Authors: Hugo Ninou, Jonathan Kadmon, N. Alex Cayco-Gajic
Accepted at: NeurIPS 2025 (Spotlight)

Paper link: NeurIPS 2025 Proceedings

This repository contains the code needed to generate and analyze the data for reproducing the figures of the paper Curl Descent: Non-Gradient Learning Dynamics with Sign-Diverse Plasticity.

The codebase is structured to facilitate reproducing the empirical results from the paper, simulating Teacher-Student learning paradigms under various non-gradient plasticity rules, including variants of DFA (Direct Feedback Alignment).

• Data_generation/: Contains Python scripts to simulate the learning dynamics and generate data for the figures presented in the paper. The scripts utilize joblib for parallel execution across different initializations and hyperparameter variations.
• Eigval_distrib_theory_vs_empirical/: Contains a Jupyter notebook for comparing the empirical eigenvalue distributions of random matrices of interest in our analytical derivations with theoretical distributions predicted by elements from Random Matrix Theory.

The simulations rely on standard Python scientific libraries. Ensure you have the following installed:

• numpy
• tqdm
• joblib

You can install these dependencies via pip:

pip install numpy tqdm joblib