Scalability of On-chip Diffractive Optical Neural Networks

TL;DR

This paper investigates the scalability limitations of on-chip diffractive optical neural networks, revealing they struggle with complex tasks and are limited to classifying only a few categories due to fundamental design constraints.

Contribution

It provides a thorough analysis of the scalability challenges in on-chip diffractive optical neural networks, highlighting their fundamental limitations compared to electronic neural networks.

Findings

Limited to classifying 3-4 classes despite optimizations

Performance degradation with increasing classification categories

Design parameters have minimal impact on scalability

Abstract

This short report focuses on the scalability challenges of the on-chip diffractive optical neural networks. It addresses an emerging gap in the literature, specifically around the limitations and challenges of scaling optical neural networks on a chip. A thorough investigation of diffractive optical neural networks provides evidence that such networks are not capable of performing complex tasks and exhibit significant performance degradation as the number of classification categories increases. Despite optimizations, these networks classify only 3-4 classes, suggesting fundamental limitations in their computational scale. The inherent scalability challenges in these systems are underscored by the fact that the design parameters, such as the number of diffractive layers, the number of neurons per layer, and the inter-layer distances, cannot substantially change the performance. Therefore, the on-chip diffraction-based approach provides a limited number of controllable degrees of freedom compared to electronic neural networks, restricting the complexity of functions an on-chip diffractive neural network can learn.