covariance-reservoir

---

This repository contains basic scripts to generate and train models and synthetic data samples as analyzed in:

1. "Covariance Features Improve Low-Resource Reservoir Computing Performance in Multivariate Time Series Classification", S. Lawrie, R. Moreno-Bote and M. Gilson. (https://link.springer.com/chapter/10.1007/978-981-16-9573-5_42)
2. "Covariance-based information processing in reservoir computing systems", by S. Lawrie, R. Moreno-Bote and M. Gilson (https://doi.org/10.1101/2021.04.30.441789).

Dependencies

The condaenvironment specifications used for this project can be found in environment.yml. However, not all dependencies listed there are strictly necesary. Main required Python libraries have widespread use (matplotlib, scikit-learn, scipy, numpy, pandas, os, time), so the scripts can be ran on any environment provided those libraries are installed.

Synthetic data

File synthetic_data.py contains all code relevant to produce synthetic datasets as analyzed in the articles. The instantiations we specifically analyzed are also available in Zenodo (https://zenodo.org/record/4906349#.YaX4F9DMI2w)

Real data

The spoken Arabic digits dataset can be freely downloaded from the UCI Machine Learning Repository website (https://archive.ics.uci.edu/ml/datasets/Spoken+Arabic+Digit). File auxiliary_functions.pycontains functions to read the files and zero-pad them to produce samples with consistent length.

digits_analysis_perceptron.ipynb

This Jupyter notebook displays the code used to train a mean/covariance perceptron readout, with and without reservoir. As example, we use the spoken digits dataset, located in folder /dataset.