Backprop Alternative: Augmented Lagrangian Predictive Coding

Modest empirical performance – Commenters note the proposed method achieves only ~85 % on MNIST and ~74 % on CIFAR‑10, far below what standard backpropagation reaches.

“~85% accuracy on MNIST. Sigh.” – cs702
“~74% on CIFAR-10. Still a far cry from backprop.” – cs702

Not positioned as a drop‑in replacement for backprop – The authors explicitly say they aren’t trying to supplant backprop, and discussants see it as a possible complementary or exploratory tool rather than a superior optimizer.

“They state replacing backprop is not their goal.” – qarl
“It might be beneficial while not being optimal on its own.” – Lerc
“[It] could be an alternate pathway out [of local minima] … doing training with radically different approaches … would avoid any method‑specific artifacts.” – Lerc

Motivation: understanding learning in distributed, brain‑like systems – The work is framed as a way to study how learning can occur without the precise error‑signal propagation that backprop requires, offering insight into biological learning.

“Their goal is to understand how distributed systems which cannot do backprop (the brain) can still do learning.” – qarl