Experimental Demonstrations of Optical Neural Computers

Ken Hsu, David Brady, Demetri PsaltisNeurIPS 1987 · OpenReview

Introduction

In the late 1980s, the primary challenge for neural computing was interconnectivity. Electronic circuits are essentially two-dimensional, making it physically difficult and highly inefficient in terms of space and power to connect thousands of neurons to one another.

Optics offers a natural solution to this bottleneck: light beams can pass through each other in 3D space without interference, resulting in exceptionally low cross-coupling. However, optics is generally linear, whereas neural networks require non-linear "thresholding"—similar to a biological neuron firing only when a signal is strong enough. This paper explores how to combine optical interconnections with optoelectronic devices to create functional, learning-capable neural hardware, demonstrating two foundational building blocks for optical AI: an associative memory loop and a photorefractive perceptron.

Experimental Methods

The authors present two distinct experimental setups to tackle different aspects of neural computing:

1. The Optical Associative Loop

This system is designed to "fill in the blanks" of a partial image. It uses two cascaded Vander Lugt correlators, which form a standard optical setup for pattern matching.

Optical Associative Loop Setup
  • Neural Plane: A Liquid Crystal Light Valve (LCLV) acts as the "neural plane" to provide the necessary non-linear thresholding and gain.
  • Feedback Filter: A pinhole array acts as a filter in the feedback loop, ensuring that only strong correlations (matches) are fed back into the system to "latch" onto a stored memory.
  • Dual Holograms: The system cleverly uses two holograms: one high-pass filter for sharp discrimination of the input, and one broadband filter for high-fidelity image reconstruction.

2. The Photorefractive Perceptron

This setup implements a single-neuron learning algorithm using a photorefractive crystal (SBN60:Ce).

  • Weight Storage: Synaptic weights are stored as holograms within the crystal's volume.
  • Learning Mechanism: A piezoelectric mirror is used to electronically control the phase and modulation depth of the light. This allows the system to "add" to the weight (learning) or "subtract" from it (erasure/un-learning) based on the perceptron error-correction rule.

Key Insights and Innovations

  • 3D Wiring via Holography: The authors emphasize that holography allows for "superimposed" channels in three dimensions. This means an optical system can maintain millions of interconnections far more simply than a physical wire-based electronic system.
  • The Fidelity-Discrimination Trade-off: A major breakthrough in the associative loop was the use of two different holograms. By optimizing one hologram to recognize the input (high-pass) and the other to reconstruct a clear image (broadband), the system achieves both stability and accuracy.
  • Dynamic Weight Storage: Using photorefractive crystals allows for "real-time recording and slow erasure." This mimics biological long-term memory, where traces of the learning process are stored and modified over time without needing a separate "fixing" or development step.

Key Results

Instead of modern large-scale datasets, the experiments utilized face images for the associative loop and simple 10x10 pixel grids for the perceptron.

Associative Recall: The optical loop successfully recalled a complete face when presented with only a partial (half-face) input. It also demonstrated "discrete shift invariance," meaning it could recognize the face even if it was slightly moved or rotated, provided it aligned with the pinhole array.

Associative Recall of Faces

Learning Convergence: The photorefractive perceptron successfully learned to group four geometric patterns into two classes. After just four training cycles, the system achieved 100% accuracy on the training set, requiring only seven total weight corrections (four recordings and three erasures).

High Neuron Density: The LCLV setup provided approximately 160,000 "neurons" on a mere 1-inch aperture, powerfully demonstrating the massive scaling potential of optical planes.

Limitations and Future Work

While groundbreaking, the systems had notable limitations:

  • Shift Invariance: The associative loop's shift invariance was "discrete" rather than continuous. If the input image shifted by an amount that didn't align with the pinhole array, the feedback signal was lost.
  • Depth of Focus: For the photorefractive perceptron, the number of channels is limited by the "depth of focus" required to isolate individual pixels within the crystal.
  • Non-linearity Challenges: The authors acknowledge that "good optical nonlinearities are hard to obtain." Their reliance on optoelectronic Spatial Light Modulators (SLMs) was a practical compromise, but they suggest future systems should strive for all-optical processes to increase speed and reduce complexity.

The paper concludes that the ultimate next step is integration: combining the robust recall dynamics of the associative loop with the adaptive learning capabilities of the photorefractive perceptron to create a truly adaptive, fully optical computer.

Summary updated

Abstract

We describe two expriments in optical neural computing. In the first a closed optical feedback loop is used to implement auto-associative image recall. In the second a perceptron-Iike learning algorithm is implemented with photorefractive holography.
BibTeX
@inproceedings{hsu1987experimental,
  author = {Ken Hsu and David Brady and Demetri Psaltis},
  title = {{Experimental Demonstrations of Optical Neural Computers}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1987},
  url = {https://openreview.net/forum?id=SkbeTLW_bS},
}
Figure