针对数据不平衡的自监督学习，提出多图像多视图框架提升表征质量。
Maximally Useful and Minimally Redundant: The Key to Self Supervised Learning for Imbalanced Data
- 引入多图像多视图输入，突破传统多视图假设限制
- 设计可逆与不可逆增强结合策略，保留完整图像信息
- 基于信息瓶颈原理优化表征，显著提升不平衡数据表现
对比式自监督学习（CSSL）通常依赖多视图假设，即所有相关信息必须在所有视图间共享。其目标是最大化不同视图表示间的互信息，同时压缩每个表示中的无关信息。近期，Schwartz Ziv与Yan LeCun指出，当多视图假设不成立时，如何基于替代假设分离相关信息与无关信息，是自监督学习的关键挑战。本文据此提出以下贡献：1）构建多图像多视图（MIMV）输入框架，突破传统多视图限制；2）采用新型增强策略，结合可逆（可逆）与非可逆（硬增强）视图，确保单张图像信息完整保留，同时对另一图像使用强增强；3）为MIMV框架提出信息瓶颈（IB）原则以生成最优表示；4）设计新损失函数，通过过滤极端特征提升表示能力；5）在数据不平衡场景下验证框架鲁棒性，在Cifar10-LT（ResNet-18）上提升2%，Cifar100-LT（ResNet-18）提升5%，Imagenet-LT（1k，ResNet-50）提升3%，达到新最先进水平。
原文摘要 · Abstract (English)
Contrastive self supervised learning(CSSL) usually makes use of the multi-view assumption which states that all relevant information must be shared between all views. The main objective of CSSL is to maximize the mutual information(MI) between representations of different views and at the same time compress irrelevant information in each representation. Recently, as part of future work, Schwartz Ziv & Yan LeCun pointed out that, when the multi-view assumption is violated, one of the most significant challenges in SSL is in identifying new methods to separate relevant from irrelevant information based on alternative assumptions. Taking a cue from this intuition we make the following contributions in this paper: 1) We develop a CSSL framework wherein multiple images and multiple views(MIMV) are considered as input, which is different from the traditional multi-view assumption 2) We adopt a novel augmentation strategy that includes both normalized (invertible) and augmented (non-invertible) views so that complete information of one image can be preserved and hard augmentation can be chosen for the other image 3) An Information bottleneck(IB) principle is outlined for MIMV to produce optimal representations 4) We introduce a loss function that helps to learn better representations by filtering out extreme features 5) The robustness of our proposed framework is established by applying it to the imbalanced dataset problem wherein we achieve a new state-of-the-art accuracy (2% improvement in Cifar10-LT using Resnet-18, 5% improvement in Cifar100-LT using Resnet-18 and 3% improvement in Imagenet-LT (1k) using Resnet-50).
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。