通过分解显著与非显著特征，提升模型对关键信息的捕捉能力。
H-SPLID: HSIC-based Saliency Preserving Latent Information Decomposition
- 将显著与非显著特征显式分离到不同空间，实现可解释性增强
- 实验显示模型对背景等非显著区域扰动敏感度降低
- 适用于需要鲁棒性和可解释性的图像分类任务
我们提出H-SPLID，一种通过将显著与非显著特征显式分解到独立空间来学习显著特征表示的新算法。研究表明，H-SPLID能促进学习低维、任务相关特征。我们证明，在输入扰动下，预测偏差的期望值被显著子空间维度和输入与表示间的希尔伯特-施密特独立性准则（HSIC）所上界控制，建立了鲁棒性与潜在表示压缩之间的联系。在图像分类任务上的实证评估表明，使用H-SPLID训练的模型主要依赖于显著输入成分，表现为对影响非显著特征（如图像背景）的扰动敏感度下降。代码已开源：https://github.com/neu-spiral/H-SPLID。
原文摘要 · Abstract (English)
We introduce H-SPLID, a novel algorithm for learning salient feature representations through the explicit decomposition of salient and non-salient features into separate spaces. We show that H-SPLID promotes learning low-dimensional, task-relevant features. We prove that the expected prediction deviation under input perturbations is upper-bounded by the dimension of the salient subspace and the Hilbert-Schmidt Independence Criterion (HSIC) between inputs and representations. This establishes a link between robustness and latent representation compression in terms of the dimensionality and information preserved. Empirical evaluations on image classification tasks show that models trained with H-SPLID primarily rely on salient input components, as indicated by reduced sensitivity to perturbations affecting non-salient features, such as image backgrounds. Our code is available at https://github.com/neu-spiral/H-SPLID.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。