Photonic CNN achieves high accuracy and energy efficiency

Researchers have developed a fully photonic convolutional neural network (PCNN) capable of executing image classification tasks entirely within the optical domain. This novel architecture integrates convolution, max-pooling, nonlinear activation, and fully connected layers using optical components like Mach-Zehnder Interferometer meshes and microring resonators. While achieving 94.49% accuracy on the MNIST dataset, a digital twin with ex situ pre-training reached 97.45% accuracy. The PCNN demonstrates significant potential for energy efficiency, with an estimated 220 to 330 times greater efficiency than state-of-the-art electronic GPUs for single-image inference, alongside a low inference latency of 843 ns. AI

IMPACT Offers a path toward significantly more energy-efficient and faster AI inference hardware.

RANK_REASON Academic paper detailing a novel hardware architecture for AI. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• CMOS
• graphics processing unit
• Mach–Zehnder interferometer
• MNIST database
• Photonic convolutional neural network
• Saurabh R Sinha

• paper
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →