Optical Neural Networks for Image Classification: A GPU-Accelerated Two-Stage Architecture

We present a novel optical neural network architecture that achieves state-of-the-art performance on the MNIST digit classification task by simulating light propagation through learnable optical elements. Our approach combines two cascaded optical stages with learnable amplitude and phase modulation, achieving 98.657 % accuracy on the Kaggle digit recognizer leaderboard with only 1000 epochs of training completed in 20 minutes. The architecture leverages GPU-accelerated Fast Fourier Transforms (FFTs) to efficiently simulate optical diffraction and interference phenomena, demonstrating that physics-inspired neural architectures can achieve competitive performance with remarkable training efficiency.