‘A lot of papers attempt to understand model behavior through the scientific approach to Machine Learning. Using this approach will ensure whatever you learn will never go absolute. Very often newcomers to the field & experienced researchers experience stress due to the rapid
development in the AI field where new findings that are presented become obsolete in a month with a new algorithm. Many contemplate is the effort of investment into this space from a scientist or engineer perspective worthwhile?
A solution is to combine what you’re doing with an understanding of why things are working. If you can understand the reason for your model improving, that understanding will outlive that particular model and how widely it will be used.’

Human wisdom on machine learning from Professor Andrew Wilson from NYU on ‘The Machine Learning Street Talk’ podcast