Unlearning-based Neural Interpretations

2024-10-10

arXiv.org

AI Summary

Gradient attribution methods rely on baselines to mitigate feature saturation, yet static baselines—such as constants, means, or blurred images—introduce strong priors (e.g., color and texture), leading to biased, fragile, and manipulable attributions. To address this, we propose UnlearnBase: the first framework to introduce “unlearning” into interpretability. It dynamically generates learnable, input-adaptive debiased baselines via gradient ascent perturbations on the input, locally smoothing high-curvature decision boundaries. The method comprises differentiable baseline parameterization, curvature-driven optimization of the decision boundary, and a robustness-aware attribution evaluation. Extensive experiments demonstrate that UnlearnBase significantly outperforms mainstream static baselines in faithfulness, stability, and adversarial robustness. It consistently improves attribution quality across multiple benchmarks while maintaining computationally tractable overhead.

Technology Category

• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Gradient-based interpretations often require an anchor point of comparison to avoid saturation in computing feature importance. We show that current baselines defined using static functions--constant mapping, averaging or blurring--inject harmful colour, texture or frequency assumptions that deviate from model behaviour. This leads to accumulation of irregular gradients, resulting in attribution maps that are biased, fragile and manipulable. Departing from the static approach, we propose UNI to compute an (un)learnable, debiased and adaptive baseline by perturbing the input towards an unlearning direction of steepest ascent. Our method discovers reliable baselines and succeeds in erasing salient features, which in turn locally smooths the high-curvature decision boundaries. Our analyses point to unlearning as a promising avenue for generating faithful, efficient and robust interpretations.

Problem

• Improve gradient-based interpretations accuracy
• Eliminate harmful static function assumptions
• Generate reliable and robust neural interpretations

Innovation

• Unlearning-based baseline computation
• Debiased adaptive baseline generation
• High-curvature boundary smoothing