Hugging Face's logo

Datasets:
xhLiu
/
PMRL_data

ArXiv:
License:
Dataset card Data Studio Files
xet
 Community
1
PMRL_data
File size: 2,618 Bytes
9e073f2
 
 
 
 
 
 
 
 
 
 
 
877117b
9e073f2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
---
license: mit
---



<div align="center">

<h1><a color="red" href="https://arxiv.org/pdf/2507.17343">Principled Multimodal Representation Learning (PMRL)</a></h1>

[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
![License](https://img.shields.io/badge/Accepted-TPAMI'2026-purple)
[![License: MIT](https://img.shields.io/badge/Github-PMRL-black.svg)](https://github.com/Xiaohao-Liu/PMRL)

*A Novel Framework for Representation Learning Across Multiple Modalities*

</div>

---

## ✨ Overview

![](img/top.png)

**Principled Multimodal Representation Learning (PMRL)** addresses the fundamental challenges in multimodal representation learning by proposing a novel framework that achieves simultaneous alignment of multiple modalities without anchor dependency.

### 💡 Our Solution

PMRL introduces a principled approach grounded in **theoretical insights**:

> **Key Insight**: Full alignment corresponds to a rank-1 Gram matrix

Our framework optimizes the dominant singular value of the representation matrix to align modalities along a shared leading direction.

---

## 🎯 Key Features

🔄 **Simultaneous Multi-Modal Alignment**
- No predefined anchor modality required
- Unified representation space for all modalities

🧮 **Softmax-based Loss Function**
- Treats singular values as logits
- Prioritizes the largest singular value for stable optimization

🎯 **Instance-wise Contrastive Regularization**
- Maintains inter-instance separability
- Prevents representation collapse

⚡ **Distributed Training Support**
- Multi-GPU training capabilities
- Efficient data parallel processing

📊 **Comprehensive Evaluation**
- Extensive benchmarking across diverse tasks
- Quantitative and qualitative analysis tools

---

## 🏗️ Architecture

![](img/framework.png)

The PMRL framework consists of three main components:

1. **🔧 Multi-Modal Encoder**: Processes different input modalities
2. **🎯 Singular Value Optimization**: Aligns representations via dominant singular value
3. **🔄 Principled Regularization**: Maintains instance-level discrimination


## Citation

If you find this work useful, please consider citing:

```bibtex
@article{liu2026principled,
  title={Principled multimodal representation learning},
  author={Liu, Xiaohao and Xia, Xiaobo and Ng, See-Kiong and Chua, Tat-Seng},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2026},
  publisher={IEEE}
}
```



<div align="center">


**[🔝 Back to Top](#-principled-multimodal-representation-learning-pmrl)**

</div>