Exponential Quantum Advantage in Communication for Distributed Learning

We show that distributed, parameterized models can be trained using gradient descent with exponentially less communication if realized using quantum computers that can exchange quantum states, with additional privacy benefits.