Team Members:

• Sarvesh Shashidhar(24M2152)
• Ritik (24M2155)
• Ankish Chandresh(24M2163)

Link to PPT:

PPT

Link to Referenced repository and code:

jenkspt/deepinfomax

Steps for reproducing our results:

1. Train a dim or vae model by running trainer.py script or trainVAE.py script respectively
2. Once the model/encoder is trained train a classifier that uses representations from the model trained in step 1 as input
3. Note the train and validation accuracy

Run The following for DIM-Global:

DIM-Global refers to the setting where purely Global Mutual Information maximisation is performed

python trainer.py --alpha=1 --beta=0 --gamma=0 --epochs=50 --batch-size=128

Run The following for DIM-Local:

DIM-Global refers to the setting where purely Local Mutual Information maximisation is performed

python trainer.py --alpha=0 --beta=1 --gamma=0 --epochs=50 --batch-size=128

Run the following for Training VAE:

python trainVAE.py

For training the classifier:

Set model flag as 'vae' for using a VAE model as feature vector and set it to 'dim' for using Deep Infomax feature vector for final classification downstream task

python classifier.py --epochs=200 --lr=0.01 --batch_size=128 --model='vae'

Link to our paper:

Learning deep representations by mutual information estimation and maximization

Results

References

Causal models on probability spaces

What Makes for Good Views for Contrastive Learning?

Analyzing Inverse Problems with Invertible Neural Networks

A Simple Framework for Contrastive Learning of Visual Representations

Contrastive Learning Inverts the Data Generating Process