Description
Nonlinear mappings lie at the core of modern machine learning models, underpinning their success in highly complex tasks. In optical neural networks (ONNs), nonlinearities are traditionally generated through higher-order light–matter interactions, commonly by encoding information in the electric field and measuring the response of specific optical materials. However, nonlinear encoding can also be achieved through the recurrent linear modulation of a probe beam using a spatial light modulator (SLM), relying solely on the geometry of the optical setup.
This form of nonlinearity, referred to as structural nonlinearity, has previously been explored in the training of ONNs for image-based tasks using an integrating sphere and a digital micromirror device (DMD). Nevertheless, such systems are highly sensitive to environmental conditions, which limits their scalability. In addition, precisely engineering a desired nonlinear transformation in these platforms remains a practical challenge.
In this work, we propose a controllable optical arrangement capable of simultaneously generating linear and nonlinear optical mappings via structural nonlinearity to address the challenge of training physical machine learning models. We demonstrate the advantage of this nonlinear encoding by training optical reservoir models for chaotic time series forecasting, significantly enhancing model performance in both short- and long-term predictions. This method contributes toward simple-to-implement optical nonlinear functions, advancing the development of physical optical neural networks.
| I am the presenting author | Yes |
|---|