Original title: Artificial Intelligence in a Photonic Temporal Processor
Youlve Chen, Jinlong Xiang, Yimin Hu, Yuchen Yin +12 · August 2026
Most optical neural networks use light to run artificial intelligence programs quickly and with little energy. However, older designs spread calculations out across physical space using tiny channels or mirrors. This setup locks the number of artificial brain cells, or neurons, to the amount of physical hardware built on the chip. To solve this, researchers moved the calculations entirely into the time domain. They used special thin-film lithium niobate devices as time lenses and masks. Spreading light out across time let nearby time-based signals interact. This allowed the team to calculate complex, high-order math using just a single input and output port. By adding optical feedback loops, they created fully programmable networks that can be trained directly in place. The system successfully tackled classification, image, and video generation tasks. It even ran an entirely analog generative pipeline using raw optical noise without any digital electronics.
Imagine you have a single conveyor belt in a factory. Normally, to assemble many different products, you need many separate assembly lines, each with its own set of machines. But if you instead send products one after another on the same belt and quickly reconfigure the machines between each product, you can handle many products with just one line. In this paper, the conveyor belt is time, and the machines are optical components that can be rapidly reconfigured to process different parts of the computation sequentially, so a single optical port can perform the work of many.
Current optical neural networks are limited by physical hardware—each neuron needs its own component. This work breaks that limit by using time instead of space, so a single optical port can handle many neurons. That means faster, more efficient AI hardware that can scale without adding more components, and it even works for generating images and videos entirely in analog, without digital conversion.
The abstract does not discuss integration with digital electronics for training or data input/output, nor does it quantify energy savings compared to electronic systems. It also does not address the physical size, cost, or long-term stability of the setup, or scalability to networks with millions of neurons beyond what was demonstrated.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Artificial Intelligence in a Photonic Temporal Processor — Youlve Chen, Jinlong Xiang, Yimin Hu et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.