A Random Matrix Approach to Extreme Learning Machine
This page contains a simple demo using Python 3 of the theoretical results in the following paper:
A Random Matrix Approach to Neural Networks
where recent advances in matrix matrix theory are used to analyze the performance of randomly-connected single-layer neural nets (also referred in literatures as extreme learning machines).
Comparison between theory and practice is available for data from
- MNIST database
- Gaussian mixture model
for a dozen of commonly-used activation functions.
To be able to test this code requires the following:
- Python: tested with version 3.6
- Numpy and Scipy
- Matplotlib for visulazation
- Scikit-learn for MNIST dataset
We strongly recommend you to use Jupyter nootbook to have a direct illustration within your web browsers: here.
- Zhenyu LIAO
- Ph.D. student at CentraleSupelec, Paris, France
- Website: https://zhenyu-liao.github.io/
- E-mail: zhenyu.liao@l2s.centralesupelec.fr
- Prof. Romain COUILLET
- Professor at CentraleSupelec, Paris, France
- Website: http://romaincouillet.hebfree.org/
- E-mail: romain.couillet@centralesupelec.fr