[Paper Review] Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision

Chao Jia, Yinfei Yang+8 more|arXiv (Cornell University)|Feb 11, 2021

Multimodal Machine Learning Applications75 references1,195 citations

TL;DR

ALIGN trains visual and vision-language embeddings from a large-scale noisy image-alt-text corpus using a dual-encoder with a contrastive loss, achieving state-of-the-art zero-shot and fine-tuned performance on vision and cross-modal retrieval tasks.

ABSTRACT

Pre-trained representations are becoming crucial for many NLP and perception tasks. While representation learning in NLP has transitioned to training on raw text without human annotations, visual and vision-language representations still rely heavily on curated training datasets that are expensive or require expert knowledge. For vision applications, representations are mostly learned using datasets with explicit class labels such as ImageNet or OpenImages. For vision-language, popular datasets like Conceptual Captions, MSCOCO, or CLIP all involve a non-trivial data collection (and cleaning) process. This costly curation process limits the size of datasets and hence hinders the scaling of trained models. In this paper, we leverage a noisy dataset of over one billion image alt-text pairs, obtained without expensive filtering or post-processing steps in the Conceptual Captions dataset. A simple dual-encoder architecture learns to align visual and language representations of the image and text pairs using a contrastive loss. We show that the scale of our corpus can make up for its noise and leads to state-of-the-art representations even with such a simple learning scheme. Our visual representation achieves strong performance when transferred to classification tasks such as ImageNet and VTAB. The aligned visual and language representations enables zero-shot image classification and also set new state-of-the-art results on Flickr30K and MSCOCO image-text retrieval benchmarks, even when compared with more sophisticated cross-attention models. The representations also enable cross-modality search with complex text and text + image queries.

Motivation & Objective

• Motivate scalable visual and vision-language representation learning without expensive data curation.
• Propose a simple dual-encoder architecture trained with a contrastive loss on a billion-plus noisy image-alt text pairs.
• Demonstrate that scale can compensate for noise to achieve strong transfer performance across vision and cross-modal tasks.

Proposed method

• Use EfficientNet as the image encoder and BERT as the text encoder with a shared embedding space.
• Train with a normalized softmax contrastive loss in both image-to-text and text-to-image directions.
• Apply minimal frequency-based filtering to construct a 1.8B image-alt-text dataset from Conceptual Captions without heavy post-processing.
• Evaluate zero-shot and fine-tuned retrieval on Flickr30K and MSCOCO, plus cross-modal benchmarks like Crisscrossed Captions (CxC).
• Show zero-shot ImageNet classification by querying with class name prompts in the text encoder.

Experimental results

Research questions

• RQ1 Can a simple dual-encoder trained on a very large noisy image-text dataset achieve state-of-the-art cross-modal retrieval without heavy filtering?
• RQ2 How does scale and data quality trade off in learning visual and vision-language representations?
• RQ3 What transfer performance can be achieved in zero-shot and fine-tuned settings for image classification and image-text retrieval?
• RQ4 Do multilingual extensions generalize cross-l modality retrieval to non-English data?
• RQ5 What are the qualitative properties of the learned embeddings (compositionality, text+image query capabilities)?

Key findings

• ALIGN achieves state-of-the-art results on Flickr30K and MSCOCO for image-text retrieval in zero-shot and fine-tuned settings.
• In zero-shot image classification on ImageNet, ALIGN attains 76.4% top-1 accuracy using class-name prompts, comparable to CLIP.
• On ImageNet, ALIGN reaches 88.64% top-1 accuracy with the image encoder alone in visual classification tasks.
• CxC retrieval and SITS metrics show sizable gains over previous VSE and cross-attention models, with especially large improvements in image-to-text and text-to-image recall.
• A multilingual ALIGN model (ALIGN mling) trained on 100+ languages outperforms some baselines in zero-shot multilingual image-text retrieval on Multi30K, demonstrating cross-lingual capabilities.