HoloGraph: All-Optical Graph Learning via Light Diffraction

TL;DR

HoloGraph tackles the challenge of graph-structured tasks in optical neural networks by introducing a free-space, all-optical graph learning framework that uses domain-specific message passing and trainable optical skip channels. It preprocesses inputs with PCA and PPRGo to capture neighborhood information, encodes features in a complex light wavefunction, and performs end-to-end training across six DONN layers with optical skips that mitigate information loss. On standard benchmarks like Cora-ML, Citeseer, and Amazon Photo, HoloGraph achieves competitive accuracies with significantly improved energy efficiency over CPU/GPU baselines, demonstrating the viability of scalable, energy-efficient optical graph learning. The work also provides thorough design-space analyses and discusses limitations and future directions toward monolithic integration and dynamic-graph processing in optical hardware.

Abstract

As a representative of next-generation device/circuit technology beyond CMOS, physics-based neural networks such as Diffractive Optical Neural Networks (DONNs) have demonstrated promising advantages in computational speed and energy efficiency. However, existing DONNs and other physics-based neural networks have mostly focused on exploring their machine intelligence, with limited studies in handling graph-structured tasks. Thus, we introduce HoloGraph, the first monolithic free-space all-optical graph neural network system. It proposes a novel, domain-specific message-passing mechanism with optical skip channels integrated into light propagation for the all-optical graph learning. HoloGraph enables light-speed optical message passing over graph structures with diffractive propagation and phase modulations. Our experimental results with HoloGraph, conducted using standard graph learning datasets Cora-ML and Citeseer, show competitive or even superior classification performance compared to conventional digital graph neural networks. Comprehensive ablation studies demonstrate the effectiveness of the proposed novel architecture and algorithmic methods.

Figures (7)

• Figure 1: Overview of all-optical multi-layer DONNs system.
• Figure 2: Illustration of the proposed HoloGraph network with 3 feature embedding layers and 3 prediction layers. The optical skip channel is connecting the input (layer 0) to the first prediction layer (layer 4).
• Figure 3: (a) The system architecture exploration for setups of optical skip connections in the proposed system. (b) Test accuracy w.r.t training epochs on Cora-ML under different optical skip connection setups in Table \ref{['tbl:optical_skip_setup']}.
• Figure 4: Accuracy and Loss convergence plots of HoloGraph on Cora, CiteSeer and Amazon Photo benchmarks.
• Figure 5: Confusion Matrix of HoloGraph classification result on three datasets.