Skip to content

Latest commit

 

History

History
96 lines (72 loc) · 4.22 KB

README.md

File metadata and controls

96 lines (72 loc) · 4.22 KB

Minimum Description Length Hopfield Networks

Python 3.11 MIT License arXiv

Code for Minimum Description Length Hopfield Networks paper.

Getting Started

Setup

Note that it has been tested with Python 3.11, but any version after 3.9 should work.

virtualenv venv
source venv/bin/activate
pip install -r requirements.txt  

Before training or running experiments, please make sure to run python prepare_data.py to obtain all the required datasets. Additionally, note that all the scripts below rely on configuration parameters specified in config.py. For instance, the default training of a specific model is performed with the following parameters:

  • DIGITS_TO_TEST = [0, 1, 2, 3, 4]
  • NOISE_TYPE = NoiseType.DISCRETE
  • NOISE_LEVEL_TO_TRAIN = NoiseLevel.LOW
  • NUM_NOISE_VARIATIONS_TO_TRAIN = 5

Which means running on 5 digits (0, 1, 2, 3, 4) with discrete low noise, and having 5 noisy examples per digit.

Training a Modern Hopfield Network

Adapted from eqx-hamux, use the train_mhn.py script to train a Modern Hopfield Network based on the provided configuration.

Training a Minimum Description Length Modern Hopfield Network

You can train using Simulated Annealing either single-processed in train_sa.py or multi-processed using multiple_train_sa.py. The number of workers is determined by NUM_WORKERS in config.py.

Running Experiments

To run the experiments as mentioned in the paper, one should execute run_experiments.py. This script runs both MDL-MHN experiments and MHN experiments. The results are saved in a default directory named results and can be parsed using parse_results.py.

Parsing Results

Parsing results will provide the following statistics and graphs for each dataset you ran against in your results directory. By default, this includes golden, discrete noise, balanced discrete noise, continuous noise, and balanced continuous noise.

  • If show_samples=True:
    • For MHN experiments, two consecutive experiments of MHN will be sampled: one with the correct number of pre-determined clusters for training, and one with an incorrect number.
    • For MDL-MHN experiments, 2 random sub-experiments will be chosen. These samples will be displayed as organized plots along with the golden solution.
  • Distance distributions per noise level: These distributions show the distances from the trained MHN memories for both the golden digits and the training set digits. We can observe the shift in MDL-MHN, where the distances concentrate towards lower values after training. The statistics of mean and median distances per noise level are also provided.
  • MDL-MHN only:
    • A bar plot of clusters in the trained MHNs, along with the golden baseline.
    • A line plot of the number of memories vs. the number of digits tested will also be shown, with color representing the number of training examples and line type representing the level of noise.
    • Accuracies - hard accuracy denotes the number of times MDL-MHN correctly identified the number of clusters, and soft accuracy denotes the number of times it correctly identified the number of clusters or a lower number.

Citation

@article{abudy2023minimum,
      title={Minimum Description Length Hopfield Networks}, 
      author={Matan Abudy and Nur Lan and Emmanuel Chemla and Roni Katzir},
      year={2023},
      eprint={2311.06518},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}