PotentialEP: Backprop-Free Training of CIFAR-Scale Generative Models via Equilibrium Propagation

Key points

• Two things are needed: a generative algorithm whose computation is a relaxation, and a way to train it without backpropagation, whose global gradient transport is hard to realise in analog.1
• We prove that a predictive coding network with a single linear injection term added to its energy relaxes to a state whose visible layer carries that gradient inherently, with an error that vanishes for small injection strengths.1
• The network can also be trained without backpropagation: Equilibrium Propagation reproduces backpropagation's parameter updates to a cosine similarity of 0.9998, and training end to end with those updates reaches an FID of 3.64 on unconditional CIFAR-10 against 3.54 for backpropagation.1
• This quality does not require group normalisation or attention: an architecture of only convolutions and pointwise nonlinearities, both native to an analog crossbar, reaches an FID of 3.51 when trained by backpropagation.1

Sources (1)

• [1]PotentialEP: Backprop-Free Training of CIFAR-Scale Generative Models via Equilibrium Propagation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 12:06 AM

  Generating an image with a diffusion or flow-matching model typically involves integrating a velocity field for hundreds of sequential steps.

  Two things are needed: a generative algorithm whose computation is a relaxation, and a way to train it without backpropagation, whose global gradient transport is hard to realise in analog.

Extractive summary: sentences quoted from the sources.