# Efficient Low-rank Multimodal Fusion with Modality-Specific Factors

**Authors:** Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, Louis-Philippe Morency  
**arXiv:** [1806.00064](https://arxiv.org/abs/1806.00064)  
**Date:** 2018-05-31  
**Summary page:** https://www.summarizepaper.com/en/paper/1806.00064v1/

## Key points

- Multimodal research is gaining prominence in the field of artificial intelligence
- Multimodal fusion is a key challenge in this field
- Previous studies have used tensors for multimodal representation, but face issues such as exponential increase in dimensions and computational complexity
- The paper proposes a method called Low-rank Multimodal Fusion to address these challenges
- The proposed model achieves competitive results in multimodal sentiment analysis, speaker trait analysis, and emotion recognition tasks
- It significantly reduces computational complexity compared to other methods using tensor representations
- The model performs well under various low-rank settings and is robust
- Authors of the paper: Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency

## Summary

In the field of artificial intelligence, multimodal research is gaining prominence. One of the key challenges in this field is multimodal fusion, which involves integrating multiple unimodal representations into a single compact multimodal representation. Previous studies have utilized tensors for multimodal representation, but these methods often face issues such as exponential increase in dimensions and computational complexity when transforming input into tensors. To address these challenges, this paper proposes a method called Low-rank Multimodal Fusion. This approach leverages low-rank tensors to improve efficiency in performing multimodal fusion. The authors evaluate their model on three different tasks: multimodal sentiment analysis, speaker trait analysis, and emotion recognition. Remarkably, the proposed model achieves competitive results across all these tasks while significantly reducing computational complexity. Furthermore, additional experiments demonstrate that the model can robustly perform well under various low-rank settings. It also outperforms other methods that utilize tensor representations in terms of both training and inference efficiency. The authors of this paper are Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh and Louis-Philippe Morency.