Electrical and Computer Engineering

Lighting the path to faster, smarter AI

May 2025 | Highlights Community

Follow CDE

PDF Download

Image above was created with the assistance of DALL·E 2

Bringing non-linearity into the light, a new photonic accelerator allows AI to compute and respond entirely in the optical domain.

As the world thirsts for faster, ever more efficient artificial intelligence (AI), the limits of silicon are beginning to rear their heads. Power-ravenous processors. Clogged interconnects. Algorithms that outpace the hardware meant to run them. It’s a bottleneck that’s hard to ignore — one that puts a cap on tech giants, and one that researchers are racing to break through.

Professor Aaron Thean led a team to develop a reconfigurable, light-responsive solution that brings true nonlinearity into the photonic fold.

Light may offer the way forward. Photonic computing, where data is processed using light rather than the usual electrons, promises blistering speed, minimal energy loss and vast computing parallelism. In fact, integrated photonics can already perform key AI tasks like matrix multiplications with astonishing efficiency.

But there is a missing piece of the puzzle. One that has kept photonic neural networks from stepping fully into the spotlight.

Without a way to mimic the non-linearity of the brain — how biological neurons respond to signals in nuanced, varied ways — photonic systems remain in the dark. They can add and multiply, but they cannot decide. Processing is possible, but interpreting information is a no-go. And that crucial function, it turns out, has been incredibly challenging to implement.

Research led by Professor Aaron Thean from the Department of Electrical and Computer Engineering, College of Design and Engineering (CDE), National University of Singapore, has shone a light on the path forward, introducing a reconfigurable, light-responsive solution that brings true non-linearity into the photonic fold.

Breaking the linear trap

In a typical artificial neural network, two elements work in tandem. First, weighted sums are computed — multiplying inputs by adjustable values and summing the results. Then comes the non-linear activation function, which introduces flexibility and complexity to model the messiness of the real world, where relationships rarely follow straight lines. It’s this non-linear step, which determines how strongly a neuron should respond, that transforms a simple model into one capable of learning.

“Photonic neural networks have the first part down,” says Prof Thean. “Integrated devices like Mach–Zehnder Interferometer (MZI) meshes can perform the weighted matrix multiplications at blistering speeds, using light as the medium. But introducing non-linearity — in situ, without converting optical signals back into electrical form — has been a long-standing challenge.”

"By embedding these ORS switches into a hybrid structure alongside MZIs and low-power control units, we created a reconfigurable nonlinear activation accelerator."

Existing approaches have made some headway. Some incorporate lasers or photodetectors with built-in non-linear behaviours, but these tend to be bulky, power-hungry or limited to fixed responses that can’t adapt from task to task.

The CDE researchers approached the problem from a different angle. Instead of forcing existing devices to bend to non-linear rules, they designed a new kind of component: a light-sensitive Opto-Resistive RAM (ORS) Switch built from solution-processed molybdenum disulfide. This two-dimensional material responds to incoming light by abruptly changing its electrical resistance, a behaviour that can be precisely tuned and repeated over many cycles. They published their findings in the journal Light: Science & Applications.

"By embedding these ORS switches into a hybrid structure alongside MZIs and low-power control units, we created a reconfigurable nonlinear activation accelerator."

“By embedding these ORS switches into a hybrid structure alongside MZIs and low-power control units, we created a reconfigurable non-linear activation accelerator,” explains Prof Thean. “It takes in optical signals, converts them into a voltage-driven non-linear response and feeds that signal back into the photonic circuit — completing the neural computation loop without ever leaving the optical domain.”

When tested in simulation on a standard image recognition task using the MNIST handwritten digit dataset, the system delivered a classification accuracy of 91.6%, matching performance benchmarks, while using 20 times less energy and occupying 40% less space than previous photonic architectures.

Towards smarter light

One of the key advantages of the ORS-based accelerator is its reconfigurability. Unlike other non-linear optical elements that are hardwired to one behaviour, the researchers’ system can be programmed to replicate different activation functions, such as ReLU, sigmoid or softplus, depending on the task at hand. This adaptability lays the groundwork for a broader class of AI models running efficiently on light-based platforms.

The team’s compact, energy-efficient design also makes photonic AI hardware practical for real-world deployment, especially in applications where speed and power budgets are critical: autonomous systems, real-time data processing or edge AI devices.

Looking ahead, the team aims to integrate lithium niobate photonic components with the silicon photonics platform, combining the maturity and high performance of silicon photonics technology with the enhanced functionality of lithium niobate. This integration is essential for enabling the next generation of advanced AI applications.

You heard it here first: AI, at the speed of light!