Unlearning-based Neural Interpretations

TL;DR

The paper addresses the problem of unreliable gradient-based attributions caused by static baselines and high-curvature decision surfaces. It introduces UNI, a framework that learns an unlearning-based baseline by perturbing inputs toward an unlearning direction and matching activations to obtain a featureless reference, which yields low-curvature, monotonic attribution paths. UNI improves faithfulness and robustness of explanations across multiple backbones and datasets, including ImageNet-C, and shows potential extensions to NLP and mechanistic interpretability. The work demonstrates that principled baseline definitions via unlearning can markedly enhance the trustworthiness and practicality of post-hoc explanations in high-stakes AI systems.

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.

Figures (26)

• Figure 1: Left: Confidence of original model $\theta$ at image $x$ and baseline $x^{\prime}$. Right: Confidence of unlearned model $\hat{\theta}$ at image $x$. After unlearning in the model space $\theta \longmapsto \hat{\theta}$, we optimise the baseline to match the unlearned input confidence, such that $F_c(x^{\prime}; \theta) \approx F_c(x; \hat{\theta})$.
• Figure 2: We visualise post-hoc biases imposed by static baselines---black baseline (colour), blurred (texture), random (frequency). UNI learns to mask out predictive features used by the model, generating reliable attributions.
• Figure 3: When the brightness or saturation is altered, IG with a black baseline fails to identify dark features, such as the boat's hull (R3) or the top of the boot (R1).
• Figure 4: Under gaussian or defocus blur, IG with a blurred baseline suffers from saturation; has overly smooth texture; does not yield meaningful features.
• Figure 5: Gaussian and shot noise create visual artifacts prominent in noised-baseline IG. Frequency bias leads to disparate scores for adjacent pixels.