12dash / Minority-Attribute-Image-Generation-with-Causal-Dependence-for-Supervised-Tasks

Generating images of minority groups using latent SCM model in the Bidirectional Generative Model.

Fair Image Generation of Minority Groups

The project aims to answer the question :

Can the use of SCM model for disentangling the minority groups in the latent variable be used in a bidirectional generative model to create new images of the minority group for the training of supervised image classification model ?

Specifically in this project the bias is introduced in the dataset which entails:

Women are more likely to smile compared to males.

References

For creating the BiGAN and disentangled representation learning, I have used the following papers and code based :

• Weakly Supervised Disentangled Generative Causal Representation Learning.
• Github Link for DEAR

The report for the project can be found at :

• Overleaf View Link

Datasets

The dataset used in project is the CelebA dataset

• CelebA Dataset