Deconstructing Predictive Coding Networks
Explore the biology of belief, local energy minimization, and pure PyTorch implementations of biologically plausible learning without backpropagation.
Part 1
Biology of Belief
Rethinking backpropagation through the lens of top-down predictions, local prediction errors, and Hebbian energy minimization.
Part 2
PyTorch Implementation
Building PCN inference loops and local Hebbian weight updates from scratch in
pure PyTorch — no loss.backward() required.
View Code on GitHub
Part 3
Benchmark Showdown
Pitting the biological PCN against a standard MLP on nonlinear regression and MNIST classification to reveal the cost of biological plausibility.