Homepage

Blog

Causal Structures in Neural Networks

Do-calculus as an epistemic tool for reading model decisions — not a dashboard feature.

A neural network can be accurate and still leave us without a usable account of why a decision obtained. Causal language promises more than correlation heatmaps: interventions, counterfactuals, and graphs that make assumptions explicit.

This essay is forthcoming. The working claim is simple: causal structure is valuable when it constrains what we are allowed to say — not when it merely decorates a prediction.