Optical logic convolutional neural network

Area of Science:

• Optical Computing
• Artificial Intelligence
• Machine Learning Hardware

Background:

• Optical computing offers high-speed potential but faces challenges with analog methods and digital configurations.
• Current optical digital computing lacks flexibility for applications like AI inference.
• Environmental perturbations and reliance on converters limit optical analog computing.

Purpose of the Study:

• To introduce and demonstrate an optical logic convolutional neural network (OLCNN) for efficient AI computation.
• To overcome limitations of existing optical computing paradigms for AI tasks.
• To pioneer a logic-driven approach for optical hardware in artificial intelligence.

Main Methods:

• Proposed and demonstrated an optical logic convolutional neural network (OLCNN) architecture.
• Implemented optical logic convolutional operators (OLCOs) of varying sizes (1x3, 2x2, 3x3).
• Validated OLCOs for pattern generation, image edge extraction, and MNIST dataset classification.

Main Results:

• Achieved high-speed optical computing at 20 Gbit/s with a 1x3 OLCO.
• Successfully performed image edge extraction using a 2x2 OLCO.
• Attained 95.1% average test accuracy for four-class classification on MNIST using a 3x3 OLCO within an OLCNN.

Conclusions:

• The proposed OLCNN offers a high-speed, energy-efficient solution for AI hardware.
• Synergizing optical logic devices with neural networks creates a new paradigm for optical computing.
• This logic-driven approach advances the development of optical hardware for artificial intelligence applications.