Efficient Discovery of Approximate Causal Abstractions via Neural Mechanism Sparsification

Researchers have developed a new method to extract interpretable causal mechanisms from neural networks using structured pruning as a search technique. The approach reframes network pruning as finding approximate causal abstractions, yielding closed-form criteria for simplifying networks while maintaining their causal structure under interventions.

• →The method treats trained neural networks as deterministic Structural Causal Models to discover interpretable mechanisms.
• →An Interventional Risk objective is derived whose second-order expansion provides closed-form criteria for network simplification.
• →Under uniform curvature conditions, the scoring method reduces to activation variance, explaining when variance-based pruning works or fails.
• →The technique efficiently extracts sparse, intervention-faithful abstractions from pretrained networks without requiring retraining.
• →The approach was validated through interchange interventions, demonstrating practical applicability for neural network interpretability.