On the Mechanistic Interpretability of Neural Networks for Causality in Bio-statistics
It addresses the need for interpretability in high-stakes health applications, but is incremental as it applies existing MI techniques to a specific domain.
This work tackles the challenge of interpreting neural networks for causal inference in bio-statistics by applying mechanistic interpretability techniques, demonstrating that these tools can probe internal representations, visualize computational pathways, and compare insights across models to enhance validation and trust.
Interpretable insights from predictive models remain critical in bio-statistics, particularly when assessing causality, where classical statistical and machine learning methods often provide inherent clarity. While Neural Networks (NNs) offer powerful capabilities for modeling complex biological data, their traditional "black-box" nature presents challenges for validation and trust in high-stakes health applications. Recent advances in Mechanistic Interpretability (MI) aim to decipher the internal computations learned by these networks. This work investigates the application of MI techniques to NNs within the context of causal inference for bio-statistics. We demonstrate that MI tools can be leveraged to: (1) probe and validate the internal representations learned by NNs, such as those estimating nuisance functions in frameworks like Targeted Minimum Loss-based Estimation (TMLE); (2) discover and visualize the distinct computational pathways employed by the network to process different types of inputs, potentially revealing how confounders and treatments are handled; and (3) provide methodologies for comparing the learned mechanisms and extracted insights across statistical, machine learning, and NN models, fostering a deeper understanding of their respective strengths and weaknesses for causal bio-statistical analysis.