[论文解读] Learning Transferable Visual Models From Natural Language Supervision

Alec Radford, Jong Wook Kim+10 more|arXiv (Cornell University)|Feb 26, 2021

Multimodal Machine Learning Applications参考文献 195被引用 5,296

一句话总结

tldr: CLIP 共同训练图像编码器和文本编码器，以预测图像–文本对的配对，基于 400M（图像，文本）对，从而通过自然语言形成分类器，实现对 30+ 视觉基准的零-shot 转移。

ABSTRACT

State-of-the-art computer vision systems are trained to predict a fixed set of predetermined object categories. This restricted form of supervision limits their generality and usability since additional labeled data is needed to specify any other visual concept. Learning directly from raw text about images is a promising alternative which leverages a much broader source of supervision. We demonstrate that the simple pre-training task of predicting which caption goes with which image is an efficient and scalable way to learn SOTA image representations from scratch on a dataset of 400 million (image, text) pairs collected from the internet. After pre-training, natural language is used to reference learned visual concepts (or describe new ones) enabling zero-shot transfer of the model to downstream tasks. We study the performance of this approach by benchmarking on over 30 different existing computer vision datasets, spanning tasks such as OCR, action recognition in videos, geo-localization, and many types of fine-grained object classification. The model transfers non-trivially to most tasks and is often competitive with a fully supervised baseline without the need for any dataset specific training. For instance, we match the accuracy of the original ResNet-50 on ImageNet zero-shot without needing to use any of the 1.28 million training examples it was trained on. We release our code and pre-trained model weights at https://github.com/OpenAI/CLIP.

研究动机与目标

• 以自然语言监督直接学习可视化表示，以克服固定的事后类别限制。
• 将模型扩展到网页规模的图像-文本数据，研究在不进行数据集特定微调的情况下向多样化视觉任务的转移。
• 展示零-shot CLIP 在许多数据集上可与或超过完全监督基线，并分析鲁棒性与提示（prompts）的作用。

提出的方法

• 使用对比目标在批量对上联合训练图像编码器和文本编码器，以对齐图像和文本的嵌入。
• 使用一个大型 400M（图像，文本）数据集（WIT），提供英语语言的标题/描述。
• 通过文本编码器生成的文本派生分类器对图像进行分类，评估零-shot 转移。
• 对 ResNet 基于的和 Vision Transformer 的图像编码器以及 Transformer 文本编码器进行实验。
• 对文本提示进行提示工程与集成，以提升零-shot 精度。
• 使用大批量对比学习进行训练；使用线性投影映射到多模态嵌入空间；对匹配与不匹配对在对称交叉熵损失上进行优化。

实验结果

研究问题

• RQ1 Can a model learned from natural language supervision transfer effectively to a wide range of vision tasks without dataset-specific labels?
• RQ2 How does zero-shot CLIP performance compare to supervised baselines across diverse datasets and tasks?
• RQ3 What factors (architecture, data scale, prompts) influence zero-shot transfer performance and robustness?

主要发现

• CLIP achieves non-trivial zero-shot transfer to 30+ vision datasets, often competitive with fully supervised baselines without dataset-specific training.
• On ImageNet zero-shot, CLIP matches the accuracy of the original ResNet-50 without using the 1.28 million ImageNet training examples.
• Prompt engineering and ensembling substantially boost zero-shot accuracy (nearly 5-point gain on ImageNet across datasets).
• Zero-shot CLIP can outperform a fully supervised ResNet-50 baseline on a majority of general object datasets and several action recognition datasets (e.g., Kinetics700, UCF101).
• Compared to Visual N-Grams, CLIP significantly improves zero-shot ImageNet and other datasets; CLIP-ResNet-50 trained on 400M image-text pairs matches Visual N-Grams results on ImageNet when trained from scratch on a similar data scale.
• Zero-shot CLIP exhibits greater robustness than equivalent-accuracy supervised ImageNet models, suggesting zero-shot evaluations better reflect broad capability.