ViP-CBM: Reducing Parameters in Concept Bottleneck Models by Visual-Projected Embeddings

With the widespread application of deep neural networks (DNN) in key areas of personal and property security, the interpretability and trustworthiness of DNN models become increasingly significant. Concept Bottleneck Models (CBM) are popular interpretable models in which hidden layer neurons are trained to predict human-understandable concepts of the final task. However, existing CBMs encounter low efficiency and interpretability in the multi-label classification (MLC) of concepts: correlations of concepts are ignored by many CBMs, while others expressing correlations with complex models have very limited improvements in the model’s performance. To address the challenge of massive parameters and limited interpretability in concept MLC problem, we propose a novel Visual-Projected CBM (ViP-CBM), which transforms MLC of concepts into an input-dependent binary classification problem of concept embeddings using visual features for projection. Our ViP-CBM model reduces the training parameter set by more than 50% with the number of training parameters when compared to other embedding-based CBMs. Experimental results show that our ViP-CBM achieves comparable or even better performance in concept prediction.