Backprop Alternative: Augmented Lagrangian Predictive Coding

PC-ALM replaces backpropagation with layer-local dynamical systems that utilize PI feedback controllers to distribute supervision credit. By incorporating dual neurons as Lagrange multipliers, this method successfully trains 1000-layer residual MLPs. This approach overcomes signal decay issues found in standard predictive coding and provides a biologically plausible alternative for gradient computation on neuromorphic hardware.