Holographic neural network tackles multiple AI jobs simultaneously with a single beam of light

Artificial intelligence has transformed nearly every corner of modern technology, but the electronic processors that power it are straining under the load. As models grow larger and inference demands climb, conventional computing architectures are running into hard physical limits on energy consumption, latency, and bandwidth. Light, which moves information faster and with far less heat than electrons flowing through silicon, has long been touted as a way out of this impasse. Now a research collaboration led by China Information Communication Technologies Group Corporation (CICT), together with Peng Cheng Laboratory, Wuhan University, and Nanyang Technological University, has unveiled a photonic computing design that pushes the concept in a striking new direction: a single optical system that can run several artificial intelligence tasks at the same time, without changing hardware, swapping lasers, or reconfiguring a single component.

The work, published in the peer-reviewed journal Opto-Electronic Technology as “Neural holography enabled multi-beam steering for multi-task classifier,” centers on a device class known as a Diffractive Optical Neural Network, or DONN. In a DONN, neural network computations are carried out not by transistors but by the physics of light propagation itself. As a wavefront passes through carefully engineered diffractive layers, each point of the wave interferes with every other point, performing what amounts to a massive parallel matrix multiplication in a single pass of light. The result is information processing at the speed of light propagation, with energy costs that are orders of magnitude below those of comparable electronic operations.

DONNs have shown impressive promise for ultra-fast, low-energy AI inference, but they have carried a persistent limitation. To make one optical platform handle multiple different tasks, such as recognizing handwritten digits and classifying photographs of clothing at the same time, researchers have typically had to resort to wavelength multiplexing, polarization multiplexing, or mechanical hardware reconfiguration. Each approach demands additional optical components, precise laser tuning, or moving parts that must be realigned whenever the task mix changes. That translates directly into higher hardware costs, greater system complexity, and delicate alignment procedures that make practical deployment difficult. The question of how to accomplish many tasks within one unified, compact optical architecture has therefore become a central challenge in the field of optoelectronic computing.

The CICT-led team, working within the State Key Laboratory of Optical Communication Technologies and Networks (OCTN) and the National Information Optoelectronics Innovation Center (NOEIC), answered that challenge by borrowing a technique from an unexpected quarter: computer-generated holography, the same family of methods used to shape and steer laser beams for displays and optical trapping. By integrating neural holography into an optoelectronic neural network, the researchers constructed an end-to-end Holographic Neural Network (HNN) built around a single phase-only spatial light modulator, or SLM. This one device plays two distinct roles at once. At the input layer, it encodes electronic information onto light by imposing phase patterns on an incoming wavefront. In its computing role, it acts as a diffraction-based modulation layer, sculpting the light as it propagates through the optical system.

The elegance of this design lies in what it eliminates. Because the single SLM handles both input encoding and diffraction modulation, there is no need for the cascades of physical diffractive layers that conventional DONNs require, and no need for wavelength or polarization multiplexing to separate tasks. Instead, the system performs intelligent beam splitting and steering: the holographic layer directs multiple independent wavefronts through the same optical path, each carrying a different task’s computation, and each landing on a different region of the camera sensor at the output. The researchers describe the effect as “one brain for multiple uses” — a unified optical architecture that behaves like several distinct neural networks simultaneously, while remaining physically simple enough to align and operate reliably.

Training such a hybrid system demands careful coordination between the electronic and optical domains. The team adopted an end-to-end optoelectronic joint training framework in which the input encoding and the holographic diffraction modulation are optimized together, rather than being designed and tuned separately. The process fully exploits the nonlinear characteristics of the electrical-to-optical conversion that occurs at the input layer, a detail the researchers identify as important to overall performance. Light propagation then completes the complex parallel computations, and the camera at the far end directly reads out the classification results, with no intermediate electronic processing required before the answer appears. A notable practical consequence of the system’s low optical complexity is that offline simulation training aligns closely with experimental inference results, meaning networks can be trained digitally and deployed optically with minimal performance degradation — a long-standing pain point for diffractive photonic processors.

The multi-task capability itself rests on a scheme the authors call phase multiplexing, inspired by the cross-stitch architecture known from multi-task machine learning. Rather than assigning each task its own dedicated wavelength, polarization, or physical hardware, the holographic scheme superimposes task-specific phase patterns onto a single modulated wavefront, and a corresponding joint training algorithm learns how to allocate the optical resources among them. Crucially, the algorithm introduces a task correlation factor during error backpropagation. This factor governs how strongly the learning of one task influences another, and tuning it effectively reduces crosstalk — the unwanted interference that occurs when multiple computations share the same optical channel.

The experimental results highlight why that control matters. When two related tasks are trained together, such as recognizing two different sets of handwritten digits, the multi-task network actually surpasses the accuracy of comparable single-task networks, demonstrating that shared optical computation can be genuinely synergistic rather than merely tolerant of interference. For unrelated task types, such as classifying handwritten digits alongside fashion item images, optimizing the task correlation factor suppresses the mutual interference that would otherwise degrade performance. The team also demonstrated that the approach scales: a single holographic neural network simultaneously carried out four distinct recognition tasks while maintaining a classification accuracy of approximately 90 percent, all within the same fixed optical architecture.

For the broader field, the significance is twofold. First, the work removes one of the main engineering obstacles that has kept diffractive optical neural networks confined to laboratories: the need for reconfigurable or multiplexed hardware to serve multiple purposes. A platform that performs several AI inference jobs concurrently through pure beam shaping, with no moving parts and no laser tuning, points toward optical processors that could sit alongside electronic accelerators in data centers, handling high-throughput, low-latency inference workloads such as image classification at a fraction of the energy cost. Second, the marriage of neural holography with optoelectronic joint training provides a general recipe for squeezing more capability out of minimal optical hardware, which is precisely the kind of resource efficiency that photonic computing needs to become commercially competitive.

The research emerged from the Optoelectronic Devices Research Department of OCTN under CICT, a group dedicated to basic and applied research in computational optics and holography, computing-oriented advanced optoelectronic devices, and intelligent optical wireless communications. The team sees the holographic multi-task classifier as foundational for next-generation intelligent beam control and optoelectronic AI information processing technologies. If diffractive photonic processors are to graduate from demonstrations to deployments, they will need architectures that are simple to build, simple to align, and versatile enough to justify their cost. This holographic neural network, which turns a single light-shaping device into a multi-task AI brain, offers a compelling answer to all three demands at once — and suggests that the future of fast, efficient machine intelligence may be written not in silicon, but in the phase of light itself.

Subject of Research: Holographic diffractive optical neural networks for simultaneous multi-task AI inference

Article Title: New holographic neural network uses light to perform multiple AI tasks at once

Article References: New holographic neural network uses light to perform multiple AI tasks at once. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: photonic neural network, diffractive optical neural network, holography, spatial light modulator, multi-task learning, optical computing, AI inference, beam steering, phase multiplexing, optoelectronics, pattern classification, low-energy computing