arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2410.08069cs.LGcs.AIcs.CV
收藏

基于遗忘的神经解释

Unlearning-based Neural Interpretations

Massachusetts Institute of Technology(麻省理工学院);University of Oxford(牛津大学);University of Copenhagen(哥本哈根大学)

Ching Lam Choi, Alexandre Duplessis, Serge Belongie

AI总结:

针对基于梯度的神经解释中静态基线注入有害假设导致归因图有偏差的问题,提出UNI方法,通过输入朝遗忘方向的最陡上升扰动计算自适应基线,消除显著特征、平滑决策边界,提升解释的忠实性与鲁棒性。

摘要:

基于梯度的解释通常需要一个比较锚点来避免特征重要性计算中的饱和问题。我们发现,当前使用静态函数(常量映射、平均或模糊)定义的基线会注入偏离模型行为的有害颜色、纹理或频率假设,导致不规则梯度累积,使归因图存在偏差、脆弱且可操纵。我们脱离静态方法,提出UNI,通过将输入朝遗忘方向的最陡上升方向扰动,来计算一个(非)可学习、无偏且自适应的基线。该方法能发现可靠基线并成功消除显著特征,进而局部平滑高曲率决策边界。分析表明,遗忘是生成忠实、高效且鲁棒解释的有前景途径。

英文摘要:

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.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑