tourdeml / robust-representation-random-convolutions

Robust and Generalizable Visual Representation Learning via Random Convolutions in PyTorch

Robust and Generalizable Visual Representation Learning via Random Convolutions

Re-implementation of the ICLR'21 paper "Robust and Generalizable Visual Representation Learning via Random Convolutions" that introduces Random Convolutions and Consistency loss. This is an interesting paradigm that opens up avenues with respect to shape and texture.

Related research:

1. Shape vs Texture (Add link) (Imagenet paper)
2. Mixup

Adding this to a pipeline that automatically searches for image augmentation would be interesting to see where this augmentation would ideally fit.

Progress

• Functions.
• Docs, Format code.
• Add MNIST example with RandConv and consistency loss.