I'm the author here! Thanks for sharing :) Feel free to ask me any questions.

I wanted to highlight some of the libraries we used: Gradio [1] to create the interfaces in the GIFs, and PAIR-code's [2] Tensorflow implementation of the papers.

[1]: https://github.com/gradio-app/gradio [2]: https://github.com/PAIR-code/saliency

Really quality write up here. That being said, I’m not convinced that interpretation offers any real value in industry ML - are there any examples of these techniques actually being used by companies and for what purpose?

Useful for finding label leaks in medical images (burned in radiologist marks for example)