Four-quadrant optical matrix-vector multiplication machine as a neural-network processor

Area of Science:

• Optoelectronics
• Artificial Intelligence
• Computer Science

Background:

• Optical processors offer parallelism for neural networks but often lack bipolar neuron output support.
• Existing systems are limited to two-quadrant multiplication, hindering network capabilities and learning rates.

Purpose of the Study:

• To propose and demonstrate an opto-electronic four-quadrant matrix-vector multiplier.
• To enable bipolar neuron outputs for enhanced feed-forward neural network performance.

Main Methods:

• Development of an opto-electronic system for four-quadrant matrix-vector multiplication.
• Utilizing common commercial components for practical implementation.

Main Results:

• Successful demonstration of a four-quadrant matrix-vector multiplier.
• Experimental validation of the system's utility for feed-forward neural network recall and learning.

Conclusions:

• The proposed opto-electronic multiplier provides a novel, useful, and reliable method for four-quadrant multiplication.
• This approach significantly advances the capabilities of optical neural networks for training and recall.