NBDT: Neural-Backed Decision Trees for Interpretable Image Classification

This article introduces NBDT (Neural-Backed Decision Trees), a novel model proposed by UC Berkeley and Boston University in April 2020. NBDT combines neural networks with decision trees to provide interpretable image classification results. The model's key advantage is balancing model accuracy with interpretability, allowing it to not only classify images but also explain the hierarchical reasoning process.

The NBDT framework consists of three main steps: (1) Pretraining a CNN model and using the final layer weights as category embeddings, (2) Using hierarchical clustering and WordNet to form an induced hierarchy tree structure, and (3) Incorporating the tree structure's classification loss into the total loss and fine-tuning the model. The model uses either "Hard" or "Soft" modes for both prediction and loss calculation.

The article provides detailed explanations of the NBDT algorithm, including code analysis of the core functions such as building the induced graph, forward probability calculation, and total loss computation. Experimental results show that NBDT achieves state-of-the-art performance while maintaining interpretability. The model can be used for various classification tasks beyond image classification, making it valuable for applications requiring explainable AI decisions.