Optically excited synapse for neural networks.

Area of Science:

• Optoelectronics
• Artificial Intelligence
• VLSI Design

Background:

• Neural networks heavily rely on matrix multiplication, a computationally intensive task.
• All-optical logic devices are currently underdeveloped, limiting purely optical neural network approaches.
• Existing electronic neural networks face limitations with current synapse implementations.

Purpose of the Study:

• To investigate optically addressed neural networks compatible with Very-Large-Scale Integration (VLSI).
• To propose and evaluate an optical matrix multiplier circuit using photodiodes for synapse implementation.
• To compare the performance of the proposed optical approach with electronic matrix multipliers.

Main Methods:

• Designed and analyzed an optical synapse using back-to-back photodiodes, leveraging their linear and nonlinear regions.
• Explored various optical addressing methods for the synapse matrix, including static (shadow mask) and dynamic (light valve, scanned laser).
• Compared the proposed optical matrix multiplier with an electronic analog-voltage vector matrix multiplier using MOSFETs in CMOS VLSI.

Main Results:

• The photodiode synapse exhibits suitable linear (for multiplication) and nonlinear (for neural networks) characteristics.
• Optical addressing allows for adjustable synapse strength via optical power, with anticipated microsecond settling times.
• A static, shadow mask-addressed optical neural network is identified as a particularly promising architecture.

Conclusions:

• Optically addressed neural networks offer a viable alternative to purely electronic implementations, especially when leveraging VLSI.
• The proposed photodiode-based optical synapse is effective for matrix multiplication and programmable neural network functions.
• Static optical neural networks, particularly those using shadow mask addressing, present a promising direction for future research and development.