Hybrid Quantum-Classical Photonic Neural Networks
Loading...
Authors
Austin, Tristan
Date
2025-05-14
Type
thesis
Language
Keyword
Quantum computing , Neuromorphic computing , Photonics
Alternative Title
Abstract
The integration of neuromorphic and quantum photonic computing offers a promising route toward scalable, energy-efficient, and high-capacity information processing systems. In this work, I explore hybrid continuous-variable quantum-classical neural networks (CVQNNs) that offer a potential path to scaling photonic computing by leveraging quantum resources to improve trainability and reduce parameter count. Specifically, I benchmark the performance of hybrid CVQNNs against classical photonic networks on a classification task, using parameter count (the network size) as a proxy for hardware complexity. I find that hybrid networks consistently outperform classical networks of comparable size, achieving similar or superior accuracy with fewer trainable parameters. A 120-parameter hybrid network, for example, achieves equivalent performance to a 235-parameter classical network.
I then extend this analysis to more realistic settings by introducing noise, limited bit precision, and photonic loss. Using this framework, I determine the minimum experimental requirements, such as effective number of bits (ENOB) and shot count, for accurate CVQNN performance. I find that most gate operations remain within reach of current integrated photonic platforms, with the exception of the Kerr non-linearity, which remains a key bottleneck. Finally, I estimate the physical footprint of quantum photonic layers and compare them to classical implementations, showing that although CVQNNs currently require more space, there is significant potential to shrink hardware components. These results underscore the ability of hybrid quantum photonic networks to perform complex tasks with fewer parameters, offering a new route to efficient photonic computation.