Space-variant interconnections based on diffractive optical elements for neural networks: architectures and...

Area of Science:

• Optoelectronics
• Optical computing
• Artificial intelligence hardware

Background:

• Investigating optical architectures for fixed-connection multilayer neural networks is crucial for advancing AI hardware.
• Diffractive optical elements (DOEs) offer potential for complex interconnections but face challenges with crosstalk and system volume.
• Comparing fully connected and limited-fan-out architectures is essential for optimizing performance.

Purpose of the Study:

• To investigate and compare optical architectures for space-variant weighted interconnections using DOEs.
• To address high interconnection crosstalk in limited-fan-out architectures.
• To propose and verify a novel crosstalk reduction technique for DOE-based neural networks.

Main Methods:

• Comparative analysis of fully connected and limited-fan-out architectures based on propagation lengths, system volumes, connection densities, and crosstalk.
• Development of a modified DOE design procedure to rearrange reconstruction patterns and minimize noise.
• Simulation of a single-layer interconnection system with 128x128 input/output nodes and 5x5 fan-out connections.

Main Results:

• Limited-fan-out architectures offer higher node capacity within smaller system volumes.
• Limited-fan-out space-variant architectures exhibit significant interconnection crosstalk due to DOE reconstruction noise.
• The proposed modified DOE design procedure significantly reduces crosstalk.
• The technique also leads to reduced propagation length and overall system volume.

Conclusions:

• A novel crosstalk reduction technique for DOE-based optical neural networks has been developed and verified.
• This technique enhances the feasibility of limited-fan-out architectures by mitigating crosstalk.
• The optimized DOEs contribute to more compact and efficient optical interconnection systems for neural networks.