Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MAMA: Meta-optimized Angular Margin Contrastive Framework for Video-Language Representation Learning

Thong Nguyen, Yi Bin, Xiaobao Wu, Xinshuai Dong, Zhiyuan Hu, Khoi Le, Cong-Duy Nguyen, See-Kiong Ng, Luu Anh Tuan

TL;DR

To adapt to the non-uniform concept distribution, MAMA utilizes a multi-layer perceptron (MLP)-parameterized weighting function that maps loss values to sample weights which enable dynamic adjustment of the model's focus throughout the training.

Abstract

Data quality stands at the forefront of deciding the effectiveness of video-language representation learning. However, video-text pairs in previous data typically do not align perfectly with each other, which might lead to video-language representations that do not accurately reflect cross-modal semantics. Moreover, previous data also possess an uneven distribution of concepts, thereby hampering the downstream performance across unpopular subjects. To address these problems, we propose MAMA, a new approach to learning video-language representations by utilizing a contrastive objective with a subtractive angular margin to regularize cross-modal representations in their effort to reach perfect similarity. Furthermore, to adapt to the non-uniform concept distribution, MAMA utilizes a multi-layer perceptron (MLP)-parameterized weighting function that maps loss values to sample weights which enable dynamic adjustment of the model's focus throughout the training. With the training guided by a small amount of unbiased meta-data and augmented by video-text data generated by large vision-language model, MAMA improves video-language representations and achieve superior performances on commonly used video question answering and text-video retrieval datasets. The code, model, and data have been made available at https://nguyentthong.github.io/MAMA.

MAMA: Meta-optimized Angular Margin Contrastive Framework for Video-Language Representation Learning

TL;DR

To adapt to the non-uniform concept distribution, MAMA utilizes a multi-layer perceptron (MLP)-parameterized weighting function that maps loss values to sample weights which enable dynamic adjustment of the model's focus throughout the training.

Abstract

Data quality stands at the forefront of deciding the effectiveness of video-language representation learning. However, video-text pairs in previous data typically do not align perfectly with each other, which might lead to video-language representations that do not accurately reflect cross-modal semantics. Moreover, previous data also possess an uneven distribution of concepts, thereby hampering the downstream performance across unpopular subjects. To address these problems, we propose MAMA, a new approach to learning video-language representations by utilizing a contrastive objective with a subtractive angular margin to regularize cross-modal representations in their effort to reach perfect similarity. Furthermore, to adapt to the non-uniform concept distribution, MAMA utilizes a multi-layer perceptron (MLP)-parameterized weighting function that maps loss values to sample weights which enable dynamic adjustment of the model's focus throughout the training. With the training guided by a small amount of unbiased meta-data and augmented by video-text data generated by large vision-language model, MAMA improves video-language representations and achieve superior performances on commonly used video question answering and text-video retrieval datasets. The code, model, and data have been made available at https://nguyentthong.github.io/MAMA.
Paper Structure (18 sections, 2 theorems, 25 equations, 11 figures, 10 tables, 1 algorithm)

This paper contains 18 sections, 2 theorems, 25 equations, 11 figures, 10 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. MAMA Framework
  4. Video-Language Representation Learning
  5. Meta-optimized Angular Margin Contrastive Framework
  6. Large Vision-Language Model for Augmentation
  7. Experiments
  8. Experimental Settings
  9. Main Results
  10. Ablation Study
  11. Analysis
  12. Conclusion
  13. Proof of Theorem 1
  14. Derivation of the Update Rule in Theoretical Analysis
  15. Datasets
  16. ...and 3 more sections

Key Result

theorem thmcountertheorem

Let $\lambda_{i,j}$ denote the angle between the representation of two samples $i, j$, $\mathcal{L}^{v,t}_{\textup{angular}, i}$ and $\mathcal{L}^{v,t}_{\textup{contrastive}, i}$ denote the training objectives with and without the angular margin, respectively. Then, if $\lambda_{i,i} \leq \frac{\pi}

Figures (11)

  • Figure 1: Examples of video inputs and their textual descriptions.
  • Figure 2: Topic distribution of the MSRVTT dataset. We use Latent Dirichlet Allocation (LDA) to extract topics from manually annotated descriptions of videos.
  • Figure 3: Illustration of the proposed MAMA framework and its components.
  • Figure 4: (Left) Validation videoQA accuracy on MSVD with respect to $\mu$; (Middle) Validation videoQA accuracy on MSVD with respect to $a_0$; (Right) Relationship between loss values and weight values generated by our MLP-parameterized weighting function.
  • Figure 5: a) Relative R@1 improvement in the text-video retrieval task on MSRVTT for each topic; b) Relative accuracy improvement in the videoQA task on MSRVTT with respect to the proportion of questions for which each label is the answer.
  • ...and 6 more figures

Theorems & Definitions (4)

  • theorem thmcountertheorem
  • Theorem 1.
  • proof
  • proof