Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data-Efficient Contrastive Language-Image Pretraining: Prioritizing Data Quality over Quantity

Siddharth Joshi, Arnav Jain, Ali Payani, Baharan Mirzasoleiman

TL;DR

This work proposes the first theoretically rigorous data selection method for CLIP and shows that subsets that closely preserve the cross-covariance of the images and captions of the full data provably achieve a superior generalization performance.

Abstract

Contrastive Language-Image Pre-training (CLIP) on large-scale image-caption datasets learns representations that can achieve remarkable zero-shot generalization. However, such models require a massive amount of pre-training data. Improving the quality of the pre-training data has been shown to be much more effective in improving CLIP's performance than increasing its volume. Nevertheless, finding small subsets of training data that provably generalize the best has remained an open question. In this work, we propose the first theoretically rigorous data selection method for CLIP. We show that subsets that closely preserve the cross-covariance of the images and captions of the full data provably achieve a superior generalization performance. Our extensive experiments on ConceptualCaptions3M and ConceptualCaptions12M demonstrate that subsets found by \method\ achieve over 2.7x and 1.4x the accuracy of the next best baseline on ImageNet and its shifted versions. Moreover, we show that our subsets obtain 1.5x the average accuracy across 11 downstream datasets, of the next best baseline. The code is available at: https://github.com/BigML-CS-UCLA/clipcov-data-efficient-clip.

Data-Efficient Contrastive Language-Image Pretraining: Prioritizing Data Quality over Quantity

TL;DR

This work proposes the first theoretically rigorous data selection method for CLIP and shows that subsets that closely preserve the cross-covariance of the images and captions of the full data provably achieve a superior generalization performance.

Abstract

Contrastive Language-Image Pre-training (CLIP) on large-scale image-caption datasets learns representations that can achieve remarkable zero-shot generalization. However, such models require a massive amount of pre-training data. Improving the quality of the pre-training data has been shown to be much more effective in improving CLIP's performance than increasing its volume. Nevertheless, finding small subsets of training data that provably generalize the best has remained an open question. In this work, we propose the first theoretically rigorous data selection method for CLIP. We show that subsets that closely preserve the cross-covariance of the images and captions of the full data provably achieve a superior generalization performance. Our extensive experiments on ConceptualCaptions3M and ConceptualCaptions12M demonstrate that subsets found by \method\ achieve over 2.7x and 1.4x the accuracy of the next best baseline on ImageNet and its shifted versions. Moreover, we show that our subsets obtain 1.5x the average accuracy across 11 downstream datasets, of the next best baseline. The code is available at: https://github.com/BigML-CS-UCLA/clipcov-data-efficient-clip.
Paper Structure (13 sections, 21 equations, 3 figures, 10 tables, 1 algorithm)

This paper contains 13 sections, 21 equations, 3 figures, 10 tables, 1 algorithm.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. PROBLEM FORMULATION
  4. FINDING THE MOST GENERALIZABLE SUBSETS
  5. Preserving the Cross-Covariance of Data
  6. Deriving the Final Objective for Finding the Most Generalizable Subset
  7. ClipCov: Efficiently Finding the Most Generalizable Subset
  8. EXPERIMENTS
  9. CONCLUSION
  10. Experimental Details
  11. Accuracy on downstream tasks
  12. Additional Training Details
  13. Visualization of CC12M Results

Figures (3)

  • Figure 1: Visualization of examples selected by $\!F_{\text{class}}(S) \!+\! F_{\text{self}}(S)\!$ in cross-modal similarity space. ClipCov selects central examples that are representative of different subgroups in every latent class.
  • Figure 2: Performance across subset of different sizes selected from ConceptualCaptions3M. Gray region indicates accuracy within 90% of that of full data.
  • Figure 3: Performance across subset of sizes 5% and 10% from CC12M

Theorems & Definitions (1)

  • Definition 4.1: Cross-Modal Similarity