Visual Auditor for Bias in CNN Image Classifier

What is VisCUIT-logo VISCUIT ?

CNN image classifiers often suffer from biases that impede their practical applications. Most existing bias investigation techniques are either inapplicable to general image classification tasks or require significant user efforts in perusing all data subgroups to manually specify which data attributes to inspect.

We present VISCUIT, an interactive visualization system that reveals how and why a CNN classifier is biased.

Interactive Features

  1. Select an underperforming image subgroup to explore by clicking an underperforming subgroup from the list. Then, the well-performing subgroup which is the most similar with the selected underperforming subgroup, confusion matrices of the two subgroups, and neuron activations are displayed.
  2. Grad-CAM window is displayed when an image in the selected underperforming or its similar well-performing subgroup is clicked. You can close the window by using X button in the window or ESC key.
  3. Adjust neuron activation score threshold by using the slider in the header.
  4. Highlight the neurons for similar concepts by hovering over a neuron.
  5. Neuron Concept Window for each neuron is displayed when a neuron is clicked. It shows the neuron's activation score and the concept patches that activate the neuron the most.

Video Tutorial

How is VISCUIT implemented?

We implemented VISCUIT using the standard HTML/CSS/JavaScript web technology stack and the D3.js visualization library. CNN model training and inference are all implemented using PyTorch.

Who Developed VISCUIT ?

Led by Seongmin Lee, VisCuit is created by Machine Learning and Human-computer Interaction researchers at Georgia Tech. The team includes Seongmin Lee, Jay Wang, Judy Hoffman, and Polo Chau.

How Can I Contribute?

If you have any questions or feedback, feel free to open an issue or contact Seongmin Lee. We’d love to hear your experience with VISCUIT! If you’d like to share (e.g., why you use VISCUIT, what features you find most helpful), please reach out to us. VISCUIT is an open-source project, and we welcome pull requests for new feature implementations and bug fixes, etc.