Speaker
Description
All-optical diffractive neural networks are designed to perform inference using light. In practice, however, designing these systems typically requires solving a computationally demanding inverse problem on classical hardware, with long training times driven by the large number of optimization parameters. The core computational bottleneck is the repeated simulation of optical wave propagation, commonly implemented through Fast Fourier Transforms (FFT) or discrete Fourier transforms (DFT). Here, we show that qubit-based platforms can be leveraged to efficiently learn the underlying phase masks, provided access to an efficient implementation of the Quantum Fourier Transform (QFT) is available. Since the QFT can be implemented with a complexity of O((log N)^2), compared to the classical O(NlogN) scaling of the FFT/DFT, this approach opens a pathway to accelerating inverse design in optical systems beyond what is feasible classically. We demonstrate this idea by implementing a coherent classifier on a qubit platform that mirrors the functionality of its optical analogue. The classifier can sort structured photon patterns with different symmetries, both Cartesian and azimuthal, achieving accuracies as high as 100% with 8 layers and 10 qubits. Beyond classification, this framework also enables new coherent feature-embedding strategies, in which tailored unitary transformations map input fields onto chosen computational basis states.
| Apply for student award at which level: | MSc |
|---|---|
| Consent on use of personal information: Abstract Submission | Yes, I ACCEPT |