The power of quantum neural networks

Fault-tolerant quantum computers offer the promise of dramatically improving machine learning through speed-ups in computation or improved model scalability. In near-term, however, benefits are not so clear. Understanding expressibility and trainability models-and neural networks particular-requires further investigation. this work, we use tools from information geometry to define a notion for classical models. The effective dimension, which depends on Fisher information, is used prove novel generalisation bound establish robust measure expressibility. We show that able achieve significantly better dimension than comparable networks. To then assess models, connect spectrum barren plateaus, problem vanishing gradients. Importantly, certain can resilience phenomenon train faster models due their favourable optimisation landscapes, captured by more evenly spread spectrum. Our work first demonstrate well-designed an advantage over higher training ability, verify real hardware.