Coherent correlator design analysis for the implementation of deep learning networks.

Area of Science:

• Optoelectronics
• Optical Computing
• Signal Processing

Background:

• Optical signal processing presents an alternative to reduce power consumption compared to graphical processing units (GPUs).
• Significant challenges exist in optical systems, including higher noise levels and difficulties in implementing operations like bias addition in free-space processors.

Purpose of the Study:

• To analyze a proposed optical design for implementing convolutional layers using a camera and electronic processing.
• To evaluate the performance of the proposed design against an ideal system and a comparable architecture.
• To investigate the impact of speckle noise and propose mitigation strategies.

Main Methods:

• Simulations were conducted to compare the performance of the proposed camera-based optical design.
• The system's performance was benchmarked against an idealized, physically unrealizable system.
• An analysis of speckle noise impact and potential reduction techniques was performed.

Main Results:

• The proposed optical design shows potential for efficient convolutional layer implementation.
• Speckle noise was identified as a significant factor affecting system performance.
• Methods for mitigating speckle noise were investigated and proposed.

Conclusions:

• The camera-based optical processing design offers a viable approach to reduce power consumption in computing.
• Addressing noise, particularly speckle noise, is crucial for the practical implementation of optical convolutional layers.
• Further research into noise reduction techniques can enhance the feasibility of optical computing systems.