Experts

TL;DR

This paper introduces RIDE, a novel long-tailed recognition method that employs diverse experts and dynamic routing to improve accuracy across all classes, addressing bias and variance issues inherent in existing approaches.

Contribution

The paper proposes RIDE, a new classifier that reduces bias and variance in long-tail recognition by using multiple experts, a diversity loss, and dynamic routing, achieving state-of-the-art results.

Findings

RIDE outperforms existing methods by 5-7% on benchmark datasets.

It effectively reduces bias and variance in long-tail classification.

The framework is versatile across different networks and training setups.

Abstract

Natural data are often long-tail distributed over semantic classes. Existing recognition methods tackle this imbalanced classification by placing more emphasis on the tail data, through class re-balancing/re-weighting or ensembling over different data groups, resulting in increased tail accuracies but reduced head accuracies. We take a dynamic view of the training data and provide a principled model bias and variance analysis as the training data fluctuates: Existing long-tail classifiers invariably increase the model variance and the head-tail model bias gap remains large, due to more and larger confusion with hard negatives for the tail. We propose a new long-tailed classifier called RoutIng Diverse Experts (RIDE). It reduces the model variance with multiple experts, reduces the model bias with a distribution-aware diversity loss, reduces the computational cost with a dynamic expert routing module. RIDE outperforms the state-of-the-art by 5% to 7% on CIFAR100-LT, ImageNet-LT and iNaturalist 2018 benchmarks. It is also a universal framework that is applicable to various backbone networks, long-tailed algorithms, and training mechanisms for consistent performance gains. Our code is available at: https://github.com/frank-xwang/RIDE-LongTailRecognition.