On Measuring Intrinsic Causal Attributions in Deep Neural Networks

TL;DR

This paper introduces a new framework for measuring intrinsic causal contributions in neural networks, treating them as structural causal models, and shows that it provides more intuitive explanations than existing methods.

Contribution

It proposes an identifiable generative post-hoc framework for quantifying intrinsic causal contributions and relates ICC to Sobol' indices, advancing causal interpretability in neural networks.

Findings

ICC provides more intuitive explanations than existing methods

The framework is applicable to synthetic and real-world datasets

ICC is related to Sobol' indices for causal attribution

Abstract

Quantifying the causal influence of input features within neural networks has become a topic of increasing interest. Existing approaches typically assess direct, indirect, and total causal effects. This work treats NNs as structural causal models (SCMs) and extends our focus to include intrinsic causal contributions (ICC). We propose an identifiable generative post-hoc framework for quantifying ICC. We also draw a relationship between ICC and Sobol' indices. Our experiments on synthetic and real-world datasets demonstrate that ICC generates more intuitive and reliable explanations compared to existing global explanation techniques.