Training deep generative models typically relies on backpropagation, but a new arXiv preprint introduces PotentialEP, a method that avoids backprop entirely. The paper reports that PotentialEP can train generative models at CIFAR scale using equilibrium propagation, a learning rule based on the physics of dynamical systems.
The authors motivate the work by noting that generating images with diffusion or flow-matching models usually requires integrating a velocity field over hundreds of sequential steps. That kind of computation is a poor fit for GPUs but well suited to analog neuromorphic substrates, where the sequential nature maps naturally onto physical dynamics.
PotentialEP appears to be a step toward closing the gap between modern generative models and neuromorphic hardware. The abstract does not provide full experimental details, but the claim of CIFAR-scale, backprop-free training is notable for researchers interested in energy-efficient, non-standard computing platforms.