当前位置:网站首页>ICML2022 | 长尾识别中分布外检测的部分和非对称对比学习
ICML2022 | 长尾识别中分布外检测的部分和非对称对比学习
2022-07-05 18:32:00 【智源社区】

现有的分布外(OOD)检测方法通常以具有平衡类分布的训练集为基准。然而,在实际应用中,训练集通常具有长尾分布。在这项工作中,我们首先证明了现有的OOD检测方法在训练集为长尾分布时通常会出现显著的性能下降。通过分析,我们认为这是因为模型很难将分布中的少数尾类样本与真实的OOD样本区分开来,使得尾类更容易被错误地检测为OOD。为了解决这一问题,提出了部分和非对称监督对比学习(Partial and Asymmetric Supervised contrast Learning, PASCL),明确鼓励模型区分尾部类内分布样本和OOD样本。为了进一步提高分布内分类精度,我们提出了辅助分支微调(Auxiliary Branch Finetuning),它使用BN和分类层的两个独立分支分别进行异常检测和分布内分类。直觉上,分布内异常数据和OOD异常数据具有不同的底层分布。我们的方法在CIFAR10-LT、CIFAR100-LT和ImageNet-LT上的异常检测假阳性率(FPR)分别为1.29%、1.45%、0.69%,分布内分类准确率为3.24%、4.06%、7.89%。代码和预训练的模型可以在https://github.com/amazon-research/long-tailed-ood-detection找到。
边栏推荐
- 进程间通信(IPC):共享内存
- Is it safe to make fund fixed investment on access letter?
- Memory management chapter of Kobayashi coding
- Matlab built-in function how different colors, matlab subsection function different colors drawing
- Can communication of nano
- [use electron to develop desktop on youqilin]
- 苹果手机炒股安全吗?打新债是骗局吗?
- [paddleclas] common commands
- Cronab log: how to record the output of my cron script
- Generate classes from XML schema
猜你喜欢

Reptile 01 basic principles of reptile

Take a look at semaphore, the current limiting tool provided by JUC

Image classification, just look at me!

达梦数据库udf实现

U-Net: Convolutional Networks for Biomedical Images Segmentation

Tupu software digital twin | visual management system based on BIM Technology

Failed to virtualize table with JMeter

使用JMeter录制脚本并调试

buuctf-pwn write-ups (9)

《ClickHouse原理解析与应用实践》读书笔记(5)
随机推荐
FCN: Fully Convolutional Networks for Semantic Segmentation
7-1 链表也简单fina
兄弟组件进行传值(显示有先后顺序)
@Extension, @spi annotation principle
Reptile 01 basic principles of reptile
图片数据不够?我做了一个免费的图像增强软件
让更多港澳青年了解南沙特色文创产品!“南沙麒麟”正式亮相
Clickhouse (03) how to install and deploy Clickhouse
爬虫01-爬虫基本原理讲解
Is it safe to make fund fixed investment on access letter?
Multithreading (I) processes and threads
SAP feature description
7-2 保持链表有序
Trust counts the number of occurrences of words in the file
瞅一瞅JUC提供的限流工具Semaphore
The easycvr platform reports an error "ID cannot be empty" through the interface editing channel. What is the reason?
基于can总线的A2L文件解析(3)
Pytorch yolov5 training custom data
ConvMAE(2022-05)
buuctf-pwn write-ups (9)