Class_Activation_Map (Tensorflow)

This project includes three scripts that describe how to extract "Activation Map" to explain the performance of the neural network

1. Class_Activation_Map_MNIST.ipynb uses Fashion MNIST dataset dataset.
   The model is built by several Conv blocks and the validation accuracy reaches 85% after a few epochs. The result:

where the darker areas in backgound imply more attention the neural network paid.

2. Class_Activation_Map_CatandDogs.ipynb uses Cats vs Dogs in the tensorflow_datasets, which is a binary classification problem.
   The validation accuracy reaches 0.87 after 25 epochs and a sample result is shown as,

Highlighted areas such as eyes and nose play an important role in classifying objects using neural networks

3. Saliency_Map.ipynb uses Inception V3 model to plot saliency maps, which also tells us what parts of the image the model is focusing on when making its predictions.

• The main difference is in saliency maps, we are just shown the relevant pixels instead of the learned features.
• one could generate saliency maps by getting the gradient of the loss with respect to the image pixels.
• changes in certain pixels that strongly affect the loss will be shown brightly in your saliency map. Here is a sample result of a running golden retriever,