Missed Causes and Ambiguous Effects: Counterfactuals Pose Challenges for Interpreting Neural Networks

TL;DR

This paper critically examines the limitations of using counterfactuals for interpreting neural networks, highlighting issues like missed causes and non-transitive dependencies that complicate causal analysis.

Contribution

The paper identifies key problems with counterfactual approaches in neural network interpretability and discusses their implications, offering suggestions for future research directions.

Findings

Counterfactual methods often miss causes due to multiple sufficient causes.

Counterfactual dependencies in neural networks are generally non-transitive.

These issues bias causal interpretations and complicate causal graph extraction.

Abstract

Interpretability research takes counterfactual theories of causality for granted. Most causal methods rely on counterfactual interventions to inputs or the activations of particular model components, followed by observations of the change in models' output logits or behaviors. While this yields more faithful evidence than correlational methods, counterfactuals nonetheless have key problems that bias our findings in specific and predictable ways. Specifically, (i) counterfactual theories do not effectively capture multiple independently sufficient causes of the same effect, which leads us to miss certain causes entirely; and (ii) counterfactual dependencies in neural networks are generally not transitive, which complicates methods for extracting and interpreting causal graphs from neural networks. We discuss the implications of these challenges for interpretability researchers and propose concrete suggestions for future work.