# Meta Module Network for Compositional Visual Reasoning

**Authors:** Wenhu Chen, Zhe Gan, Linjie Li, Yu Cheng, William Wang, Jingjing Liu  
**arXiv:** [1910.03230](https://arxiv.org/abs/1910.03230)  
**Date:** 2019-10-08  
**Summary page:** https://www.summarizepaper.com/en/paper/1910.03230v5/

## Key points

- The Meta Module Network (MMN) is a new architecture designed to address scalability and generalizability issues in Neural Module Networks (NMNs) for compositional visual reasoning tasks.
- NMNs have strong interpretability and compositionality, but their customized modules make them impractical when scaling up to larger sets of functions in complex tasks.
- MMN uses a meta module that can take in function recipes and dynamically morph into diverse instance modules, which are woven into an execution graph for complex visual reasoning while inheriting the strong explainability and compositionality of NMN.
- With this flexible instantiation mechanism, the parameters of instance modules are inherited from the central meta module, retaining the same model complexity as the function set grows, promising better scalability.
- Functions are encoded into the embedding space, allowing unseen functions to be represented based on their structural similarity with previously observed ones ensuring better generalizability.
- Experiments on GQA and CLEVR datasets validate the superiority of MMN over state-of-the-art NMN designs. Synthetic experiments on held-out unseen functions from GQA dataset also demonstrate the strong generalizability of MMN.
- The authors Wenhu Chen, Zhe Gan, Linjie Li, Yu Cheng, William Wang and Jingjing Liu have released their code and model on Github for practical use.
- The paper has been accepted for oral presentation at WACV 21 conference.

## Summary

The Meta Module Network (MMN) is a novel architecture proposed to address the scalability and generalizability issues of Neural Module Networks (NMNs) in compositional visual reasoning tasks. While NMNs have strong interpretability and compositionality, their customized modules for specific functions make them impractical when scaling up to larger sets of functions in complex tasks. To overcome these limitations, MMN centers on a novel meta module that can take in function recipes and dynamically morph into diverse instance modules. The instance modules are then woven into an execution graph for complex visual reasoning, inheriting the strong explainability and compositionality of NMN. With such a flexible instantiation mechanism, the parameters of instance modules are inherited from the central meta module, retaining the same model complexity as the function set grows which promises better scalability. Furthermore, as functions are encoded into the embedding space, unseen functions can be readily represented based on their structural similarity with previously observed ones ensuring better generalizability. Experiments on GQA and CLEVR datasets validate the superiority of MMN over state-of-the-art NMN designs. Synthetic experiments on held-out unseen functions from GQA dataset also demonstrate the strong generalizability of MMN. The authors Wenhu Chen, Zhe Gan, Linjie Li, Yu Cheng, William Wang and Jingjing Liu have released their code and model on Github for practical use. The paper has been accepted for oral presentation at WACV 21 conference.