Fisher Rao

The Fisher-Rao framework, utilizing a geometrically invariant metric, offers a powerful approach to analyzing and improving machine learning models. Current research focuses on leveraging the Fisher-Rao norm to regularize deep neural networks, particularly within adversarial training, aiming to enhance robustness without sacrificing generalization performance. This involves exploring its application in various model architectures and loss functions, such as the cross-entropy loss and the evidence lower bound (ELBO) in variational autoencoders. The resulting advancements promise improved model stability and accuracy, particularly in challenging scenarios like noisy datasets.