Research map / Graph learning / Graph algorithms and clustering
Learning on Arbitrary Graph Topologies via Predictive Coding
Tommaso Salvatori, Luca Pinchetti, Beren Millidge, Yuhang Song, Tianyi Bao, Rafal Bogacz, Thomas Lukasiewicz
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a fully connected PC graph instead of the tested fully connected baselines | MNIST classification · MNIST | test accuracy: 91.76 $\pm$ 0.02 % vs 79.23 $\pm$ 0.15 % | Improved |
| Use a fully connected PC graph instead of the tested fully connected baselines | FashionMNIST classification · FashionMNIST | test accuracy: 83.72 $\pm$ 0.33 % vs 69.63 $\pm$ 1.64 % | Improved |
| Use a fully connected PC graph instead of the tested fully connected baselines | SVHN classification · SVHN | test accuracy: 84.51 $\pm$ 0.11 % vs 59.14 $\pm$ 2.64 % | Improved |
| Train a multilayer PCN rather than a BP multilayer perceptron | MNIST classification · MNIST | test accuracy: 98.47 $\pm$ 0.12 vs 98.41 $\pm$ 0.18 | No clear effect |
| Train a multilayer PCN rather than a BP multilayer perceptron | FashionMNIST classification · FashionMNIST | test accuracy: 89.92 $\pm$ 0.23 vs 90.29 $\pm$ 0.33 | No clear effect |
| Train a multilayer PCN rather than a BP multilayer perceptron | SVHN classification · SVHN | test accuracy: 88.99 $\pm$ 0.26 vs 89.31 $\pm$ 0.09 | No clear effect |
| Train a multilayer PCN rather than a BP multilayer perceptron | CIFAR10 classification · CIFAR10 | test accuracy: 56.23 $\pm$ 3.36 vs 59.11 $\pm$ 2.47 | Hurt |
| Add hierarchical structure instead of using a fully connected PC graph (ablation) | PC graphs on generation tasks · MNIST and FashionMNIST generation | generation quality | Improved |
| Add recurrent connections to a feedforward PC graph (ablation) | PC graphs reconstructing noisy or unfamiliar inputs · MNIST reconstruction | reconstruction quality | Improved |
| Add extra feedforward layers beyond the tested best-depth range (ablation) | Feedforward PC graph · Generation tasks | convergence time | Hurt |
| Increase feedforward network width beyond the best-performing width (ablation) | Feedforward PC graph · Generation tasks | generation quality | Hurt |
| Add strong weight decay rather than use low weight decay (ablation) | Recurrent PC graph · MNIST conditional reconstruction with labels | reconstruction quality | Improved |
Closest papers by meaning
- Understanding and Improving Optimization in Predictive Coding Networks (AAAI 2024 · similarity 0.89)
- Understanding Sample Efficiency in Predictive Coding (NeurIPS 2026 · similarity 0.89)
- Predictive Coding beyond Correlations (ICML 2024 · similarity 0.88)
- Divide-and-Conquer Predictive Coding: a structured Bayesian inference algorithm (NeurIPS 2024 · similarity 0.88)
- A Stable, Fast, and Fully Automatic Learning Algorithm for Predictive Coding Networks (ICLR 2024 · similarity 0.88)
- $\mu$PC: Scaling Predictive Coding to 100+ Layer Networks (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
