Neural networks owe much of their power to universal function approximation — the ability to represent arbitrary nonlinear relationships. But translating that capability into optical hardware is not straightforward, because it requires deliberately engineering the computational features carried by light.
A paper posted on arXiv takes up this challenge, arguing that wavelength diversity can expand the expressive power of optical computing. By exploiting multiple wavelengths, the approach aims to give optical systems more degrees of freedom for encoding and processing information.
The abstract stresses that realizing universal approximation in optics depends on how computational features are designed into the light itself. The paper's contribution is to identify wavelength diversity as a concrete route toward that goal, though the full details of the proposed method are not included in the abstract.