.png){kind=icon}
Return
TopicCL: Visual concept explanation with topic modeling
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2026.114757.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Attribution-based explanation methods are employed to generate saliency heatmaps, thereby highlighting the regions in an image that have a significant impact on model decisions. However, these methods only reveal the regions attended to by the model, and the semantic content perceived by the model is difficult to clarify. To address this challenge, the potential advantages of integrating topic modeling with visual concept explanation are explored, and an unsupervised concept explanation framework called Topic Concept Learning (TopicCL) is proposed. Within TopicCL, a Concept Variational Autoencoder (C-VAE) is designed based on the four fundamental properties of concepts, including causality, interactivity, consistency, and completeness, and a hyperspherical constraint is imposed on the latent space by using a spherical Cauchy (spCauchy) distribution. Concepts are then clustered on the hypersphere through variational inference to reflect the semantic preferences of the model. To quantify the contribution of concepts to model decisions, a method named Spherical Ablation Scoring (SAS) is proposed, which can measure concept importance by changes in model outputs induced by tangential perturbations on the hypersphere. In the experiments, a completeness metric termed mean Reconstruction Error (mRE) is introduced to quantify the predictive information contained in concepts. Experimental results demonstrate that TopicCL exhibits broad applicability and can be seamlessly integrated into various visual neural networks, achieving excellent performance in both faithfulness and purity evaluations.
Keywords:
Visual concept explanation
Topic concept learning
Visual neural network
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
Cited Papers
No cited papers available