GINN: Geometric Illustration of Neural Networks

TL;DR

GINN provides a geometric visualization of decision boundaries in neural networks, illustrating how ReLU units activate during training, aiding understanding of neural network behavior.

Contribution

This work introduces GINN, a tool for visualizing ReLU activation boundaries in neural networks, enhancing interpretability of decision boundary formation.

Findings

ReLU units switch states at specific input points during training

Visualization reveals geometric patterns in decision boundaries

Tracking activation points aids understanding of network learning dynamics

Abstract

This informal technical report details the geometric illustration of decision boundaries for ReLU units in a three layer fully connected neural network. The network is designed and trained to predict pixel intensity from an (x, y) input location. The Geometric Illustration of Neural Networks (GINN) tool was built to visualise and track the points at which ReLU units switch from being active to off (or vice versa) as the network undergoes training. Several phenomenon were observed and are discussed herein. This technical report is a supporting document to the blog post with online demos and is available at http://www.bayeswatch.com/2018/09/17/GINN/.