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

Name: PotentialEP: Backprop-Free Training of CIFAR-Scale Generative Models via Equilibrium Propagation
Uploaded: 2026-10-19T00:57:46.051154+00:00
Duration: 369 s
Description: The paper introduces Equilibrium Propagation (EP), a backpropagation-free training method suitable for analog hardware. It shows that a generative model trained with EP achieves competitive quality on CIFAR-10, matching backpropagation's performance. Furthermore, it demonstrates that the resulting model can be sampled purely through relaxation on analog devices without any retraining.

The paper introduces Equilibrium Propagation (EP), a backpropagation-free training method suitable for analog hardware. It shows that a generative model trained with EP achieves competitive quality on CIFAR-10, matching backpropagation's performance. Furthermore, it demonstrates that the resulting model can be sampled purely through relaxation on analog devices without any retraining.

Open the full episode