Photonic Neuromorphic Computing with an Optically-Programmable Processing Unit

Deep learning has revolutionized computer vision, natural language processing, and autonomous systems. The increasing complexity of neural networks now demands specialized hardware to improve performance and energy efficiency. Photonic computing offers a promising, high-speed, low-power alternative, but a key challenge remains: achieving the nonlinearity required for deep neural networks. Furthermore, neuromorphic hardware is typically task-specific, making it difficult to adapt once training is complete, often requiring new hardware for each task.
In this work, we propose an optically reprogrammable neuromorphic processing unit that overcomes these limitations. Our system employs phase-shaped laser pulses interacting with a programmable control pulse within a white-light photonic fiber, generating the nonlinear response critical for neural network operations. Further, this interaction allows the system to be dynamically reconfigured through the programming pulse, which essentially determines the weights of the neural network, enabling task-specific learning and classification.
Unlike traditional reservoir computing, our approach offers real-time adaptability without the need for digital backend training and can be scaled for solving increasingly complex tasks by adapting the employed materials. Using guided design of the photonic fiber material, with the aim for generating complex non-linear response in the phase and spectral modes of the computing light pulses, the platform supports parallel processing, by taking advantage of multiple interaction patterns between input and programming pulses, as a promising approach for energy-efficient solutions in high-performance neural network hardware.