Interpretable AI Series (1): overview

Machine learning and artificial intelligence are hype words today. Companies, universities, institutions, and governments are all busy making their products, services, and themselves more "smart". Recently the hottest topic in lunchtime talks is how to use chatGPT to write source codes for us programmers. However, as we become more powerful and intelligent with the help of AI products, there's always more or less dubiousness in our hearts. Is AI reliable? Should I take its suggestion or conclusion even disagreed with my intuition? How to challenge AI's decision when it's surely wrong?

Ajay Thampi's book, "Interpretable AI", published by Manning last year, was trying to answer these questions.

The first chapter gave an overview of the interpretability and explainability of AI products. Simply put, data scientists and engineers concern mainly about "interpretability": why my model emits such a classification result on this sample? Which features dominate the classification results? Is there any bias in my model? Can this model be used on that dataset confidently? How to deal with concept drift as time goes by? While business stakeholders and end users care more about "explainability":: provide reasons about the classification or regression results in a human-readable format. Ensure ML products' regulatory compliance in the target market.

How to implement interpretable ML? Besides the "normal" steps of the ML pipeline: training, validation, and deploying, this book provides some extra steps: understanding after model validation, and explaining and deploying after deploying to the production environment.

In the 3 main fields of ML, this book mainly focuses on supervised learning, which is mostly used in our ML practices and products. However, the techniques introduced in this book can be easily extended into unsupervised and reinforcement machine learning.