Causal Abstractions of Neural Networks
This provides a formal tool for interpretability in AI, addressing the need to understand model-internal mechanisms, though it is incremental as it builds on existing structural analysis methods.
The authors tackled the problem of analyzing neural network representations by proposing a causal abstraction method that aligns neural representations with interpretable causal models and verifies their causal properties through interventions. In a case study on a complex natural language inference dataset, they found that a high-performing BERT model successfully encodes the compositional structure, while a simpler baseline does not.
Structural analysis methods (e.g., probing and feature attribution) are increasingly important tools for neural network analysis. We propose a new structural analysis method grounded in a formal theory of causal abstraction that provides rich characterizations of model-internal representations and their roles in input/output behavior. In this method, neural representations are aligned with variables in interpretable causal models, and then interchange interventions are used to experimentally verify that the neural representations have the causal properties of their aligned variables. We apply this method in a case study to analyze neural models trained on Multiply Quantified Natural Language Inference (MQNLI) corpus, a highly complex NLI dataset that was constructed with a tree-structured natural logic causal model. We discover that a BERT-based model with state-of-the-art performance successfully realizes parts of the natural logic model's causal structure, whereas a simpler baseline model fails to show any such structure, demonstrating that BERT representations encode the compositional structure of MQNLI.