HyperAIHyperAI超神经
首页算力平台文档资讯论文教程数据集百科SOTALLM 模型天梯GPU 天梯顶会
全站搜索
关于
HyperAIHyperAI超神经
4 个月前

基于视觉-语言模型的零样本分类标签传播

查看论文详情
StojnićVladan ; KalantidisYannis ; ToliasGiorgos
基于视觉-语言模型的零样本分类标签传播

摘要

视觉-语言模型（VLMs）在零样本分类任务中展现了令人印象深刻的表现，即仅提供类别名称列表时的分类能力。本文研究了在存在未标记数据的情况下进行零样本分类的问题。我们利用未标记数据的图结构，引入了一种基于标签传播（LP）的方法——ZLaP，该方法利用测地距离进行分类。我们将标签传播技术应用于同时包含文本和图像特征的图，并进一步提出了一种基于对偶解和稀疏化步骤的有效归纳推理方法。我们进行了广泛的实验，评估了该方法在14个常用数据集上的有效性，并展示了ZLaP优于最新的相关工作。代码：https://github.com/vladan-stojnic/ZLaP

代码仓库

vladan-stojnic/zlap
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
transductive-zero-shot-classification-onZLaP
Top 1 Accuracy: 72.7
transductive-zero-shot-classification-on-1ZLaP
Accuracy: 60.9
transductive-zero-shot-classification-on-2ZLaP
Accuracy: 73.4
transductive-zero-shot-classification-on-3ZLaP
Accuracy: 92.8
transductive-zero-shot-classification-on-4ZLaP
Accuracy: 71.9
transductive-zero-shot-classification-on-5ZLaP
Accuracy: 72.1
transductive-zero-shot-classification-on-6ZLaP
Accuracy: 83.7
transductive-zero-shot-classification-on-7ZLaP
Accuracy: 93.6
transductive-zero-shot-classification-on-8ZLaP
Accuarcy: 73.3
transductive-zero-shot-classification-on-9ZLaP
Accuracy: 77.7
transductive-zero-shot-classification-on-cubZLaP
Accuracy: 64.1
transductive-zero-shot-classification-on-dtdZLaP
Accuracy: 51.8
transductive-zero-shot-classification-on-fgvcZLaP
Accuracy: 28.4
transductive-zero-shot-classification-on-foodZLaP
Accuracy: 87.9
zero-shot-learning-on-caltech-101ZLaP
Accuracy: 84
zero-shot-learning-on-caltech-101ZLaP*
Accuracy: 83.1
zero-shot-learning-on-cifar-10ZLaP*
Accuracy: 93.6
zero-shot-learning-on-cifar-10ZLaP
Accuracy: 93.4
zero-shot-learning-on-cifar-100ZLaP*
Accuracy: 74.2
zero-shot-learning-on-cifar-100ZLaP
Accuracy: 74
zero-shot-learning-on-cub-200-2011ZLaP*
Accuracy: 64.2
zero-shot-learning-on-cub-200-2011ZLaP
Accuracy: 64.3
zero-shot-learning-on-dtdZLaP*
Accuracy: 51
zero-shot-learning-on-dtdZLaP
Accuracy: 51.2
zero-shot-learning-on-eurosatZLaP*
Accuracy: 63.2
zero-shot-learning-on-fgvc-aircraftZLaP
Accuracy: 29.1
zero-shot-learning-on-fgvc-aircraftZLaP*
Accuracy: 29
zero-shot-learning-on-flowers-102ZLaP*
Accuracy: 75.5
zero-shot-learning-on-flowers-102ZLaP
Accuracy: 75.9
zero-shot-learning-on-food-101ZLaP*
Accuracy: 87.9
zero-shot-learning-on-food-101ZLaP
Accuracy: 87.8
zero-shot-learning-on-imagenetZLaP
Top 1 Accuracy: 72.1
zero-shot-learning-on-imagenetZLaP*
Top 1 Accuracy: 72.1
zero-shot-learning-on-oxford-iiit-petsZLaP*
Accuracy: 89
zero-shot-learning-on-oxford-iiit-petsZLaP
Accuracy: 90
zero-shot-learning-on-stanford-carsZLaP*
Accuracy: 71.8
zero-shot-learning-on-stanford-carsZLaP
Accuracy: 71.2
zero-shot-learning-on-sun397ZLaP*
Accuracy: 71.4
zero-shot-learning-on-sun397ZLaP
Accuracy: 71
zero-shot-learning-on-ucf101ZLaP*
Accuracy: 76.3
zero-shot-learning-on-ucf101ZLaP
Accuracy: 76.3

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
基于视觉-语言模型的零样本分类标签传播 | 论文 | HyperAI超神经