论文阅读:Trustworthy Long-Tailed Classification CVPR2022

论文:[2111.09030] Trustworthy Long-Tailed Classification (arxiv.org)

code:GitHub - lblaoke/TLC: PyTorch implementation of the paper "Trustworthy Long-Tailed Classification" (CVPR 2022)

Abstract

首先提出了目前集成学习处理长尾问题的两个限制:首先,对于failure-sensitive applications预测是不可信的(尾部的预测不够准确);其次,对于简单样本使用多专家预测是冗余的。基于此,作者提出了一种可信长尾分类(TLC)方法,在多专家框架下联合进行分类不确定性评估,以识别硬样本。

该方法首先会获取每一位专家网络的证据基于证据的不确定度(EvU),然后会基于Dempster-Shafer Evidence Theory (DST)理论融合这些证据和不确定度。此外作者还提出了一个动态专家参与策略,以减少预测简单样本的专家数量。

Introduction

现有的解决长尾问题的方法主要是:重新训练数据并赋予尾部更大的权重、在头部和尾部之间传递知识、互补集成(多个teacher网络蒸馏出一个student网络)和冗余集成(多专家共同决策)。冗余集成效果最好,主要是通过降低模型的方差来获取稳健的预测。但主要有两个局限性:首先是容易过度自信,这也防止集合方法感知错误的预测或者ODD样本。由于尾类的预测更多,对尾类更有害。其次是冗余集成通常假设所有分类器都应该在所有样本上进行训练,这是静态的,并且会通过将专家分配给所有类别从而增加计算成本

对于这些问题,作者提出了一种新的可信长尾分类方法(TLC)在统一的框架内联合进行分类和不确定性估计。首先,作者介绍了DST下的证据及相关的不确定性。在EvU的帮助下,模型可以感知困难样本提升对尾类和OOD样本的可信度并识别潜在的错误预测。其次,作者在Dempster规则下将不同专家的evidence和基于不确定性的多专家融合策略结合,获得准确的不确定性和稳健的预测。同时作者提出动态减少expert的数量以提高效率(头类需要较少的expert)。贡献如下:

1)引入EvU提高长尾分类的可信度,这是第一个在长尾分类中主张可信度的工作。

2)提出了一种基于DST的每个专家的不确定性的多专家融合策略,可以通过可靠的感知困难样本提升分类可信度。

3)通过动态减少不确定专家的参与提升了效率,达到了SOTA。

Related Work

Long-tailed classification

传统的长尾分类方法包括欠采样、过采样和数据增强。重新平衡方法更关注尾部类。OLTR 和inflated memory在不同的类别区域之间传递知识。BBN 分别学习头部和尾部的模式。LFME 分别提取类别区域的多个教师模型。RIDE和 ACE 集合了多位专家以获得稳健的预测。TDE 采用随机推理来消除尾类的偏差。这些方法并没有充分探索在预测中感知硬样本的不确定性。

Uncertainty estimation

Evidence theory.

Preliminaries

长尾数据集由不平衡训练集和平衡测试集组成。数据集可以按照类别数量的不同分为头 中 尾 三个部分。

对于测试集,类频率是相等的。

Method

Estimating Evidence-based Uncertainty

Combining Experts with Dempster's Rule

Combining uncertainties

Combining evidences.

Learning Experts with Dynamic Engagement

  • 0
    点赞
  • 1
    收藏
    觉得还不错? 一键收藏
  • 打赏
    打赏
  • 0
    评论
Casola, V., & Castiglione, A. (2020). Secure and Trustworthy Big Data Storage. Springer. Corriveau, D., Gerrish, B., & Wu, Z. (2020). End-to-end Encryption on the Server: The Why and the How. arXiv preprint arXiv:2010.01403. Dowsley, R., Nascimento, A. C. A., & Nita, D. M. (2021). Private database access using homomorphic encryption. Journal of Network and Computer Applications, 181, 103055. Hossain, M. A., Fotouhi, R., & Hasan, R. (2019). Towards a big data storage security framework for the cloud. In Proceedings of the 9th Annual Computing and Communication Workshop and Conference (CCWC), Las Vegas, USA (pp. 402-408). Rughani, R. (2019). Analysis of Security Issues and Their Solutions in Cloud Storage Environment. International Journal of Computer Trends and Technology (IJCTT), 67(6), 37-42. van Esbroeck, A. (2019). Zero-Knowledge Proofs in the Age of Cryptography: Preventing Fraud Without Compromising Privacy. Chicago-Kent Journal of Intellectual Property, 19, 374. Berman, L. (2021). Watch out for hidden cloud costs. CFO Dive. Retrieved from https://www.cfodive.com/news/watch-out-for-hidden-cloud-costs/603921/ Bradley, T. (2021). Cloud storage costs continue to trend downward. Forbes. Retrieved from https://www.forbes.com/sites/tonybradley/2021/08/27/cloud-storage-costs-continue-to-trend-downward/?sh=6f9d6ade7978 Cisco. (2019). Cost optimization in the multicloud. Cisco. Retrieved from https://www.cisco.com/c/dam/en/us/solutions/collateral/data-center-virtualization/cloud-cost-optimization/cost-optimization_in_multicloud.pdf IBM. (2020). Storage efficiency solutions. IBM. Retrieved from https://www.ibm.com/blogs/systems/storage-efficiency-solutions/ Microsoft Azure. (n.d.). Azure Blob storage tiers. Microsoft Azure. Retrieved from https://azure.microsoft.com/en-us/services/storage/blobs/#pricing Nawrocki, M. (2019). The benefits of a hybrid cloud strategy for businesses. DataCenterNews. Retrieved from https://datacenternews.asia/story/the-benefits-of-a-hybrid-cloud-strategy-for,请把这一段reference list改为标准哈佛格式
05-29
Casola, V. & Castiglione, A. (2020) 'Secure and Trustworthy Big Data Storage', Springer. Corriveau, D., Gerrish, B. & Wu, Z. (2020) 'End-to-end Encryption on the Server: The Why and the How', arXiv preprint arXiv:2010.01403. Dowsley, R., Nascimento, A. C. A. & Nita, D. M. (2021) 'Private database access using homomorphic encryption', Journal of Network and Computer Applications, 181, p.103055. Hossain, M. A., Fotouhi, R. & Hasan, R. (2019) 'Towards a big data storage security framework for the cloud', in Proceedings of the 9th Annual Computing and Communication Workshop and Conference (CCWC), Las Vegas, USA, pp. 402-408. Rughani, R. (2019) 'Analysis of Security Issues and Their Solutions in Cloud Storage Environment', International Journal of Computer Trends and Technology (IJCTT), 67(6), pp. 37-42. van Esbroeck, A. (2019) 'Zero-Knowledge Proofs in the Age of Cryptography: Preventing Fraud Without Compromising Privacy', Chicago-Kent Journal of Intellectual Property, 19, p.374. Berman, L. (2021) 'Watch out for hidden cloud costs', CFO Dive. [online] Available at: https://www.cfodive.com/news/watch-out-for-hidden-cloud-costs/603921/ (Accessed: 5 October 2021). Bradley, T. (2021) 'Cloud storage costs continue to trend downward', Forbes. [online] Available at: https://www.forbes.com/sites/tonybradley/2021/08/27/cloud-storage-costs-continue-to-trend-downward/?sh=6f9d6ade7978 (Accessed: 5 October 2021). Cisco. (2019) 'Cost optimization in the multicloud', Cisco. [online] Available at: https://www.cisco.com/c/dam/en/us/solutions/collateral/data-center-virtualization/cloud-cost-optimization/cost-optimization_in_multicloud.pdf (Accessed: 5 October 2021). IBM. (2020) 'Storage efficiency solutions', IBM. [online] Available at: https://www.ibm.com/blogs/systems/storage-efficiency-solutions/ (Accessed: 5 October 2021). Microsoft Azure. (n.d.) 'Azure Blob storage tiers', Microsoft Azure. [online] Available at: https://azure.microsoft.com/en-us/services/storage/blobs/#pricing (Accessed: 5 October 2021). Nawrocki, M. (2019) 'The benefits of a hybrid cloud strategy for businesses', DataCenterNews. [online] Available at: https://datacenternews.asia/story/the-benefits-of-a-hybrid-cloud-strategy-for (Accessed: 5 October 2021).

“相关推荐”对你有帮助么?

  • 非常没帮助
  • 没帮助
  • 一般
  • 有帮助
  • 非常有帮助
提交
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包

打赏作者

小小强2021

你的鼓励将是我创作的最大动力

¥1 ¥2 ¥4 ¥6 ¥10 ¥20
扫码支付:¥1
获取中
扫码支付

您的余额不足,请更换扫码支付或充值

打赏作者

实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值