Ilyas et al. (2019)

New research explains how adversarial training reduces superposition in neural networks

Researchers have explored the relationship between adversarial examples and superposition in neural networks. Building on prior work that linked adversarial examples to superposition and showed adversarial training redu…