Abstract: Diffractive neural networks have recently emerged as a promising framework for all-optical computing. However, these networks are typically trained for a single task, limiting their potential adoption in systems requiring multiple functionalities. Existing approaches to achieving multi-task functionality either modify the mechanical configuration of the network per task or use a different illumination wavelength or polarization state for each task. In this work, we propose a new control mechanism, which is based on the illumination's angular spectrum. Specifically, we shape the illumination using an amplitude mask that selectively controls its angular spectrum. We employ different illumination masks for achieving different network functionalities, so that the mask serves as a unique task encoder. Interestingly, we show that effective control can be achieved over a very narrow angular range, within the paraxial regime. We numerically illustrate the proposed approach by training a single diffractive network to perform multiple image-to-image translation tasks. In particular, we demonstrate translating handwritten digits into typeset digits of different values, and translating handwritten English letters into typeset numbers and typeset Greek letters, where the type of the output is determined by the illumination's angular components. As we show, the proposed framework can work under different coherence conditions, and can be combined with existing control strategies, such as different wavelengths. Our results establish the illumination angular spectrum as a powerful degree of freedom for controlling diffractive networks, enabling a scalable and versatile framework for multi-task all-optical computing.
Introduction
All-optical diffractive neural networks (DNNs) represent a significant leap in computing, offering the potential for information processing at the speed of light with minimal power consumption. However, a persistent challenge is the "hard-wiring" of these networks; once fabricated, a physical DNN is typically restricted to a single predefined task. Previous attempts to introduce multi-functionality required mechanical adjustments, wavelength shifts, or polarization changes—methods that are often slow or limited in scalability. This paper introduces a novel control mechanism: Illumination Angular Spectrum Encoding, which allows a single static network to perform multiple tasks by simply altering the angular spectrum of the incoming light.
Methods and System Architecture
The core of the proposed approach lies in shaping the illumination using a programmable amplitude mask. The system follows a specific optical pipeline: a plane wave passes through an amplitude mask placed in a "2f" lens configuration. This setup performs a Fourier transform, mapping the spatial pattern of the mask into specific angular components that illuminate the input object.
As the light propagates through a series of trainable diffractive phase masks, the network responds to these varied angles. By changing the pattern on the initial mask, the user can trigger different "programs" or functionalities within the same physical network. The researchers explored both predefined geometric masks (e.g., rings or squares) and "learned" masks optimized via deep learning to maximize task performance.
Key Insights
A significant finding of this study is that effective task control can be achieved within the paraxial regime, using an angular range of less than 0.5 degrees. This ensures compatibility with standard optical components. Furthermore, the study demonstrates that "learned" masks—those optimized alongside the network layers—significantly outperform intuitive geometric designs. The method is also highly versatile; it can be combined with wavelength encoding to further expand the network's task capacity and remains robust under broadband or spatially incoherent light conditions.
Experimental Results
The researchers validated their approach using standard benchmarks like MNIST and EMNIST for image-to-image translation and classification. In a digit translation task, where a handwritten input is converted into different typeset values based on the illumination mask, the learned masks provided a notable 0.7 dB PSNR (Peak Signal-to-Noise Ratio) improvement over simple ring masks.
Quantitatively, the angular spectrum method outperformed traditional wavelength encoding by approximately 2 dB in average PSNR. In multi-dataset classification tests across MNIST, FashionMNIST, KMNIST, and EMNIST, a single network achieved an average accuracy of 82.35%, surpassing both wavelength-encoded (79.95%) and non-encoded (73.5%) baselines.
Conclusion and Future Outlook
While the proposed framework offers a scalable and versatile path for multi-task all-optical computing, challenges remain regarding energy efficiency. Currently, high-fidelity results come at the cost of low light throughput (1% to 8%). Future research will focus on optimizing this trade-off and transitioning from numerical simulations to physical hardware implementations. By establishing the illumination angular spectrum as a powerful degree of freedom, this work paves the way for more complex, real-time optical processing systems.
Summary updated