Code of paper "Pseudo-likelihood produces associative memories able to generalize, even for asymmetric couplings"

Francesco D'Amico, Dario Bocchi, Luca Maria Del Bono, Saverio Rossi, Matteo Negri

ArXiv: https://arxiv.org/abs/2507.05147

The code released has been produced to run simulations for pseudolikelihood training of a binary, two-bodies interaction model. This code prensents two custom dataset: random binary data and random features binary data.

---

Installation

Clone the repository:

git clone https://github.com/Francill99/PseudoLikelihood_Analysis.git

Install:

pip install -e .

Dependencies: to create the same conda (or miniconda) environment we used

conda env create -f environment.yml
conda activate your_env_name

---

Structure of the repository

In "Pseudolikelihodd_Analysis" folder there are codes used. In "Graphs" folder there are data and code to reproduce the plots in the paper.

---

Basic Example

Run a single Pseuodlikelihood training with random data for 400 steps:

chmod +x simple_training.sh
./simple_training.sh

Parameters can be changed inside file simple_training.sh. It will be created a file training_log.txt containing the log of the training process, and at end of training in folder "savings" it will be saved a checkpoint of the model at last step. Metrics in the log file corresponds to final overlaps of dynamics with respect to training data, features and generalization data. Set D=0 to simulate model without features. In that case all final overlaps are related to training data.