Nanophotonic neural networks are an exciting emerging technology which promises low-energy, ultra high-throughput machine learning systems implemented purely optically. Our lab has previously done work on these devices, and our new paper which extends programmable photonics to the quantum domain is now on arXiv! In this paper, we describe a photonic architecture for a quantum programmable gate array (QPGA) which can be dynamically reprogrammed to perform any quantum computation. We show how to exactly prepare arbitrary quantum states and operators on the device, and we apply machine learning techniques to automatically implement highly compact approximations to important quantum circuits. Below is an animation of a simulated QPGA being trained to implement a quantum Fourier transform on five qubits.