A research program in thermodynamic machine learning — proving that energy-based models on thermodynamic hardware can actually be trained.
A disciplined loop from analytical claim to reproducible result. Research anyone can re-run, not just the person who wrote it.
Analytical framing of when energy-based models train: spectral gaps, gradient SNR, and the plateaus between them.
A single spectral gap can over-predict gradient SNR by up to 10³⁰×.
EXPERIMENTSEvery claim tested against a frozen pre-commitment — exact diagonalization, RBM block-Gibbs, GPU-scale denoising models.
Corrected predictor tracks 45/48 exact-diag cells — 92–99% in RBMs.
HARDWAREp-bits and sMTJ devices as physical Gibbs samplers — what makes the thing that runs on thermal noise learnable.
~1 fJ per spin flip on an sMTJ p-bit, vs ~10 pJ in software.
STACKThe exact toolchain behind every run — open libraries, pinned versions, rented accelerators.