Neuromorphic (brain-inspired) photonics leverages photonic chips to accelerate neural networks, offering high-speed and energy efficient solutions for use in datacom, autonomous vehicles, or other time sensitive applications. However, the limited size of photonic neural networks limits the complexity of solvable tasks. A natural candidate to provide increased complexity is quantum computing and its exponential speedup capabilities. Specifically, we explore photonic continuous variable (CV) quantum computation. Combining classical networks with trainable CV quantum circuits yields hybrid networks that provide significant trainability and accuracy improvements. On a classification task, hybrid networks achieve the same accuracy as fully classical networks that are twice the size. When noise is applied to the network parameters, the hybrid and classical networks maximize accuracy below the expected on-chip noise level. These results demonstrate that hybrid networks can achieve increased performance with smaller network sizes, providing a promising route to scalable neuromorphic photonic processing.
|