Residual calibration for high-precision optical neural networks.

Area of Science:

• Optical computing
• Computational optics
• Photonics

Background:

• Optical processors offer high speed and energy efficiency for matrix-vector multiplications (MVMs), crucial for optical neural networks (ONNs).
• Existing analog optical architectures suffer from computational inaccuracies and limited scalability due to inherent errors.

Purpose of the Study:

• To develop and validate a novel residual calibration method for improving the precision of optical matrix computations.
• To address the critical challenge of achieving high-accuracy matrix products in optical computing platforms.

Main Methods:

• Proposed a residual calibration technique that iteratively refines optical computations using multiple low-precision multiplications.
• Conducted theoretical analysis to demonstrate exponential error reduction rates under specific conditions (maximum singular value < 1).
• Experimentally validated the method on fabricated optical processors.

Main Results:

• Achieved significant error reduction in optical computations through successive calibration iterations.
• Demonstrated substantial performance improvements in ONNs for semantic segmentation tasks.
• A single calibration iteration matched digital implementation accuracy, yielding a 24% increase in mean intersection-over-union and a 22% enhancement in pixel accuracy.

Conclusions:

• The residual calibration method offers a flexible and scalable solution for high-precision optical computations.
• This advancement significantly enhances the practicality of deploying optical processors in demanding computational applications.
• The study overcomes key limitations in optical computing, paving the way for more robust and accurate optical neural networks.