Optical system performs feature extraction with unprecedented low latency

Many modern artificial intelligence (AI) applications, such as surgical robotics and real-time financial transactions, rely on the ability to rapidly extract key features from streams of raw data. Currently, this process is bottlenecked by traditional digital processors. The physical limitations of traditional electronics prevent the reduction in latency and increase in throughput required by emerging data-intensive services.

The answer may lie in harnessing the power of light. Optical computing, or the use of light to perform demanding calculations, has the potential to significantly accelerate feature extraction. In particular, optical diffraction operators, which are plate-like structures that perform calculations as light propagates, hold great promise due to their energy efficiency and parallel processing capabilities.

However, actually increasing these systems to operating speeds above 10 GHz remains a technical challenge. This is mainly due to the difficulty of maintaining stable coherent light required for optical calculations.

To tackle this problem, a research team led by Professor Hongwei Chen from Tsinghua University in China has devised a surprising solution. As reported in Advanced Photonics Nexus, they developed an optical feature extraction engine (called OFE2) that performs optical feature extraction for a variety of practical applications.

The core innovation is in the OFE2 data preparation module. Providing high-speed, parallel optical signals to optical cores operating in a coherent environment is extremely challenging because the use of fiber-based components for power division and delay introduces strong phase perturbations. The team solved this problem by developing an integrated on-chip system with an adjustable power splitter and a high-precision delay line.

This module effectively deserializes the data stream by sampling the input signal into multiple stable parallel branches. Additionally, the tunable integrated phased array allows OFE2 to be reconfigured as needed.

Once the data is prepared, the light waves pass through a diffraction operator. This process can be modeled mathematically as a matrix-vector multiplication that performs feature extraction. The key to this operation is how the diffracted light forms a focused “bright spot” at the output, which can be partially deflected toward a specific output port by adjusting the phase of the parallel input light. This movement and corresponding change in output power allows OFE2 to effectively capture features related to changes in the input signal over time.

Operating at a rate of 12.5 GHz, OFE2 can perform a single matrix-vector multiplication in less than 250.5 ps. This is the lowest latency of any similar optical computing implementation.

“We believe this work will advance integrated optical diffraction computing and provide an important benchmark beyond 10 GHz rates in real-world applications,” said Chen.

The research team successfully demonstrated the functionality of the proposed system across a variety of tasks. Regarding image processing, OFE2 was able to extract edge features from the input image and create two complementary “relief and engraving” feature maps.

Features generated by OFE2 improved image classification performance and improved pixel accuracy for semantic segmentation (e.g., organ identification in computed tomography scans). In particular, our AI network using OFE2 requires fewer electronic parameters than the baseline, proving that optical preprocessing can lead to a lightweight and efficient hybrid AI system.

Furthermore, the team obtained similar results in a digital trading task where OFE2 received time-series market data and suggested profitable trading actions based on an optimized strategy. In this task, traders input real-time price signals into OFE2. After pre-training, an optimally configured OFE2 produces output signals that can be directly translated into buy or sell actions through a simple decision-making process, resulting in stable profitability. The entire process runs at the speed of light, which significantly reduces waiting times and allows you to reap profits with minimal delay.

Taken together, these results point to a new paradigm in which the most demanding computational loads are shifted from power-hungry electronics to ultra-fast, low-energy photonics, leading to a new generation of real-time decision-making AI systems.

“The advances demonstrated in our research will increase the speed of integrated diffraction operators and provide support for compute-intensive services in areas such as image recognition, medical assistance, and digital finance. We look forward to working with partners with data-intensive computational needs,” concluded Chen.