Concept Bottleneck Models (CBMs) are a novel neural network architecture proposed to address the “black box” problem in deep learning. The core idea is to introduce an intermediate layer of human-interpretable concepts between the input and the final prediction, making the model’s decision process transparent and explainable.
This paper addresses a critical flaw in CNN predictions: spurious correlations. The authors propose the Reasonability Matrix and the GRADIA algorithm as solutions, enabling human experts to interactively correct model attention through visual annotations.
MONET (Medical Concept Retriever) is an improved version of OpenAI’s CLIP model, specifically designed as a medical concept extractor. It assigns interpretable concept scores to medical images and enables transparent AI diagnosis through Concept Bottleneck Models (CBMs).
This paper investigates how individual neurons in deep neural networks encode semantic concepts. The core finding: whether in CNNs (for recognition) or GANs (for generation), interpretable mechanisms emerge spontaneously—individual units become carriers of specific semantic concepts.