Only Strict Saddles in the Energy Landscape of Predictive Coding Networks?

This repo holds code for the NeurIPS 2024 paper Only Strict Saddles in the Energy Landscape of Predictive Coding Networks?. You can read my blog post for the key ideas of the paper.

Reproducibility

All the results and plots from the paper can be reproduced from the included notebooks. All notebooks except for the convergence experiments can be run in reasonable time (< 15min) on a CPU. For the heavier convergence experiments, we recommend using a GPU.

• To reproduce Figure 1, run Theoretical_Equilibrated_Energy.ipynb.
• To reproduce the toy examples in Figure 2 (and the statistics in Figure 7), run Linear_Chains_Analysis.ipynb and Hessian_Analysis_of_DLNs.ipynb.
• To reproduce Figure 3 & 4 (as well as 8-10), run Hessian_Analysis_of_DLNs.ipynb.
• To reproduce Figure 5 (as well as 11-12), run PC_vs_BP_Convergence_Experiments_on_DNNs.ipynb.
• Finally, to reproduce Figure 6, run Matrix_Completion_Experiment.ipynb.