Discriminative Feature Attributions: Bridging Post Hoc Explainability and Inherent Interpretability

TL;DR

Distractor Erasure Tuning (DiET) is proposed, a method that adapts black-box models to be robust to distractor erasure, thus providing discriminative and faithful attributions and naturally combines the ease of use of post hoc explanations with the faithfulness of inherently interpretable models.

Abstract

With the increased deployment of machine learning models in various real-world applications, researchers and practitioners alike have emphasized the need for explanations of model behaviour. To this end, two broad strategies have been outlined in prior literature to explain models. Post hoc explanation methods explain the behaviour of complex black-box models by identifying features critical to model predictions; however, prior work has shown that these explanations may not be faithful, in that they incorrectly attribute high importance to features that are unimportant or non-discriminative for the underlying task. Inherently interpretable models, on the other hand, circumvent these issues by explicitly encoding explanations into model architecture, meaning their explanations are naturally faithful, but they often exhibit poor predictive performance due to their limited expressive power. In this work, we identify a key reason for the lack of faithfulness of feature attributions: the lack of robustness of the underlying black-box models, especially to the erasure of unimportant distractor features in the input. To address this issue, we propose Distractor Erasure Tuning (DiET), a method that adapts black-box models to be robust to distractor erasure, thus providing discriminative and faithful attributions. This strategy naturally combines the ease of use of post hoc explanations with the faithfulness of inherently interpretable models. We perform extensive experiments on semi-synthetic and real-world datasets and show that DiET produces models that (1) closely approximate the original black-box models they are intended to explain, and (2) yield explanations that match approximate ground truths available by construction. Our code is made public at https://github.com/AI4LIFE-GROUP/DiET.

Figures (8)

• Figure 1: Illustration of our method, Distractor Erasure Tuning. DiET models exhibit robustness to distractor erasure (non-discriminative features such as backgrounds), allowing for the recovery of discriminative attributions.
• Figure 2: Visualization of datasets and attribution methods considered in this work. Row 1: Raw data samples, Row 2: Ground truth attributions, Row 3: DiET attributions, Rows 4-7: Baseline methods.
• Figure 3: Pixel perturbation tests (higher is better) for MNIST (left), Chest X-ray (middle), and CelebA (right) datasets. DiET's recommended mask sparsity is shown as a vertical dashed line. We observe that DiET performs the best overall. Refer to Section \ref{['correctness']} for details.
• Figure 4: Pixel perturbation test and example images for models trained with gradient manipulation on Hard MNIST. Results for other datasets are in the appendix. Refer to Section \ref{['sec:manip']} for details.
• Figure 5: Ablation study of pixel perturbation test with varying levels of mask downscaling for MNIST and CelebA. Example images for each factor are shown on the right.

Theorems & Definitions (9)

• Definition 1
• Definition 2
• Definition 3
• Theorem 1
• Corollary
• Theorem 2
• proof
• Theorem 3
• proof