# VMI-VAE: Variational Mutual Information Maximization Framework for VAE With Discrete and Continuous Priors

**Authors:** Andriy Serdega, Dae-Shik Kim  
**arXiv:** [2005.13953](https://arxiv.org/abs/2005.13953)  
**Date:** 2020-05-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2005.13953v1/

## Key points

- Variational Autoencoder (VAE) is a powerful method for learning latent variable models of complex data
- VAE offers a clear objective that is easily optimized and has been widely used in machine learning and artificial intelligence research
- One limitation of VAE is the lack of explicit measurement of the quality of learned representations
- The key innovation of the VMI-VAE framework lies in its objective function, which aims to maximize the mutual information between latent codes and observations
- The VMI-VAE framework acts as a regularizer that prevents VAE from ignoring important aspects of the latent code
- Researchers can selectively emphasize certain components of the latent code that are most informative with respect to the observations
- The proposed framework offers a systematic way to evaluate the mutual information between latent codes and observations within a fixed VAE model
- This capability provides valuable insights into how well the model captures relevant information from the data
- The VMI-VAE framework enhances interpretability and effectiveness by explicitly addressing representation quality assessment

## Summary

Variational Autoencoder (VAE) is a powerful method for learning latent variable models of complex data. It offers a clear objective that is easily optimized and has been widely used in machine learning and artificial intelligence research. However, one limitation of VAE is the lack of explicit measurement of the quality of learned representations. The key innovation of the VMI-VAE framework lies in its objective function, which aims to maximize the mutual information between latent codes and observations. This approach acts as a regularizer that prevents VAE from ignoring important aspects of the latent code. By doing so, researchers can selectively emphasize certain components of the latent code that are most informative with respect to the observations. Moreover, the proposed framework offers a systematic way to evaluate the mutual information between latent codes and observations within a fixed VAE model. This capability provides valuable insights into how well the model captures relevant information from the data. Overall, the VMI-VAE framework enhances interpretability and effectiveness by explicitly addressing representation quality assessment. In conclusion,  opens up new possibilities for improving latent variable models and advancing research in machine learning and artificial intelligence.