Long-tailed Recognition by Routing Diverse Distribution-Aware Experts (2020-10-05T00:00:00.000000Z)

TL;DR

RIDE aims to reduce both the bias and the variance of a long-tailed classifier by RoutIng Diverse Experts (RIDE), which significantly outperforms the state-of-the-art methods by 5% to 7% on all the benchmarks including CIFAR100-LT, ImageNet-LT and iNaturalist.

Abstract

Natural data are often long-tail distributed over semantic classes. Existing recognition methods tend to focus on tail performance gain, often at the expense of head performance loss from increased classifier variance. The low tail performance manifests itself in large inter-class confusion and high classifier variance. We aim to reduce both the bias and the variance of a long-tailed classifier by RoutIng Diverse Experts (RIDE). It has three components: 1) a shared architecture for multiple classifiers (experts); 2) a distribution-aware diversity loss that encourages more diverse decisions for classes with fewer training instances; and 3) an expert routing module that dynamically assigns more ambiguous instances to additional experts. With on-par computational complexity, RIDE significantly outperforms the state-of-the-art methods by 5% to 7% on all the benchmarks including CIFAR100-LT, ImageNet-LT and iNaturalist. RIDE is also a universal framework that can be applied to different backbone networks and integrated into various long-tailed algorithms and training mechanisms for consistent performance gains.

Authors

Ziwei Liu

27 papers

Zhongqi Miao

2 papers

Stella X. Yu

5 papers

TL;DR

Abstract

Authors

References53 items

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Feature Space Augmentation for Long-Tailed Data

Distribution-Balanced Loss for Multi-Label Classification in Long-Tailed Datasets

Inflated Episodic Memory With Region Self-Attention for Long-Tailed Visual Recognition

M2m: Imbalanced Classification via Major-to-Minor Translation

Rethinking Class-Balanced Methods for Long-Tailed Visual Recognition From a Domain Adaptation Perspective

Equalization Loss for Long-Tailed Object Recognition

Deep Representation Learning on Long-Tailed Data: A Learnable Embedding Augmentation Perspective

Learning From Multiple Experts: Self-paced Knowledge Distillation for Long-tailed Classification

BBN: Bilateral-Branch Network With Cumulative Learning for Long-Tailed Visual Recognition

Contrastive Representation Distillation

Similarity-Preserving Knowledge Distillation

Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss

Image Deformation Meta-Networks for One-Shot Learning

Large-Scale Long-Tailed Recognition in an Open World

Few-Shot Learning via Saliency-Guided Hallucination of Samples

Class-Balanced Loss Based on Effective Number of Samples

Delta-encoder: an effective sample synthesis method for few-shot object recognition

Unsupervised Feature Learning via Non-parametric Instance Discrimination

Learning Deep Representations with Probabilistic Knowledge Transfer

Low-Shot Learning from Imaginary Data

Feature Space Transfer for Data Augmentation

Learning to Compare: Relation Network for Few-Shot Learning

Focal Loss for Dense Object Detection

The iNaturalist Species Classification and Detection Dataset

Prototypical Networks for Few-shot Learning

Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

Range Loss for Deep Face Recognition with Long-Tailed Training Data

Aggregated Residual Transformations for Deep Neural Networks

Optimization as a Model for Few-Shot Learning

Meta-Learning with Memory-Augmented Neural Networks

Learning feed-forward one-shot learners

Matching Networks for One Shot Learning

Low-Shot Visual Recognition by Shrinking and Hallucinating Features

Deep Residual Learning for Image Recognition

Deep Metric Learning via Lifted Structured Feature Embedding

Distilling the Knowledge in a Neural Network

ImageNet: A large-scale hierarchical image database

Dimensionality Reduction by Learning an Invariant Mapping

Cryptosporidium Parvum Genome Project

Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

Learning From Imbalanced Data

2019), and compared the differences in Table 5. Although adding other methods along with

Learning to Model the Tail

We apply distillation from a more powerful model with more experts into a model with fewer experts. A simple way to transfer knowledge is knowledge distillation (KD

Learning Multiple Layers of Features from Tiny Images

ImageNet-LT is constructed by sampling a subset of ImageNet-2012 following the Pareto distribution with the power value α = 6

A unified bias-variance decomposition

Stage Two: Routing Diverse Experts

A.2 ADDITIONAL EXPERIMENTS

Our method also achieves lower computational cost and higher performance when compared to other methods

Since our method requires the joint decision from several experts, which raw ensembles also do, we also compare against ensembles of LDAM in Fig

On the model bias: All the long-tail methods reduce the overall, medium-shot, and few-shot bias. The reduction tends to be greater for the tail classes

Field of Study

Journal Information

Name

Volume

Venue Information

Name

Type

URL

Alternate Names