This repository contains all jobs for week 1 of bootcamp. The main paper is Axiomatic Attribution for Deep Networks .
The main idea behind this paper is to identify how much an inputs feature (pixels, for example) contribute to the prediction of a given input data (image for example). This is very important for the model interpretability. The method that we use is Integrated gradient.
In Axiomatic Attribution for Deep Networks , the implementations were done on:
- An Object Recognition Network
- Diabetic Retinopathy Prediction
- Question Classification
- Neural Machine Translation
- Chemistry Models
Whereas, up to now my experimentations stay on An Object Recognition Network . I keep working on all these use cases.
The original image is:
The interpolated images are:
Experimentation results: The left and right figures represent gradient method result and Integrated Gradient method results respectively. As you can see, Integrated Gradient method works better than Gradient method , because it the Integrated Gradient image seems the original image. See the Attention mask image et Attention mask + original image.
