Optical processor for a binarized neural network.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computer Engineering

Background:

• Binarized neural networks (NNs) require efficient multiply-accumulate operations.
• Implementing both positive and negative weights is crucial for binarized NN performance.
• Existing hardware faces challenges in high-speed, low-power NN computation.

Purpose of the Study:

• To propose and experimentally demonstrate an optical processor for binarized neural networks (NNs).
• To achieve high-speed multiply-accumulate operations using optical methods.
• To leverage optical processing for enhanced data computing capabilities.

Main Methods:

• Utilized a dual-drive Mach-Zehnder modulator (DD-MZM) to implement positive and negative weights (+1 and -1) and perform multiplication.
• Employed dispersion-induced time delays for the accumulation operation.
• Integrated a photodetector (PD) for signal detection.

Main Results:

• Successfully demonstrated a proof-of-concept optical processor for binarized NNs.
• Achieved an acceleration speed of 32 giga floating point operations/s (GFLOPS) for a binarized convolutional neural network (CNN).
• Tested the processor on benchmark image classification tasks, validating its functionality.

Conclusions:

• The proposed optical processor efficiently implements multiply-accumulate operations for binarized NNs.
• The system offers high bandwidth and parallel processing capabilities.
• This technology holds significant potential for next-generation data computing applications.