Visualization of neural network architectures and parameters.

• (23.11.2023) Master Thesis published [link] [pdf]
• (13.10.2023) Added docker (see DOCKER.md) and docker image (see here)
• (11.10.2023) Fixed bugs with demo and evaluation scripts, also added more example networks
• (30.12.2022) Added VR support (see VR_TOOL.md)

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

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.

Can be found at the university's publication server [link] [pdf]

If you use my work in your research, please cite it by using the following BibTeX entry:

@mastersthesis{Rogawski2023,
  author      = {Julian Rogawski},
  title       = {Visualization of Neural Networks},
  type        = {masterthesis},
  pages       = {ii, 48},
  school      = {Universit{\"a}t Koblenz, Universit{\"a}tsbibliothek},
  year        = {2023},
}

1. Prepare the configs/processing.json with the parameters described here.
2. Create a neural network model and process it. An example of this process is given in examples/process_mnist_model.py on MNIST data.
3. Start the visualization tool start_tool.py and select the neural network via Load Processed Network to render the representation of the neural network.
   • With Load Processed Network you can select and load the processed model with bundled edges and nodes.
   • With Load Network you can select and load the unprocessed network with its importance values, with unbundled edges and nodes.

Or

1. Run start_tool.py --demo to download data of some already processed model and importance data, loads and renders one of it.
   • With Load Processed Network you can select different processed networks and visualize them.
   • With Load Network you can select different unprocessed networks with just their importance values but no bundling of nodes and edges.

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.

A processed model can be downloaded here.

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.

See VR_TOOL.md for more info.

See DOCKER.md for more info.

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

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.

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.

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
}

Each classification is represented by one color. Nodes and edges 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.

• Windows 10
• NVIDIA GeForce RTX 3080
• AMD Ryzen 7 3700X

• Processing Times - Pocessing of a fully connected neural network with following nodes per layer: 784, 81, 49, 10 takes 3-4 minutes. So the one-time calculations are not in real-time.
• Python Version - Tested on 3.9 (3.7 and 3.8 was tested with older python dependencies)
• Dependencies - check requirements.txt