Artificial intelligence (AI) is moving toward real-time decisions in changing environments such as self-driving cars reacting to traffic or instruments handling continuous data. Each prediction depends on current inputs and earlier states, so hardware must adapt while it computes.
Photonic processors use light for high bandwidth and low energy use, yet most existing designs rely on fixed optical paths interrupted by electronic controls, limiting on-the-fly adaptation.
Researchers have developed a Temporally Plastic Photonic Processor, or TPPP. This light-based features places time-dependent optical operators. Optical feedback and rapid modulation allow computation to evolve with the data stream instead of returning constantly to electronic memory.
The design is inspired by biological plasticity. A slow kernel maintains long-term stable transformations. A fast kernel applies brief pulse-level changes. A feedback kernel routes the optical output back to the input through a recursive delay. Together they let successive light pulses experience different weights while preserving a stable structure.
Hardware platforms and experimental results
The architecture was realized on a low-loss silicon nitride chip for the all-optical path and a silicon platform with integrated semiconductor amplifiers for fast electro-optic modulation.
In a linear demonstration it predicted moisture content in flour from near-infrared spectra reduced to eight principal components. Matrix inversion reformulated as Richardson iteration, an iterative numerical method, rose from about 90 percent accuracy with a static approach to 95 percent with dynamic correction; classification improved from 88 to 95 percent.
A second test used a time-adaptive recurrent neural network for vehicle control. Four sensor inputs - speed and obstacle distance - produced throttle and steering commands. Dynamic modulation reduced trajectory error to less than 5 percent, about ten times better than a fixed network.
These results show that temporal plasticity can make photonic computing an adaptive substrate for evolving data. At 40 gigabits per second and 810 femtojoules per operation, the design projects up to sixteen times higher energy efficiency and one hundred times lower delay than advanced electronic processors. Further scaling could support real-time sensing and autonomous systems where adaptation must match the speed of arriving data.
This research is published in eLight.