Skip to content
master
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 

Interpreting Neural Networks by Reducing Nonlinearities during Training

This repo contains a short paper and sample code demonstrating a simple solution that makes it possible to extract rules from a neural network that employs Parametric Rectified Linear Units (PReLUs). We introduce a force, applied in parallel to backpropagation, that aims to reduce PReLUs into the identity function, which then causes the neural network to collapse into a smaller system of linear functions and inequalities suitable for review or use by human decision makers.

As this force reduces the capacity of neural networks, it is expected to help avoid overfitting as well.

Download the article in PDF format from the latest release at https://github.com/csirmaz/trained-linearization/releases/latest .