Bilinear Energy Substrate Enables On-Chip Diffusion Model Training
A new theoretical result shows that score-based diffusion models can be trained entirely on an analog substrate without external digital accelerators. The method, called Symmetric Equilibrium Propagation, applies to bilinearly-coupled energy landscapes previously shown to accelerate inference by three to four orders of magnitude. The key contribution is proving that the training gradient can be estimated unbiasedly in the zero-nudge limit, with a finite-nudge bias bound controlled by substrate stiffness, local curvature, and loss-gradient norm. This closes the training loop on the same physical substrate, potentially enabling ultra-low-energy on-device learning for generative AI. The work is published on arXiv under ID 2604.23806.
Key facts
- Equilibrium Propagation applied to bilinear energy yields unbiased gradient estimator for denoising score matching.
- Bias bound for finite nudging depends on substrate stiffness, local curvature, and loss-gradient norm.
- Prior work showed bilinear analog substrate achieves 3-4 orders of magnitude energy advantage for inference.
- The method eliminates need for external digital accelerator during training.
- Published on arXiv with ID 2604.23806.
- Reverse process in score-based diffusion models is equivalent to overdamped Langevin dynamics.
- Bilinearly-coupled analog substrate replaces dense skip connections with low-rank inter-module couplings.
Entities
Institutions
- arXiv