Causal Model -- VAE

CS 7290 Project: Causal Modeling with a Variational Autoencoder

Project Goal

Refactor the provided program such that there is a causal relationship between the given variables.

What would the image be like if we change something?

Confusions during Development

• Where is the DAG? How to find it?
• Which part to refactor?
• What is the use of SCM?

Timeline

Pre-launching

Week 1

• Topic Determination
• Literature Reviews

Week 2

• Literature Reviews
• Task Assignment

Developing

Week 3

• Understanding the pyro tutorial for creating, training, and making inference on VAE models
• De-noising exogenous variables for image representations in SCM (stochastic causal model)

Week 4

• Fine-tuning Encoder and Decoder in VAE model for refactoring
• Training the model so that the orignial images can be reconstructed
• Creating conditioned model on one of the labels
• Intervening on the conditioned model by applying do-operation on the other labels
• Visualizing results and drawing insights from them

Post-development

Week 5

• Formatting the notebook
• Ready for presentation

References

Understanding VAEs

Tutorial Provided By TA

Pyro Documentation

An Introduction to Variational Autoencoders

Robustly Disentangled Causal Mechanisms: Validating Deep Representations for Interventional Robustness

Causal Effect Inference with Deep Latent-Variable Models