Exploring the photonic superiority: powerful computational expressiveness and trainability in optical computing

The blossoming of artificial intelligence technologies demands greater computational power. Optical computing, with its ultra-fast operation and inherent parallel processing capabilities, is considered promising for high-performance computing tasks. However, there is no evidence that optical computing outperforms classical neural networks in machine learning tasks. In this work, we explored this issue by evaluating the expressive and trainable capabilities of optic propagation processes. This evaluation was based on the Fisher information and the effective dimension theories. Practical machine learning experiments were conducted to validate our theoretical findings. The results revealed that optical-propagation-based computing exhibited a superior effective dimension and quicker training speed than classical feedforward neural networks, underscoring optical computing’s benefits in machine learning applications.