Noise-injected analog Ising machines enable ultrafast statistical sampling and machine learning
This addresses the inefficiency of Ising machines for machine learning applications, making them competitive with digital computers for tasks like neural network training, though it is incremental as it builds on existing Ising machine concepts.
The paper tackles the problem of analog Ising machines being inefficient for neural network training due to slow statistical sampling, and introduces a noise-injection method to enable ultrafast sampling, achieving equal accuracy as software-based training and orders-of-magnitude speed improvements in simulations.
Ising machines are a promising non-von-Neumann computational concept for neural network training and combinatorial optimization. However, while various neural networks can be implemented with Ising machines, their inability to perform fast statistical sampling makes them inefficient for training neural networks compared to digital computers. Here, we introduce a universal concept to achieve ultrafast statistical sampling with analog Ising machines by injecting noise. With an opto-electronic Ising machine, we experimentally demonstrate that this can be used for accurate sampling of Boltzmann distributions and for unsupervised training of neural networks, with equal accuracy as software-based training. Through simulations, we find that Ising machines can perform statistical sampling orders-of-magnitudes faster than software-based methods. This enables the use of Ising machines beyond combinatorial optimization and makes them into efficient tools for machine learning and other applications.