Self-optimizing multichannel optical computing

TL;DR

The paper addresses the bottlenecks of conventional digital hardware by introducing a self-optimizing multichannel optical computing platform that processes RGB and other multichannel inputs directly in the optical domain. It combines a programmable multi-plane light conversion (MPLC) architecture with two gradient-free optimization strategies: Bayesian optimization of input-channel mixing and hardware-in-the-loop optimization using self-organized criticality to adapt phase masks. Across tasks including HAM10000, STL-10, Flowers-17, and Abalone, multichannel processing with random phase masks yields substantial accuracy gains (26–58 percentage points) over raw pixel baselines, with RGB generally outperforming grayscale by about 5–6 points; further gains (6–7 points) arise from the complementary optimization strategies. The work demonstrates a practical, scalable, and autonomous optical computing paradigm that can serve as a low-latency, energy-efficient feature extractor for real-world machine learning applications, with broad applicability to image, multi-modal, and structured data tasks.

Abstract

Optical computing offers ultrafast, energy-efficient alternatives to conventional digital processors, yet most implementations remain confined to single-channel processing, severely underutilizing light's information capacity. Here we demonstrate a self-optimizing multichannel optical computing architecture based on multi-plane light conversion that natively processes RGB images and structured numerical data throughout the optical domain. We introduce two complementary optimization strategies that enable autonomous performance adaptation without differentiable forward models. First, Bayesian optimization tunes channel mixing coefficients to minimize crosstalk and enhance feature separability at the input level. Second, a hardware-in-the-loop protocol based on self-organized criticality leverages avalanche dynamics to autonomously navigate the high-dimensional phase landscape, enabling the system to self-optimize through stochastic multi-scale perturbations. Across medical imaging, natural image classification, and regression tasks, multichannel processing with random phase masks improves accuracy by 26--58 percentage points over raw pixel baselines, with RGB systematically outperforming grayscale by 5--6 percentage points. Self-optimization strategies provide additional gains of 6--7 percentage points through autonomous adaptation at complementary system levels. Our work establishes self-optimizing multichannel optical computing as a practical platform for real-world machine learning applications.

Figures (5)

• Figure 1: Multichannel optical computing architecture and concept. (a) Experimental setup showing the spatial light modulator (SLM), planar mirror, and camera detection system. The laser beam undergoes multiple reflections to create cascaded phase transformations. (b) Conceptual diagram illustrating multichannel input encoding (RGB channels spatially multiplexed), phase mask cascade, and intensity detection followed by digital readout. (c) Example phase patterns displayed on the SLM active regions, demonstrating the programmable nature of the optical transformation.
• Figure 2: Single-channel and multichannel optical computing for HAM10000 dermatoscopic lesion classification. (a) Confusion matrix for raw grayscale inputs. (b) Confusion matrix for optically processed grayscale inputs. (c) LDA projection showing raw grayscale feature space. (d) LDA projection showing optically processed grayscale features with improved cluster separation. (e) Confusion matrix for raw RGB inputs. (f) Confusion matrix for optically processed RGB inputs, showing near-perfect classification. (g) LDA projection for raw RGB features. (h) LDA projection for optically processed RGB features, demonstrating tight, well-separated clusters across all seven diagnostic categories.
• Figure 3: Single-channel and multichannel optical computing for STL-10 natural image classification. (a) Confusion matrix for raw grayscale inputs showing poor discrimination across object categories. (b) Confusion matrix for optically processed grayscale inputs with substantial improvement. (c) LDA projection for raw grayscale features showing heavily overlapping clusters. (d) LDA projection for optically processed grayscale features with improved separation. (e) Confusion matrix for raw RGB inputs. (f) Confusion matrix for optically processed RGB inputs demonstrating strong category discrimination. (g) LDA projection for raw RGB features. (h) LDA projection for optically processed RGB features showing well-defined, separated clusters for all ten object categories.
• Figure 4: Input-space optimization via Bayesian channel mixing. Left: Schematic showing multichannel RGB data (R, G, B) undergoing learned linear mixing with optimized coefficients before optical encoding. (a) Confusion matrix for optically processed RGB inputs without channel mixing. (b) Confusion matrix after Bayesian optimization of mixing coefficients, showing improved diagonal dominance and reduced confusion between similar categories. (c) LDA projection for unmixed RGB optical features. (d) LDA projection after channel mixing optimization, demonstrating tighter, more separated clusters across all ten STL-10 categories.
• Figure 5: Hardware-space optimization via self-organized criticality for Oxford Flowers-17 classification. Representative flower images from the dataset are shown at top. (a) Confusion matrix for raw RGB pixel inputs processed by Ridge classifier (37% accuracy), showing poor discrimination across the 17 flower categories. (b) Confusion matrix after optical processing with random phase masks (74% accuracy), demonstrating substantial improvement from optical feature extraction alone. (c) Confusion matrix after SOC-based phase mask optimization (80% accuracy), showing further refinement with particularly strong diagonal dominance and reduced inter-class confusion.