Optical Implementation of Equilibrium Propagation Using Spatial Photonic Ising Machines

Technical Summary: Optical Implementation of Equilibrium Propagation Using Spatial Photonic Ising Machines

Problem Statement
Hardware implementations of machine learning algorithms are increasingly sought after for their potential energy efficiency, particularly within photonic systems. However, state-of-the-art digital networks rely on error backpropagation, an algorithm that is difficult to map onto analog physical systems due to the requirement for bidirectional light propagation and precise phase measurements. While contrastive energy-based approaches offer an alternative by exploiting the intrinsic dynamics of physical substrates, experimental realizations of Equilibrium Propagation (EP)—a framework that trains energy-based networks without explicit backpropagation—remain scarce. Existing physical implementations often require external digital update loops or operate as fully autonomous systems with limited scalability. There is a need for a parallel, energy-efficient physical system capable of performing the gradient-based updates required for EP.

Methodology
The authors propose and demonstrate a hybrid optical-digital implementation of Equilibrium Propagation using a Spatial Photonic Ising Machine (SPIM). The core of the methodology involves:

1. Energy Functional and Dynamics: The system trains a dynamical system that relaxes to the ground state of an energy functional $E$. The energy includes input bias, quadratic interactions, a bounding quadratic term, and a cost term for output neurons. The learnable parameters are the interaction matrices $J_{in}$ and $J_{dyn}$.
2. SPIM Encoding: To bypass the computational bottleneck of evaluating interaction gradients ($O(N_d^2)$ matrix-vector multiplications), the authors utilize a SPIM. They employ a gauge transformation method to encode continuous neuron states ($s_i \in [-\pi/2, \pi/2]$) and rank-1 binary trainable patterns ($\xi_{k,i} \in \{-1, 1\}$) as phase modulations on a Spatial Light Modulator (SLM).
3. Hamiltonian Evaluation: The SPIM computes the Hamiltonian $H_{SPIM}$ via free-space propagation and a single intensity measurement. By taking linear combinations of rank-1 Mattis problems, the system achieves full programmability of the interaction matrix.
4. Gradient Estimation:
   • Inference: The gradient with respect to neuron states is estimated using a finite difference scheme, requiring two optical measurements of the Hamiltonian with shifted phases ($\pm \pi/4$).
   • Learning: The gradient with respect to the coupling parameters ($\lambda_k$) is evaluated optically in a single shot. The gradient with respect to the binary patterns ($\xi_{k,i}$) is evaluated numerically to reduce experimental runtime, though the paper notes it could theoretically be evaluated optically.
5. Optimization Loop: The system follows a bi-level optimization: an inner loop minimizes energy (inference) and an outer loop updates parameters based on the system's sensitivity to "nudging" the output neurons toward a target. The authors utilize the Binary Optimizer (BOP) for updating binary patterns and standard optimizers (like Adam) for continuous variables in numerical simulations.

Key Contributions

• Hybrid Implementation: The paper demonstrates a concrete hybrid optical-digital architecture where the SPIM handles the computationally intensive evaluation of the Hamiltonian and specific gradients, while digital components handle the remaining terms and optimization logic.
• Continuous Neuron Representation: By adapting the gauge transformation method, the authors successfully represent continuous neuron states on a hardware platform originally designed for binary variables, overcoming a significant limitation of standard SPIMs.
• Experimental Demonstration: The system is experimentally validated on the Wine classification dataset, achieving a test accuracy of $89.7 \pm 3.2\%$.
• Numerical Scalability Analysis: The authors provide extensive numerical simulations on the MNIST dataset. They show that a future SPIM with continuous parameters (eliminating the binary restriction on $\xi$) can achieve $97.81 \pm 0.15\%$ accuracy. They also demonstrate that the architecture remains viable with constrained resources (e.g., 8-bit precision) and can support structured, deeper architectures.

Results

• Experimental (Wine Dataset): The hybrid system achieved an average test-set accuracy of $89.7 \pm 3.2\%$ over seven independent runs. The discrepancy between this result and the simulation accuracy ($98.2\%$) is attributed to experimental noise (shot noise, phase fluctuations, laser power instability), which acts as an effective temperature affecting the equilibrium state.
• Inference Acceleration: Post-training, the energy landscape featured distinct basins of attraction, allowing inference to be accelerated to a single gradient descent step while maintaining $85.71 \pm 2.3\%$ accuracy.
• Numerical (MNIST Dataset):
  • With continuous parameters, the system achieved $97.81 \pm 0.15\%$ accuracy, comparable to standard EP implementations and other photonic simulations.
  • Scaling studies indicated that a minimal configuration ($N_d=32, K=22$) could outperform linear classifiers.
  • Quantization analysis revealed that reducing neuron precision to 8-bit (consistent with SLM resolution) significantly accelerated inference steps without degrading accuracy, whereas 4-bit precision caused a notable drop in performance.

Significance and Claims
The paper claims to provide a "concrete pathway toward energy-efficient physical implementations of Equilibrium Propagation." By leveraging the parallelism of SPIMs, the approach avoids the strict hardware constraints of optical backpropagation (such as bidirectional propagation and precise phase measurement). The authors emphasize that their method allows for the direct reuse of scalable SPIM platforms to perform standard machine learning tasks without requiring complex optical systems or accurate digital twins for training.

The work broadens the potential applications of SPIMs beyond quadratic unconstrained binary optimization (QUBO) and Boltzmann machine training, positioning them as viable candidates for analog, low-energy neuromorphic computing. The authors conclude that while current experimental limitations (binary patterns, noise) affect performance on complex datasets, future hardware improvements (continuous parameters, larger SLMs, wavelength multiplexing) could fully realize the potential of this method for physical machine learning.