arXiv 2202.034822022-02cs.CVcs.AIcs.LG

在神经空间中导航：重新审视概念激活向量以克服方向发散

Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence

Fraunhofer Heinrich Hertz Institute（弗劳恩霍夫海因里希·赫兹研究所）；BIFOLD – Berlin Institute for the Foundations of Learning and Data（BIFOLD——柏林学习与数据基础研究所）

Frederik Pahde, Maximilian Dreyer, Leander Weber, Moritz Weckbecker, Christopher J. Anders, Thomas Wiegand, Wojciech Samek, Sebastian Lapuschkin

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

本文针对概念激活向量因可分离性优化而受干扰方向影响导致方向发散的问题，提出仅聚焦概念信号的基于模式的 CAV，并在多个数据集和模型架构上验证了其在概念敏感性测试和模型校正中的优势。

摘要：

随着人们对理解神经网络预测策略的兴趣日益增长，概念激活向量（CAVs）已成为在潜在空间中建模人类可理解概念的流行工具。通常，CAV 的计算方式是借助线性分类器，优化含有和不含给定概念的样本潜在表示的可分离性。然而，在本文中我们表明，这种面向可分离性的计算会导致解偏离精确建模概念方向的真实目标。这种偏差可归因于干扰方向的显著影响，即与概念无关的信号，它们被线性模型的滤波器（即权重）拾取以优化类别可分离性。为了解决这一问题，我们引入了基于模式的 CAV，仅聚焦于概念信号，从而提供更准确的概念方向。我们评估了各种 CAV 方法在它们与真实概念方向的对齐程度以及它们对 CAV 应用的影响方面的表现，包括概念敏感性测试和针对由数据伪影引起的捷径行为的模型校正。我们使用 Pediatric Bone Age、ISIC2019 和 FunnyBirds 数据集，以 VGG、ResNet、ReXNet、EfficientNet 和 Vision Transformer 作为模型架构，展示了基于模式的 CAV 的优势。

英文摘要：

With a growing interest in understanding neural network prediction strategies, Concept Activation Vectors (CAVs) have emerged as a popular tool for modeling human-understandable concepts in the latent space. Commonly, CAVs are computed by leveraging linear classifiers optimizing the separability of latent representations of samples with and without a given concept. However, in this paper we show that such a separability-oriented computation leads to solutions, which may diverge from the actual goal of precisely modeling the concept direction. This discrepancy can be attributed to the significant influence of distractor directions, i.e., signals unrelated to the concept, which are picked up by filters (i.e., weights) of linear models to optimize class-separability. To address this, we introduce pattern-based CAVs, solely focussing on concept signals, thereby providing more accurate concept directions. We evaluate various CAV methods in terms of their alignment with the true concept direction and their impact on CAV applications, including concept sensitivity testing and model correction for shortcut behavior caused by data artifacts. We demonstrate the benefits of pattern-based CAVs using the Pediatric Bone Age, ISIC2019, and FunnyBirds datasets with VGG, ResNet, ReXNet, EfficientNet, and Vision Transformer as model architectures.

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。