cifar-100-visualization

Interactive dashboard that enables some exploration of data from the CIFAR-100 dataset

Goals

1. Dimensionality Reduction: it utilizes the t-SNE algorithm over the activation layers of a CNN to create a 2D embedding of high-dimensional data
2. Explanation: it implements LIME to get the explaination of the class predicted by the deep model, it also shows the top-3 class distribution from the softmax layer
3. Visualization: for a nice, customizable and responsive user experience the UI is created using Dash

Details

A Keras model based on the VGG16 architecture is used for the predictions

To get valuable insights about the model (and dataset) weaknesses it is interesting to visualize the explanations of the wrong predictions, users can filter such images using a dropdown menu.

Examples:

Run the code

From the XAI folder:

pip install -r requirements.txt

Then run app.py and just click on the link, it might take a while (~20 seconds) depending on the hardware of your system