aharley/nn_vis

An interactive visualization of neural networks

An Interactive Node-Link Visualization of Convolutional Neural Networks

Abstract

Convolutional neural networks are at the core of state-of-the-art approaches to a variety of computer vision tasks. Visualizations of neural networks typically take the form of static node-link diagrams, which illustrate only the structure of a network, rather than the behavior. Motivated by this observation, this project presents a new interactive visualization of neural networks trained on handwritten digit recognition, with the intent of showing the actual behavior of the network given user-provided input. The user can interact with the network through a drawing pad, and watch the activation patterns of the network respond in real time.

Paper PDF

Live demos

Live demos for all models are available at https://adamharley.com/nn_vis:

1. 3d visualization of a multi-layer perceptron:

   

2. 3d visualization of a convolutional network:

   

3. 2d visualization of a multi-layer perceptron:

   

4. 2d visualization of a convolutional network:

   

FAQ

• Can I use this in my course/textbook/presentation?

  • Yes! Please just cite the work appropriately.

• How do I cite you?

  • Here is a plaintext citation:

    A. W. Harley, "An Interactive Node-Link Visualization of Convolutional Neural Networks," in ISVC, pages 867-877, 2015

    Here is a bibtex snippet:

    @inproceedings{harley2015isvc,
    title = {An Interactive Node-Link Visualization of Convolutional Neural Networks},
    author = {Adam W Harley},
    booktitle = {ISVC},
    pages = {867--877},
    year = {2015}
    }
    

• I tried to run the code locally, and I see classifications, but I do not see the network visualization. What's wrong?

  • This is probably related to json requests being blocked by something called CORS policy. The solution is to upload the code to a web address and run it from there, instead of running locally.

Contact: [email protected]

nn_vis's Issues

About CORS policy

I think Web Server for Chrome can solve this problem

Request for Guidance on Implementing Multi-Level Edge Connectivity

Hello,

I am working on a project aimed at developing educational tools to teach high school students about Convolutional Neural Networks (CNNs). I found your nn_vis tool to be a great starting point for this endeavor.

Feature Implementation Request:
I am interested in enhancing the nn_vis tool by adding a feature where clicking on a neuron or feature map box in a higher-level convolutional layer would display edges connecting it to its corresponding neurons or feature maps in the lower layer, continuing down to the first layer. This would help in illustrating the hierarchical structure of CNNs and the connection between different layers.

Purpose:
The purpose of this enhancement is to provide a more interactive and comprehensive understanding of how different layers in a CNN are interconnected, which is crucial for beginners in machine learning.

Request for Insights:
As I am planning to implement this feature myself, I would greatly appreciate any insights or guidance you could provide on how best to approach this modification. Specifically, I am interested in:

• Understanding the current structure of the nn_vis tool, particularly how neurons and feature maps are represented and handled.
• Suggestions on how to efficiently track and display connections between layers in the tool's existing framework.
• Any potential challenges or considerations I should be aware of while implementing this feature.
• It would be very helpful to know what are the functions handling calculating the edges and displaying them.

Vision

Here is an image of how I imagine the fully implemented version would look like.

Conclusion:
Your expertise and insights into the tool's design and functionality would be incredibly valuable for this enhancement. I am eager to learn from your experience and contribute to the development of this educational resource.

Thank you for your time and for creating such a valuable tool for the machine learning community.

Best regards,
Marco Qin
Brandeis University

Hi, this tool is really cool! I am wondering why each filter in the second conv layer is connected to different activation maps in the previous layer?