Artificial intelligence is moving from static, one-shot inference toward real-time decisions in environments that change continuously. For instance, autonomous vehicles must respond to traffic and obstacles, scientific instruments must interpret data streams as they arrive, and generative models must process long temporal contexts. In these tasks, the prediction at one instant depends not only on the current input, but also on earlier states and previous outputs. Hardware therefore needs more than speed: it must adapt while it computes.
Conventional electronic processors can update weights and recurrent states, but doing so in real time is expensive. Frequent memory access, data movement and clock synchronization create latency and energy costs, especially for sequential models such as recurrent neural networks, Transformers and state-space models. Photonic processors offer high bandwidth, low latency and excellent energy efficiency, yet many existing photonic accelerators are designed around fixed transformations. Their optical signal flow is often interrupted by electronic control and memory operations, which limits their ability to adapt on the fly.
In a recent article, teams of Professor Qixiang Cheng from the University of Cambridge, Professor Lu Fang from Tsinghua University and, together with collaborators from Microsoft Research, GlitterinTech, Ghent University-IMEC, Huazhong University of Science and Technology, and the University of Bath, reports a Temporally Plastic Photonic Processor (TPPP). The processor introduces time-dependent optical operators directly into the photonic domain. Instead of repeatedly returning to electronic memory for every update, the device uses optical feedback and fast modulation to let computation evolve with the input stream.
The core idea is inspired by biological plasticity across multiple timescales. A slow reconfigurable kernel provides stable long-term transformations. A fast dynamic kernel changes on short timescales and applies pulse-level modulation. A feedback kernel routes the optical output back to the input through a recursive optical delay memory. Together, these three function kernels allow different optical pulses in the same loop to experience different instantaneous weights while preserving a stable computational backbone.
The researchers implemented the architecture on two complementary photonic platforms. A low-loss silicon nitride chip demonstrated the all-optical computing link, while a compact silicon-on-insulator platform incorporated micro-transfer-printed InP semiconductor optical amplifier arrays for fast electro-optic modulation. After calibration, the processor achieved high single-pass precision, with an average dot-product error of 0.175 least significant bits and an effective number of bits of about 5.78 across eight INT8-encoded output channels. The optical loop showed power fluctuations below one percent, supporting stable multi-iteration operation.
The authors first tested the TPPP in a linear task: spectrum-based moisture prediction in flour. Near-infrared transmission spectra with 1,899 wavelength channels were compressed to eight principal components and used in a ridge-regression model. The demanding matrix inversion step was reformulated as Richardson iteration and implemented on the photonic hardware. A data-driven dynamic correction compensated for path-dependent loss, device nonidealities, environmental noise and channel skew. As a result, matrix inversion accuracy improved from about 90 percent with a static model to about 95 percent with dynamic adaptation, while a wet-dry classification example improved from 88 percent to 95 percent accuracy.
The second demonstration addressed nonlinear sequential decision-making. The team designed a time-adaptive recurrent neural network (TA-RNN), for autonomous-driving control and co-designed it with the photonic processor. The model received four sensor inputs: vehicle speed and obstacle distances in the front, left and right sectors. It then predicted throttle and steering commands. By dynamically modulating diagonal operators inside the recurrent model, the TPPP introduced time-dependent capacity beyond a fixed RNN. In the navigation test, the TA-RNN achieved less than 5 percent trajectory prediction error, about an order-of-magnitude improvement over a standard RNN baseline.
These demonstrations show that temporal plasticity can make photonic computing more than a fast static matrix engine. It can become an adaptive computing substrate for data streams that evolve in time. Operation-matched analysis at a validated 8 by 8 INT8 operating point of 40 gigabits per second and 810 femtojoules per operation projects up to 16 times higher per-operation energy efficiency and up to 100 times lower intrinsic single-pass compute delay than advanced electronic processors.
The researchers expect that larger matrices, lower-loss optical routing, better noise control and heterogeneous multi-chip integration will further expand the capability of the platform. If scaled successfully, temporally plastic photonic processors could provide a hardware route for real-time sensing, edge intelligence, autonomous control and other applications where adaptation must happen as quickly as the data arrive.
eLight
Temporally plastic photonic processor for real-time adaptive computing