Neural Network Visualization

Visualization of neural network architectures and parameters.

Description

This project was done for my master thesis. A general description can be taken from the thesis:

Abstract

Artificial neural networks is a popular field of research in artificial intelligence. The increasing size and complexity of huge models entail certain problems. The lack of transparency of the inner workings of a neural network makes it difficult to choose efficient architectures for different tasks. It proves to be challenging to solve these problems, and with a lack of insightful representations of neural networks, this state of affairs becomes entrenched. With these difficulties in mind a novel 3D visualization technique is introduced. Attributes for trained neural networks are estimated by utilizing established methods from the area of neural network optimization. Batch normalization is used with fine-tuning and feature extraction to estimate the importance of different parts of the neural network. A combination of the importance values with various methods like edge bundling, ray tracing, 3D impostor and a special transparency technique results in a 3D model representing a neural network. The validity of the extracted importance estimations is demonstrated and the potential of the developed visualization is explored.

How to use

Prepare the configs/processing.json with the parameters described here.
Create a neural network model and process it. An example of this process is given in examples/process_mnist_model.py on MNIST data.
Start the visualization tool start_tool.py and select the neural network via Load Processed Network to render the representation of the neural network.

Multiple scripts are located in examples, which can be adapted to create and process neural networks. examples/evaluation_plots.py for example can be used to recreate the evaluation data and plots of my thesis.

Rendering Tool

The visualization tool start_tool.py can be used to render and/or process neural networks. Instead of existing ones, you can also generate random networks and process them of various sizes. For neural networks the visualization results in a more structured view of a neural network in regards to their trained parameters compared to the most common ones.

Example

The parameters of the three neural networks represented above are all trained differently, while having the same architecture. The one on the left is not trained at all with randomly assigned values. The nodes and edges of this model spread further from the center. The middle one is trained with some basic settings for learning rate and achieving >90% accuracy. The third one on the right is trained in the same way but with an additional *L1* regularization, with similar accuracy rating and is the most narrow model. **The closer together the edges are the greater the generalization** of these parts of the neural network.

Controls

Key	Description
H	Toggle rotation
K	Screenshot
0-9	Switch camera position

GUI

The settings for shaders, statistics and the processing of neural networks in general is controllable by the gui.

Shader Parameters

The parameters used in the shaders rendering the neural network can be changed by either the configs/rendering.json or by changing the values in the gui. The visualization can differ vastly and different results can be seen here.

Name	Recommended	Range	Description
Size	0.02	0 - 1.0	size of the primitive objects
Base Opacity	0.0	0.0 - 1.0	base opacity of the objects
Importance Opacity	1.1	0.0 - 2.0	rate at which the importance values influence opacity
Depth Opacity	0.5	0.0 - 1.0	rate at which the distance to the camera influence opacity
Depth Exponent	0.25	0.0 - 10.0	rate at which the density at different points on a object influences opacity
Importance Threshold	0.1	0.0 - 1.0	defines the threshold at which an object is rendered based on its importance value

Processing

The above pipeline explains the bundling process of a neural network through my code.

This image shows the different stages in the processing pipeline.

Parameters

The processing can be influenced by the following parameters. The default values are in general derived from empircally tested values of related work regarding edge bundling methods. Some values have a high impact on the processing time.

Name	Recommended	Range	Description	Performance Impact
edge_bandwidth_reduction	0.9	0 - 1.0	reduction of advection range every iteration for edge samples	high
node_bandwidth_reduction	0.95	0 - 1.0	reduction of advection range every iteration for nodes	low
edge_importance_type	0	{0,1,2,3}	the calculation type for edge importance	low
layer_distance	0.5	0.0 - 1.0	the distance between the nodes of every neural network layer	medium
layer_width	1.0	0.0 - 1.0	the width of every layer slice on which the nodes reside	medium
prune_percentage	0.0	0.0 - 1.0	the percentage of edges, which should be ignored in order of their importance values, lower value means more longer processing	high
sampling_rate	15.0	5.0 - 20.0	defines the amount of samples created per distance unit, higher rate means more detailed	very high
smoothing	true	{true, false}	should smoothing of edges be applied between each iteration?, can break without	high
smoothing_iterations	8	0 - 16	smoothing iterations between every advection iteration	high

To change the parameters for processing change values in following file: configs/processing.json

{
    "edge_bandwidth_reduction": 0.9,
    "edge_importance_type": 0,
    "layer_distance": 0.5,
    "layer_width": 1.0,
    "node_bandwidth_reduction": 0.95,
    "prune_percentage": 0.0,
    "sampling_rate": 15.0,
    "smoothing": true,
    "smoothing_iterations": 8
}

Importance

Each classification is represented by one color. Nodes and egdes are colored according to their importance in the network for correctly predicting the associated class. The validity of the importance is proven by pruning the model parameters in order of their calculated importance.

Overall Importance Pruning	Class Importance Pruning

The left plot shows that pruning unimportant parameters does not influence the prediction accuracy of the model as much as the important parameters.

Also by pruning based on importance of specific classes shows the accuracy is preserved for the exact classes in the right plot. The accuracy for the focused class is always higher compared to the overall accuracy.

Used System

Windows 10
NVIDIA GeForce GTX 1660 SUPER
AMD Ryzen 7 3700X

Notes

Processing Times - Pocessing of a fully connected neural network with following nodes per layer: 784, 81, 49, 10 takes 12 minutes. So the one-time calculations are not in real-time.
Python Version - Tested on 3.7 and 3.8
Dependencies - check requirements.txt

Fleschier / nn_vis