Author: Petr Šícho, Supervisors: Niyazi Ulas Dinç, Mustafa Yildirim, Advisor: Prof. Christophe Moser (EPFL, LAPD)
Objective
Improve the scalability and performance of Optical Neural Networks (ONNs) by replacing direct parameter optimization with an implicit neural representation.
Methodology
- Implemented a SIREN (Sinusoidal Representation Network) to generate phase modulation parameters.
- The model maps 2D spatial coordinates to smooth scale and bias masks for each optical layer (Spatial Light Modulator).
- Evaluated against baseline optical models using the CIFAR-10 image classification dataset.
Key Results
- Parameter Efficiency: Achieved a 16x reduction in learnable parameters for high-resolution optical setups (e.g., reducing 262,144 parameters to just 16,384 for a 512x512 layer).
- Performance Gain: Boosted classification accuracy by ~4% on deeper architectures (up to 53.56% on a 6-layer setup).
- Scalability: Successfully overcame the performance/overfitting plateau, allowing for deeper, more reliable optical networks.