All Optical Classification Surpasses Cascaded Diffractive Networks through Dual Wavelength Differential Modulation within a Single Layer Architecture

Abstract: Diffractive deep neural networks (D2NNs), which perform computation using light instead of electrons, offer a promising pathway toward accelerating artificial intelligence by leveraging the inherent advantages of optics in speed, parallelism, and energy efficiency. However, conventional multi-layer D2NNs suffer from inter-layer misalignments that significantly increase system complexity and degrade performance, particularly under visible-light operation where optical alignment is highly sensitive. Here, we present a compact, single-layer dual-wavelength differential D2NN that combines wavelength-division multiplexing with differential intensity detection to enable high-accuracy all-optical classification while substantially reducing hardware complexity. By encoding complementary spatial frequency information at two distinct wavelengths, the proposed network overcomes non-negativity constraints and feature loss inherent to single-wavelength systems. Our numerical experiments achieve outstanding classification accuracies of 98.59% on MNIST and 90.4% on Fashion MNIST using only 40k trainable parameters surpassing the performance of conventional five layer D2NNs (91.33% and 83.67%, respectively) with merely 20% of the parameter count. Furthermore, the network maintains strong performance with only 10k parameters and demonstrates enhanced robustness against random phase perturbations, optical occlusions, and input noise. To best of our knowledge, this work represents the first demonstration of a single layer diffractive optical network that achieves such high classification accuracy, establishing a new benchmark for compact, robust, and shallow photonic computing architectures.