brajkowski / ann_visualization

A proof-of-concept, interactive visualization tool that allows for the exploration of deep artificial neural networks (ANN's).

ANN Visualization

Description

A proof-of-concept, interactive visualization tool that allows for the exploration of deep artificial neural networks (ANN's). This tool loads a pre-trained ANN capable of identifying handwritten digits (MNIST). Users can flip through the MNIST testing set and visualize neuron activations, neuron biases, and the connection weights that form the network.

System Requirements

• Supported Operating Systems

  • Windows 10 (x64)

• Minimum Hardware Requirements

  • 512 MB of RAM
  • 100 MB of available hard disk space

Build Instructions

1. Clone repository and open solution in Visual Studio
2. Install MathNet.Numerics 4.5.1: PM> Install-Package MathNet.Numerics -Version 4.5.1
3. Ensure build configuration is set to x64
4. Build in Visual Studio

Visualization Controls

• Navigate through MNIST test images

  • Left and Right arrow keys

• Change connection opacity

  • Up and Down arrow keys

• Change connection view type

  • G: View only positive connections
  • R: View only negative connections
  • B: View all connections

• Filter neurons

  • Right click on neuron to hide
  • A: Unhide all hidden neurons

• Additional Information

  • Left click on neuron to toggle its bias value
  • C: Hide all bias values