Causal Analysis Framework
Multi-stage framework for identifying and validating knowledge-critical neurons in AI models.
Neuron Identification
Isolate predictive neurons using integrated gradients and knowledge probing tasks for factual correctness.
Causal Intervention
Selectively deactivate high-impact neurons to test factual degradation and observe knowledge drift.
Compare attribution methods for consistency in identifying knowledge-critical neurons against human-annotated rankings.
Attribution Validation
Neuron Ablation
Testing factual degradation by selectively deactivating high-impact neurons for insights.
Attribution Validation
Comparing attribution methods for consistency in identifying critical knowledge neurons.