Thermodynamic computing hits 96.95% MNIST accuracy with moment-resolved readout
Physicists use thermal fluctuations as neurons, fusing multiple reservoirs for better classification.
In a new paper on arXiv (2607.14520), researchers JiZheng Duan, MingYang Zhao, YanWei Chen, and Lei Yang advance the field of thermodynamic computing by introducing moment-resolved readout. Traditional Langevin computing used only the first moment (mean) of driven reservoirs as computational signals. The team instead constructs response vectors from raw polynomial moments: the mean, the elementwise square, and the elementwise fourth power (E[x], E[x⊙2], E[x⊙4]). These align naturally with the linear, quadratic, and quartic terms of the local driven dynamics, capturing both displacement and central-shape information.
The work further proposes a heterogeneous multi-reservoir architecture where three reservoirs with distinct initialization and training histories produce a joint 2304-dimensional response representation. On the fixed MNIST 60000/10000 reproduction protocol, feature-level fusion achieved the best observed accuracy of 96.95% (9695/10000), versus 96.82% for the strongest single reservoir and 96.84% for equal-weight logit averaging. An exact paired McNemar test didn't establish statistical significance, but ablation and wrong-set overlap analysis suggests complementary classification errors. These findings point to higher-order moments and reservoir diversity as promising design principles for finite-time Langevin computing.
- Using raw polynomial moments (1st, 2nd, 4th) instead of just the mean boosts MNIST accuracy from 96.82% to 96.95%
- Three heterogeneous reservoirs with distinct training histories form a 2304-dimensional response vector for feature fusion
- Improvement not statistically significant (McNemar test), but ablation studies indicate complementary error reduction across reservoirs
Why It Matters
New design principles for thermodynamic computers could lead to ultra-low-power AI hardware using thermal noise.