Adversarial Augmentations
Daniel and Alex discuss the concept of training adversaries to generate augmented images and the benefits it brings to self-supervised learning. They explore the simplicity of the L1 norm constraint and the diverse and strange perturbations that can be generated. They also touch on the potential of combining augmentations and the importance of input-dependent views in varied data sets.In this clip
From this podcast
The Gradient
Alex Tamkin on Self-Supervised Learning and Large Language Models
Related Questions