A Blueprint for Equilibrium-Based Differentiable Continuous-Variable Thermodynamic Computing
The paper lays out thermodynamic hardware as a machine-learning stack, not just a device concept.
It centers on stochastic analog processes governed by Langevin dynamics and tunable energy potentials. The authors show how those hardware-native energy-based models could be used to build and train common ML model classes through probabilistic graphical models. They also compare expected runtime and energy use with theory and numerical studies, and describe an early superconducting-circuit realization driven by thermal noise. Source: ArXiv · AI/CL/LG's note
It centers on stochastic analog processes governed by Langevin dynamics and tunable energy potentials. The authors show how those hardware-native energy-based models could be used to build and train common ML model classes through probabilistic graphical models. They also compare expected runtime and energy use with theory and numerical studies, and describe an early superconducting-circuit realization driven by thermal noise. Source: ArXiv · AI/CL/LG's note
score 4