With the rapid advancement of generative artificial intelligence, efficiently training complex neural networks has emerged as a central challenge in deep learning. Ising machines, as specialized quantum-inspired hardware, show considerable promise for solving combinatorial optimization problems and training neural networks. However, current existing approaches are often limited to shallow architectures due to inefficient sampling strategies. In this study, we introduce a deep belief network (DBN) architecture that utilizes the sampling mechanism of a coherent Ising machine integrated with a spiking neural network (SNN-CIM). By stacking two restricted Boltzmann machines (RBMs), the model supports hierarchical feature learning while significantly enhancing training efficiency. The SNN-CIM mitigates amplitude heterogeneity through dissipative pulse injection and employs Gaussian noise to drive the system toward thermal equilibrium, enabling effective deep network training. Experimental results demonstrate that a single-layer RBM achieves $96.4\mathrm{\%}$ accuracy on the MNIST dataset, while the full DBN reaches $98.9\mathrm{\%}$ on the test set. Compared to Metropolis-Hastings-based MCMC sampling, the SNN-CIM exhibits broader sampling coverage and higher efficiency, achieving orders-of-magnitude performance improvements on large-scale problems. These findings underscore the potential of SNN-CIMs for scalable deep generative modeling and offer alternative pathways for enhancing performance in quantum-inspired computing and deep learning.

Close

PrevNext

View figure in article

• Photonics
• Physics of computation
• Quantum information processing
• Quantum optics
• Artificial neural networks
• Deep learning
• Ising model
• Optical computing