Causal Abstraction

New method diagnoses and improves neural network interpretability

Researchers have developed a new method to diagnose and improve the interpretability of neural networks, particularly for causal abstraction tasks. This approach involves identifying specific input subspaces where a pro…