New Algorithm Scales Thermodynamic AI for Low-Power Inference.

A new study presents a significant advancement in scaling thermodynamic AI models, which hold promise for low-power AI inference and edge computing using Ising model-based devices. Previous theoretical work established that the time-averaged behavior of high-temperature Gibbs-sampled Ising systems could perform feed-forward neural inference. This research translates that theory into a practical, scalable algorithm. The core contribution is a purely backpropagation-based method for training deep convolutional networks specifically designed for thermodynamic inference on Ising machine hardware. This approach has demonstrated impressive results, achieving 94.9% accuracy on CIFAR-10 and 76.0% on CIFAR-100 using binary Gibbs sampling for image classification tasks. Furthermore, the researchers developed and validated a mathematical theory that connects inference cost with accuracy and provides methods for controlling autocorrelation times. They also derived asymptotic results to bound inference cost against performance and outlined algorithms for computing optimal inference schedules, paving the way for more efficient hardware development and the future of high-temperature thermodynamic AI.