Optimizing Energy-based Neural Network Training with Coherent Ising Machine

Abstract: While Ising machines serve as advanced physical solvers for the Ising model,enabling applications in combinatorial optimization and neural network training,their scalability for large-scale neural networks remains constrained by hardware connectivity limitations and suboptimal training methodologies. In this work,we leverage a Coherent Ising Machine (CIM) to train an energy-based neural network using Equilibrium Propagation, achieving performance comparable to existing software-based implementations. We further enhance the algorithm by integrating the Adam optimizer to solve for the ground state of a Hopfield energy network, significantly improving convergence speed and solution accuracy. Additionally, we demonstrate the scalability of our approach across deeper network architectures and convolutional operations. Our results highlight the potential of CIM dynamics as a scalable platform for training complex neural networks, offering a pathway toward energy-efficient implementations via analog circuits, optoelectronics, or integrated photonics. This work establishes a novel physical framework for next-generation AI hardware development.

• The paper introduces an integrated framework combining CIM dynamics, Equilibrium Propagation, and the Adam optimizer to enhance training efficiency for energy-based neural networks.
• The paper leverages the physical parallelism of Coherent Ising Machines to map neural network optimization onto Ising energy minimization, achieving near-optimal results on Max-Cut graphs and MNIST classification.
• The paper demonstrates significant scalability and energy efficiency improvements, promising up to three orders of magnitude faster training compared to conventional CPU-based methods.

Optimizing Energy-based Neural Network Training with Coherent Ising Machine

Overview

This paper introduces an integrated framework leveraging Coherent Ising Machine (CIM) dynamics, Equilibrium Propagation (EP), and the Adam optimizer for energy-based neural network training. The approach capitalizes on the physical parallelism and scalability of CIMs—optical platforms simulating Ising spin systems via degenerate optical parametric oscillators—to address key limitations in classical hardware and software-based implementations for large-scale neural networks. By mapping neural network optimization onto Ising energy minimization and incorporating EP for biologically plausible learning, the authors demonstrate competitive performance on canonical benchmarks, enhanced computational speed, and substantial energy efficiency gains, particularly for deep and convolutional architectures.

Methodology

Coherent Ising Machine (CIM) as Physical Solver

CIMs are physical systems capable of simulating densely connected Ising models, overcoming classical computational bottlenecks by mapping neural network weights and biases onto an Ising Hamiltonian. The network energy is minimized through spin interactions governed by programmable coupling matrices. The authors utilize a measurement-feedback architecture with FPGA-modulated feedback for amplitude and phase control, allowing for arbitrary connectivity and scaling up to $10^6$ spins. This hardware-centric approach provides a highly scalable platform far surpassing current quantum annealers and classical combinatorial solvers.

Equilibrium Propagation (EP) for Learning

EP is adopted as a biologically plausible alternative to Backpropagation, relying on local measurements of equilibrium states—free and nudge phases—to compute weight updates via energy differentials, aligning with physical dynamical systems. EP is well-suited for implementation in CIMs due to its compatibility with symmetric weights and energy-based computation, and its operational principle closely mirrors that of Hopfield energy networks.

Adam Optimizer Integration

To ameliorate amplitude heterogeneity and local optima trapping, the Adam optimizer is incorporated within CIM dynamics. This hybrid algorithm—Adam-CIM—adapts learning factors per parameter using first and second moment estimates of gradients, facilitating efficient ground state search. The resultant method outperforms both conventional simulated annealing and standard CIM implementations in convergence speed and solution quality.

Experimental Results

Ground State Search: Max-Cut Graphs

Adam-CIM is benchmarked on combinatorial optimization tasks, notably Max-Cut instances from the G-set. Comparative analysis shows Adam-CIM achieving near-optimal energies with substantially reduced iterations. The final energy distribution is highly concentrated in low-energy regions, highlighting robust convergence. Simulated Annealing (SA) exhibits slower convergence and a broader energy distribution, often trapped in local minima.

Neural Network Training: MNIST Classification

An MLP with one hidden layer (256 units) is trained on the MNIST dataset using EP and Adam-CIM. Numerical results indicate test accuracy stabilizing above 95% after sufficient iterative convergence, matching or exceeding software-based training regimes. Notably, with minimal hidden layer complexity, EP suffices for accurate mapping, but deeper or wider networks demand increased $\beta$ and careful architectural tuning to mitigate gradient propagation challenges.

Energy distributions pre- and post-training illustrate a transition from high-entropy, scattered states to low-entropy, concentrated ones, mirroring phase transitions in physical systems. Adam-CIM outperforms SA and quantum annealers (e.g., D-Wave) in test accuracy for equivalent network sizes.

Structural Scalability and Convolutional Architectures

Architectural scaling reveals accuracy plateaus after threshold node counts; additional depth can lead to diminishing returns, attributable to error amplification in implicit update rules under EP. The framework is extended to CNNs, showing around 80% accuracy post 20 epochs on MNIST—a promising baseline for more complex vision tasks, though not yet matching state-of-the-art CNNs optimized with BP.

Scalability analysis with respect to physical implementation suggests potential for three orders of magnitude improvements in both time and energy efficiency compared to CPU-based training, especially when utilizing integrated photonic CIMs operating at >100 GHz.

Comparative Analysis and Claims

The Adam-CIM + EP approach exhibits several strong empirical results:

• Test accuracy on MNIST: 96.8% (±0.52%) for MLP with 256 hidden units; significantly higher than quantum annealing-based EP implementations [laydevant2024training].
• Scalability: CIM hardware supports up to 100,000 spins, several orders beyond D-Wave's 5,000.
• Computational speed: Adam-CIM achieves optimal solutions in fewer iterations than SA or s-CIM counterparts.
• Energy efficiency: Physical CIMs with integrated optics promise substantial reductions in training time and power consumption compared to digital hardware.

The authors boldly claim that increasing neural network complexity (more layers, more nodes) does not guarantee accuracy improvements with EP, and may sometimes degrade performance. Furthermore, EP cannot outperform BP in gradient descent fidelity, presenting a fundamental ceiling to physically plausible learning algorithms.

Implications and Future Directions

The confluence of CIM physics, EP learning, and adaptive optimizers marks a step toward scalable, energy-efficient AI hardware. This paradigm aligns with neuromorphic and optoelectronic computing, bridging biological plausibility and physical realizability. As optical and integrated photonic CIMs mature, deploying CIM-trained models for edge inference, probabilistic reasoning, and resource-constrained domains is increasingly viable.

Theoretical implications include deeper exploration of phase transition dynamics and error amplification in energy-based learning for deep and recurrent architectures. Practical directions involve optimizing network structures and $\beta$ scheduling, as EP's hyperparameters significantly impact convergence and robustness. Techniques such as three-phase EP may reduce estimator bias and approach BP-level performance, though the absence of explicit gradient formulas constrains ultimate accuracy.

Further research should target:

• Physical implementation benchmarks for large-scale CIM-based neural training
• Hybrid regimes combining EP and BP, or leveraging CIM for in situ sampling
• Architectural innovations for convolutional and transformer networks under EP

Conclusion

Integrating CIMs with EP and Adam optimization establishes a high-performance, scalable, and energy-efficient framework for training energy-based neural networks. While limitations exist concerning depth scalability and EP's theoretical ceiling, the approach achieves competitive numerical results and operational efficiency, laying the foundation for next-generation AI hardware platforms with broad applications in optical, analog, and quantum computing (2606.09117).