ML (Machine Learning) has enjoyed tremendous practical success in the last decade with applications ranging from e-commerce to self-driving cars. The success of deep networks in vision and speech recognition are particularly notable examples. However, the theoretical understanding and characterization of these techniques has not kept pace with the real-world achievements of ML. Traditional approaches such as generalization bounds, stability-based justifications, capacity arguments, regularization etc., have only gone part way towards rationalizing the uncanny success of modern ML methods. In addition there are a number of empirical phenomena such as adversarial examples, effectiveness of gradient descent, failure of NLP GANS, etc., deserving of deeper observational scrutiny and rigorous theoretical treatment. Lastly, there are a number of desirable properties such as explainability, debuggability, ability to selectively ignore discriminatory biases in data, verifiability, etc., that st...
