Analogue optical pattern recognition for cross-correlational CNN.

Area of Science:

• Optoelectronics
• Machine Learning Hardware
• Computational Science

Background:

• Convolutional Neural Networks (CNNs) rely on computationally intensive 2D convolutions.
• Traditional von Neumann architectures face limitations in processing power and time for these tasks.
• Pattern recognition demands significant computational resources, hindering real-time applications.

Purpose of the Study:

• To propose an analogue optical hardware system for enhancing CNN efficiency.
• To leverage light wave properties for faster 2D convolutional operations.
• To overcome the power and time limitations of current computing architectures for machine learning.

Main Methods:

• Utilizing analogue optical hardware for CNN forward propagation tasks.
• Employing light wave properties to perform 2D convolutional operations.
• Simulating the proposed system using MATLAB and COMSOL for validation.

Main Results:

• Demonstrated potential for significant improvements in CNN processing speed.
• Showcased the efficiency of optical wave operations for tasks like 2D Fourier transforms.
• Validated the feasibility of the proposed optical approach through simulations.

Conclusions:

• The proposed analogue optical system offers a path towards more efficient machine learning hardware.
• This approach can overcome the computational bottlenecks of current CNN implementations.
• Future work includes enabling full CNN training and developing commercial hardware.