Fermionic Born Machines - Classical Training of Quantum Generative Models Based on Fermion Sampling

Explore quantum generative learning through a research seminar that introduces Fermionic Born Machines as a novel approach to classically trainable quantum generative models. Discover how these models address key trainability challenges in quantum machine learning by enabling efficient classical computation of expectation values for local observables, eliminating the need for quantum gradient evaluations during training. Learn about the model's architecture utilizing parameterized magic states and fermionic linear optical transformations with learnable parameters, and understand how the decomposition of magic states into Gaussian operators facilitates efficient expectation value estimation. Examine the favorable optimization characteristics of the loss landscape induced by the specific ansatz structure, and see how fermionic linear optical circuits can be implemented on qubit architectures through fermion-to-qubit mappings for quantum sampling during inference. Review numerical experiments demonstrating the effectiveness of this training framework on systems scaling up to 160 qubits, showcasing the potential computational advantages of maintaining classical training efficiency while preserving the quantum advantage for sampling tasks.