[论文解读] Explaining Classifiers with Causal Concept Effect (CaCE)

Yash Goyal, Amir Feder+2 more|arXiv (Cornell University)|Jul 16, 2019

Explainable Artificial Intelligence (XAI)参考文献 24被引用 92

一句话总结

本论文将 CaCE 定义为人类可解释概念对分类器输出的因果效应，并提出基于 VAE 的方法来估计 CaCE，减缓全局解释中的混淆。

ABSTRACT

How can we understand classification decisions made by deep neural networks? Many existing explainability methods rely solely on correlations and fail to account for confounding, which may result in potentially misleading explanations. To overcome this problem, we define the Causal Concept Effect (CaCE) as the causal effect of (the presence or absence of) a human-interpretable concept on a deep neural net's predictions. We show that the CaCE measure can avoid errors stemming from confounding. Estimating CaCE is difficult in situations where we cannot easily simulate the do-operator. To mitigate this problem, we use a generative model, specifically a Variational AutoEncoder (VAE), to measure VAE-CaCE. In an extensive experimental analysis, we show that the VAE-CaCE is able to estimate the true concept causal effect, compared to baselines for a number of datasets including high dimensional images.

研究动机与目标

• Define the causal concept effect (CaCE) as the average causal effect of a binary or categorical concept on a classifier’s output.
• Propose a framework to estimate CaCE using generative models to approximate the image generation process.
• Show that CaCE estimates can reduce confounding compared to correlation-based methods across varied datasets.
• Provide diagnostic tests to increase confidence in CaCE estimates.
• Demonstrate CaCE estimation on high-dimensional image data and discuss its applicability to black-box classifiers.

提出的方法

• Introduce CaCE as E[f(I)|do(C0=1)] − E[f(I)|do(C0=0)], the average treatment effect of a concept on the classifier output.
• Model the image generation process with a conditional VAE (DC-VAE) conditioned on concepts and class labels to approximate p(I|C0, L).
• Propose Dec-CaCE, using only the VAE decoder to generate counterfactual images for CaCE estimation.
• Propose EncDec-CaCE, using both VAE encoder and decoder to estimate CaCE for specific images or sets of images.
• Provide diagnostic tests: (I) positive effect (concept equals label) and (II) null effect (random dummy concept).
• Evaluate CaCE with GT-CaCE in controlled settings and compare to ConExp and TCAV across datasets.

实验结果

研究问题

• RQ1 Can CaCE quantify the causal impact of a high-level concept on a classifier’s output rather than mere correlations?
• RQ2 How well can a conditional VAE approximate the true image-generating process to estimate CaCE?
• RQ3 Do Dec-CaCE and EncDec-CaCE provide unbiased or more accurate CaCE estimates compared to baselines?
• RQ4 Do diagnostic tests help identify when CaCE estimates may be confounded or unreliable?
• RQ5 How do CaCE estimates behave across synthetic and real-world high-dimensional image datasets?

主要发现

• CaCE estimates (via Dec-CaCE and EncDec-CaCE) align with ground-truth CaCE in controlled datasets and tend to be lower thanCorrelation-based baselines when confounding is present.
• Dec-CaCE generally outperforms EncDec-CaCE in matching GT-CaCE on the BARS and Colored-MNIST datasets.
• CaCE estimates using the proposed methods stay closer to ground truth than ConExp and TCAV in high-dimensional COCO-Miniplaces and CelebA settings.
• CaCE tends to increase with classifier complexity, and richer generative models (convolutional DC-VAE) yield estimates closer to GT-CaCE than simpler architectures.
• Diagnostic tests can flag potential failures of the VAE-based approach, highlighting its limitations under strong confounding.