BUPT review maps the path to optical convolution computing

A Beijing University of Posts and Telecommunications team has published a review that organizes optical convolution computing into two main approaches and lays out where the field is headed. The paper points to applications in signal processing, imaging and high-dimensional data, while flagging energy efficiency, scalability and full CNN support as the biggest hurdles.

Why it matters: - Convolution is the main workload in convolutional neural networks and can account for more than 80% of inference cost. - Optical systems could match convolution’s linear algebra in hardware and cut energy use in bandwidth-heavy computing. - The review frames hybrid opto-electronic systems as the most practical route to deployable optical convolution hardware.

What happened: - A team led by Prof. Kun Xu at the State Key Laboratory of Information Photonics and Optical Communications, Beijing University of Posts and Telecommunications, published a review titled "Optical Convolution Computation: Principle, Applications, Challenges" in Volume 2 of Intelligent Opto-Electronics on June 29, 2026. - The review splits optical convolution into two core methods: definition-based and theorem-based. - The paper also organizes the field by application area, data dimensionality and implementation style.

The details: - Definition-based optical convolution directly performs sliding-window multiply–accumulate operations in the physical domain. - In that approach, modulation through amplitude, phase or polarization sets the weights, and photodetection turns optical field superposition into accumulation. - The review lists dimension-interleaved schemes, microring resonator weight banks, coherent Mach–Zehnder interferometer meshes and spatial-projection architectures as representative designs. - Theorem-based optical convolution uses the convolution theorem to move the operation into the frequency domain. - That category includes classical 4f Fourier-optics processors, on-chip nanophotonic convolvers, ultracompact meta-imagers and frequency-comb-based spectral convolution schemes. - One silicon-photonic implementation in that line reportedly reached the TOPS regime in a single computing cell. - The review says optical convolution applications now span 1D radio-frequency and radar processing, 2D image classification and reconstruction, and 3D-plus tensor processing. - The paper cites demonstrations in video action recognition, multi-channel ECG analysis and biomedical diagnostics. - The review says optical convolution is moving from isolated operator demos toward CNN-complete accelerators.

Between the lines: - The field has momentum, but the hardware landscape remains fragmented across free-space 4f systems, microring banks and other photonic platforms. - Performance comparisons remain difficult because studies use inconsistent metrics and often omit interface and calibration costs. - The review’s emphasis on hybrid systems reflects a practical limit: photonics can do the linear math, while electronics still handles memory, nonlinear activation and control.

What's next: - The biggest technical gap is CNN-complete support, including stride, padding, multi-channel processing and cascaded layers. - The review says the field needs standard full-chain energy-efficiency metrics that include DAC and ADC interfaces, modulator drivers and calibration overhead. - Scalable integration will also require better control of phase drift, thermal crosstalk and fabrication variability. - Early deployment is most likely in real-time edge intelligence and ultra-high-throughput signal processing, where latency and bandwidth matter most.

The bottom line: - Optical convolution is no longer just a lab concept. - The next step is turning promising photonic operators into stable, scalable systems that can run full neural-network workloads.