Speaker
Description
Optical neural networks offer a path toward direct, light-speed inference. These networks have a range of functionalities, including handwritten-digit recognition, optical field phase imaging, and quantum state generation. However, their physical synthesis is computationally demanding: iterative, taxing simulations of light propagation dominate the training cycle. Our research question is then this: is there an analogous system or approach that exists that can allow us to probe these optical systems, and improve their design? Here, we propose an equivalent quantum model can be used to design better optical neural networks. We design a quantum neural network which is inspired by the physics of our optical system, using Quantum Fourier Transforms (QFTs), which are crucial to many known quantum algorithms. QFTs are used to replace the use of classical Fast Fourier Transforms (FFTs), which conventionally model wave propagation. By training this quantum neural network to distinguish between a variety of images, the optimal rotation angles are extracted from our circuit and can be physically implemented to create phase masks for a diffractive network. A variational algorithm is used to train the quantum-assisted architecture, which yields exceptional classification fidelity across all tested modal suites. The baseline quantum neural network achieved classification accuracies as high as $93.8\%$ for orthogonal families of spatial patterns, and $88.9\%$ for non-orthogonal families. Notably, integrating an upgraded entanglement topology pushes classification accuracy to 100\% with as few as 8 layers and 11 qubits. These findings highlight quantum machine learning as a viable alternative for the design of photonic processors.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Tensor methods and quantum-inspired machine learning |