Thermodynamic AI models hit 94.9% accuracy on CIFAR-10 for low-power inference
94.9% accuracy on CIFAR-10 using energy-efficient Ising hardware and backpropagation.
A new paper by Andrew G. Moore on arXiv (2607.00170) presents a practical method for training large thermodynamic AI models that run on Ising machine hardware. The key innovation is a purely backpropagation-based algorithm that turns the known theoretical link between high-temperature Gibbs-sampled Ising systems and feed-forward neural networks into a scalable training pipeline. The models are deep convolutional networks designed for image classification, achieving 94.9% on CIFAR-10 and 76.0% on CIFAR-100 under binary Gibbs sampling — impressive results for a fundamentally different computing paradigm.
The paper also develops a mathematical theory linking inference cost to accuracy and controlling autocorrelation times in the sampling process. Asymptotic results show that inference cost is bounded by a well-controlled tradeoff with performance, and algorithms for computing optimal inference schedules are provided. This work directly addresses the scalability bottleneck that has limited thermodynamic computing, and the authors discuss implications for hardware development and the future of high-temperature thermodynamic AI models for low-power edge inference.
- Achieves 94.9% on CIFAR-10 and 76.0% on CIFAR-100 using binary Gibbs sampling on Ising hardware.
- Introduces a purely backpropagation-based algorithm for training deep convolutional networks on thermodynamic devices.
- Formalizes a cost-accuracy tradeoff and provides optimal inference scheduling algorithms.
Why It Matters
Paves the way for ultra-low-power AI inference on edge devices using thermodynamic computing hardware.