# Memory Fusion Network for Multi-view Sequential Learning

**Authors:** Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria, Louis-Philippe Morency  
**arXiv:** [1802.00927](https://arxiv.org/abs/1802.00927)  
**Date:** 2018-02-03  
**Summary page:** https://www.summarizepaper.com/en/paper/1802.00927v1/

## Key points

- Multi-view sequential learning is a crucial area of study focusing on understanding and analyzing interactions between different views.
- Researchers led by Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency developed the Memory Fusion Network (MFN) to address this complex problem.
- MFN captures view-specific interactions and cross-view interactions using the System of LSTMs and Delta-memory Attention Network (DMAN) with Multi-view Gated Memory.
- Through experimentation on benchmark datasets, MFN outperformed existing approaches for multi-view sequential learning.
- The research was presented at AAAI 2018 as an oral presentation and demonstrates the potential of MFN in advancing the field of multi-view sequential learning.

## Summary

In the field of machine learning, multi-view sequential learning is a crucial area of study that focuses on understanding and analyzing multi-view sequences. These sequences involve interactions between different views, including view-specific interactions and cross-view interactions. To address this complex problem, a group of researchers led by Amir Zadeh, Paul Pu Liang, Navonil Mazumder, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency have developed a novel neural architecture known as the Memory Fusion Network (MFN). The MFN is designed to explicitly capture both types of interactions within a neural framework and dynamically model them over time. The architecture consists of two key components: the System of LSTMs and the Delta-memory Attention Network (DMAN) with Multi-view Gated Memory. In the System of LSTMs, individual LSTM functions are assigned to each view to learn view-specific interactions independently. The DMAN mechanism then identifies cross-view interactions using a specialized attention mechanism. Through extensive experimentation on various publicly available benchmark datasets, the researchers compared MFN against existing approaches for multi-view sequential learning. Remarkably, MFN outperformed all other methods in terms of performance metrics. Not only did it surpass existing multi-view approaches but it also set new state-of-the-art results for these challenging multi-view datasets. This groundbreaking research was presented at AAAI 2018 as an oral presentation and showcases the potential of the Memory Fusion Network in advancing the field of multi-view sequential learning. With its innovative design and superior performance, MFN represents a significant contribution to machine learning research and opens up new possibilities for analyzing complex multi-view data structures effectively.