A Handbook for Deep Learning with their Piecemeal Intuitions from Causal Theory

Our paper available at: “A Handbook for Deep Learning with their Piecemeal Intuitions from Causal Theory” (my Computational Learning Theory course essay at Oxford).

Deep learning has been fully adopted in various applications nowadays. On the other hand, causality, a powerful weapon to describe the relationship between causes and effects, is gaining increasing attention. Recent works begin to adopt intuitions from causal theory in order to improve deep learning, and the results are optimistic.

We first introduce causal theory basics, then classify these works as improving

1. Out-of-distribution (OOD) generalization;
2. Generation;
3. Robustness, interpretability, and fairness.

In addition, we explicitly point out the causal intuitions in these works, describing

• What causal intuitions are embraced and
• How do they help improve deep learning.

We hope this “handbook” may help both beginners and researchers understand the underlying causal principles and see the promising future of deep learning with causality.