Single-shot optical neural network.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computer Engineering

Background:

• Analog hardware offers a more efficient alternative to digital electronics for deep neural networks (DNNs).
• Previous analog DNN processors faced limitations in scalability and required model retraining, hindering practical application.
• The development of efficient hardware is crucial for advancing next-generation AI and machine learning.

Purpose of the Study:

• To present a scalable, CMOS-compatible analog DNN processor utilizing free-space optics and optoelectronics.
• To demonstrate the processor's effectiveness with standard DNN models and datasets without the need for retraining.
• To experimentally determine the throughput limitations imposed by optical bandwidth.

Main Methods:

• Developed a CMOS-compatible analog DNN processor employing free-space optics for input vector distribution.
• Utilized optoelectronics for static, updatable weighting and nonlinearity implementation.
• Tested the processor on MNIST, Fashion-MNIST, and QuickDraw datasets using standard fully connected DNNs.

Main Results:

• Achieved high classification accuracies: 95.6% (MNIST), 83.3% (Fashion-MNIST), and 79.0% (QuickDraw).
• Demonstrated single-shot-per-layer classification without preprocessing or retraining.
• Experimentally determined a fundamental throughput upper bound of approximately 0.9 exaMAC/s, limited by optical bandwidth.

Conclusions:

• The developed analog optical DNN processor offers a scalable and efficient solution for deep learning tasks.
• The system's compatibility with standard DNNs and datasets facilitates wider adoption of analog computing for AI.
• This technology paves the way for highly efficient, next-generation computing architectures by combining wide spectral and spatial bandwidths.