On the Learning Dynamics of Attention Networks

Rahul Vashisht, Harish G. Ramaswamy

Attention models are typically learned by optimizing one of three standard loss functions that are variously called -- soft attention, hard attention, and latent variable marginal…