Bio-inspired backpropagation-free training for optical neural networks

Area of Science:

• Optics
• Artificial Intelligence
• Computer Science

Background:

• Optical Neural Networks (ONNs) offer high-speed, energy-efficient computing but face challenges with traditional backpropagation algorithms.
• Physical optical systems struggle with the reciprocal paths and sensitivity required for error backpropagation in traditional ONNs (T-ONNs).

Purpose of the Study:

• To propose a novel bio-inspired optical neural network (B-ONN) that bypasses gradient backpropagation for improved optical computing.
• To demonstrate a layer-wise target propagation mechanism using trainable error convolution kernels for local learning in ONNs.

Main Methods:

• Implemented a bio-inspired backpropagation-free optical neural network (B-ONN) using layer-wise target propagation.
• Introduced trainable error convolution kernels for local learning, eliminating the need for optical conjugation.
• Validated B-ONN using programmable spatial light modulator (SLM) systems and chip-scale integration via nano printing.

Main Results:

• B-ONN achieved comparable accuracy to T-ONN on MNIST (93.25%) and Fashion-MNIST (82.28%) datasets.
• Demonstrated superior robustness against phase noise (75% accuracy at σ≈0.4π) and alignment errors (75% accuracy within ±3 pixels).
• Achieved 95% accuracy in handwritten digit recognition with SLM validation and 94% accuracy with chip-scale integration.

Conclusions:

• B-ONN offers a practical and feasible approach for deploying optical computing systems.
• The network learns smooth phase distributions, providing inherent structural robustness without noise-augmented training.
• Local learning rules facilitate asynchronous parallel updates, supporting scalable deep architectures.