Neural Networks Decoded: Targeted and Robust Analysis of Neural Network Decisions via Causal Explanations and Reasoning
Primary Area: causal reasoning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: AI Explainability, Causal Discovery, Counterfactuals Reasoning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Despite their continued success and widespread adoption, deep neural networks (DNNs) remain enigmatic “black boxes” due to their complex architectures and opaque decision-making processes taking place within, which poses significant trust challenges to critical applications. While various methods have been proposed to address this lack of interpretability, existing solutions often offer inconsistent, or overly simplified explanations, or require alterations that compromise model performance. In this work, we introduce TRACER, a novel explainability method grounded in causal inference theory and designed to shed light on the causal dynamics underpinning DNN decisions without altering their architecture or compromising their performance. We further propose an efficient methodology for counterfactual generation, offering contrastive explanations for misclassifications, thereby identifying potential model biases. Through comprehensive evaluations across diverse datasets, we demonstrate the superiority of TRACER compared to prevalent explainability methods, and underscore its ability to transcend explainability from mere associations to causal relationships. We subsequently highlight TRACER’s potential to enable the creation of highly compressed and highly efficient models, showcasing its versatility in both understanding and optimizing DNNs.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2791
Loading