From Dependence to Causation
PhD thesis — University of Cambridge, 2016. Full text: https://archive.org/details/arxiv-1607.03300
Arguments
From Dependence to Causation
Dependence is the difference between seeing and doing — and any system that acts in the world (an agent, a policy, a treatment) needs causal structure, not just predictions.
Machine learning is stuck at prediction (dependence) and can't reach machine reasoning without causal inference — uncovering cause-effect structure so systems can answer "what if" questions under interventions; the thesis advances causal inference three ways: a framework for statistical dependence from copulas and random features, causal inference reframed as distribution classification (yielding a family of novel algorithms), and causal discovery inside convolutional neural network features — with scalable algorithms, theoretical guarantees, and state-of-the-art real-world benchmarks.
Why it matters
Dependence is the difference between seeing and doing — and any system that acts in the world (an agent, a policy, a treatment) needs causal structure, not just predictions.
Use when: Prediction tells you what will happen; causation tells you what happens if you act — only the second one is useful to a decision-maker.