The Ladder of Causation
Causal Inference, developed extensively by Judea Pearl, transitions machine learning from statistical curve-fitting (correlations) to the modeling of underlying mechanisms (causality).
Levels of Causal Reasoning
- Association: Identifying statistical relationships between variables (typical of standard neural networks).
- Intervention: Predicting the effects of direct changes to a system (modeled via the
do-calculus). - Counterfactuals: Retrospective analysis of specific outcomes to determine the influence of alternate actions ("What if I had acted differently?").
Historical & Alternative Methods
- Structural Equation Modeling (SEM): A statistical technique that combines factor analysis and multiple regression analysis to examine structural relationships between variables.
- Granger Causality: A statistical hypothesis test for determining whether one time series is useful in forecasting another, widely used in economics and neuroscience.
- Instrumental Variables (IV): A method used in econometrics to estimate causal relationships when controlled experiments are not possible.
- Potential Outcomes Framework (Rubin-Neyman): An alternative mathematical approach to causality focusing on the comparison of outcomes under different treatment assignments.
Technical Objective
Standard LLMs operate primarily at the level of association. The integration of Structural Causal Models (SCMs) enables agents to predict action consequences and provide mechanistic explanations for their decision-making processes based on cause-effect relationships.
Libraries & Links
- DoWhy (GitHub): A framework for causal inference combining graphical models and potential outcomes.
- PyWhy: An ecosystem for causal machine learning research.
- Causal Inference (Wikipedia)