Optical Neural Networks (ONNs) are often described as the "holy grail" of high-speed, low-power AI. By using light instead of electricity, these systems can perform massive matrix-vector multiplications—the backbone of deep learning—at the speed of light with almost zero energy consumption. However, a fundamental physical hurdle has long kept these systems in the lab: light is "too polite."
In most media, light waves propagate linearly, meaning they pass through each other without interacting. Deep learning, however, requires nonlinearity. Without nonlinear activation functions like ReLU or Sigmoid, a neural network is just a single linear transformation, no matter how many layers it has. To solve this, researchers are finding ways to synthesize nonlinearity directly in the optical domain, moving beyond the slow and power-hungry process of converting light back into electricity.
The Bottleneck of Electronic Conversion
Traditionally, photonic chips handle the linear math (convolutions and matrix multiplications) and then pass the signal to a digital electronic processor to apply the activation function. As Florent Krzakala and Sylvain Gigan point out in Artificial intelligence: From electronics to optics, these frequent Optical-Electrical-Optical (O-E-O) conversions create a "data transfer bottleneck."
The energy saved by using light for math is often lost during these conversions. To realize the full potential of photonic computing, the signal must stay in the optical domain from input to output.
Exploiting Material Physics for Optical Neurons
One way to create an "optical neuron" is to use materials that respond nonlinearly to light intensity. When light is intense enough, it can change the physical properties of the medium it travels through, which in turn affects the light itself.
- Saturable Absorption: Albert Ryou and colleagues demonstrated this using thermal rubidium atoms. As light intensity increases, the atoms become "saturated" and stop absorbing photons, effectively creating a thresholding effect similar to a neural activation. Thomas D Barrett used a similar principle of saturable absorption to implement backpropagation entirely with light.
- Laser Dynamics: Semiconductor lasers can be "pushed" into nonlinear regimes. Guanting Liu and Cheng Wang showed that by injecting light into a laser, they could mimic a ReLU-like activation function. Similarly, Alexandre R. S. Romariz implemented the FitzHugh-Nagumo model, mimicking biological spiking neurons at speeds far beyond software simulations.
Synthesizing Nonlinearity with Linear Optics
Perhaps the most counterintuitive breakthrough is the discovery that you can achieve nonlinear processing using only linear optical components. This approach, known as nPOLO (nonlinear Processing with only Linear Optics), was pioneered by Mustafa Yildirim and Niyazi Ulas Dinç in Nonlinear Processing with Linear Optics.
The intuition behind nPOLO is based on recurrent scattering. If you pass light through a scattering medium once, it's a linear operation. But if you "re-embed" the input data multiple times—for example, by letting light bounce repeatedly inside a cavity—the resulting interference patterns synthesize high-order polynomial transformations.
Kyungduk Kim and Hui Cao leveraged this in Nonlinear optical encoding enabled by recurrent linear scattering, achieving massive 3000:1 data compression for tasks like pedestrian detection. By using the structure of the system rather than the properties of the material, these researchers avoid the need for high-power pulsed lasers, allowing nonlinear computing with low-power, continuous-wave light.
Computing in the Propagation Environment
The quest for nonlinearity isn't limited to chips; it extends to the very air around us. Kyriakos Stylianopoulos and George C. Alexandropoulos have proposed Metasurfaces-Integrated Neural Networks (MINNs).
In this framework, the wireless channel is transformed into a computational layer. By placing "Stacked Intelligent Metasurfaces" (SIM) in the environment, the physical propagation of electromagnetic waves performs the inference. While wave-domain computation is inherently linear, integrating these surfaces with passive nonlinear components allows edge devices to offload complex AI tasks to the environment itself, slashing energy consumption.
The Path to All-Optical Deep Learning
Despite these advances, challenges remain. Many current systems are still "shallow" or rely on digital backends for final classification. Jinlin Xiang and Arka Majumdar note that the sim-to-real gap—where fabrication errors ruin simulated performance—is a major hurdle.
However, the shift toward in-situ training, as demonstrated by Sunil Pai in Experimentally realized in situ backpropagation, suggests a future where photonic chips learn to account for their own physical imperfections. By combining structural nonlinearity (like nPOLO) with high-speed material effects (like the Kerr effect), the field is moving toward fully integrated, light-speed neural networks that can finally match the accuracy of their digital counterparts.