New Method Improves Deep Network Training Efficiency and Accuracy

Junlong Shen, Xingyu Li· July 16, 2026 View original

Summary

This research introduces WF-Act-PC, a novel predictive coding method that locally computes the Jacobian transpose, eliminating the need for a non-local autograd backward pass in deep networks. It achieves state-of-the-art accuracy on CIFAR-10/100 and Tiny-ImageNet, matching or exceeding backpropagation on deeper architectures.

Traditional deep learning relies on backpropagation, which requires non-local operations to route error signals. Predictive Coding (PC) offers a biologically inspired alternative, but still had one non-local dependency: the Jacobian-transpose product for error routing. This new work, WF-Act-PC, addresses this by showing how to compute this product locally for a specific class of deep network layers. The method achieves this by factoring the Jacobian transpose into three locally available terms, provided certain assumptions like weight symmetry and soft spectral-norm control are met. By restoring previously omitted corrections, WF-Act-PC closes the "transport gap" for these layers. Empirical results demonstrate that WF-Act-PC is the only PC method whose accuracy improves with depth, significantly outperforming existing PC baselines. It even matches or surpasses tuned backpropagation on deeper CIFAR-10 architectures and Tiny-ImageNet, making it a promising alternative for efficient and accurate deep network training.

Why it matters

This research offers a more biologically plausible and potentially more efficient way to train deep neural networks, which could lead to faster training times and better performance, especially for complex models.

How to implement this in your domain

  1. 1Explore the public implementation of WF-Act-PC on GitHub to understand its architecture.
  2. 2Experiment with integrating WF-Act-PC into existing deep learning frameworks for specific tasks.
  3. 3Benchmark its performance against traditional backpropagation on custom datasets and model architectures.
  4. 4Evaluate the computational overhead and memory footprint compared to current training methods.

Who benefits

AI/ML DevelopmentHigh-Performance ComputingRoboticsAutonomous Systems

Key takeaways

  • WF-Act-PC offers a local, biologically inspired alternative to backpropagation for deep network training.
  • It eliminates the non-local Jacobian-transpose dependency in predictive coding.
  • The method shows improved accuracy with depth and outperforms other predictive coding baselines.
  • It can match or exceed backpropagation performance on certain deep architectures and benchmarks.

Original post by Junlong Shen, Xingyu Li

"arXiv:2607.13380v1 Announce Type: new Abstract: Predictive Coding (PC) offers a biologically motivated alternative to backpropagation via local weight updates, yet routing error between layers still relies on an autograd Jacobian-transpose ($J^\top$) product - the last non-local…"

View on X

Originally posted by Junlong Shen, Xingyu Li on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses