In many applications, deep learning classifiers such as CNNs and Vision Transformers achieve high accuracy but operate as black box models with little transparency. Regulators and stakeholders increasingly demand interpretable predictions to ensure fairness, accountability, and compliance with industry standards. ViNCE extracts high-dimensional features from a pre-trained teacher network, applies PCA to reduce dimensionality, and then uses a visual-language model to generate human-readable descriptors for each principal component based on feature activation maps. A decision tree student model is trained on those features to match the teacher’s decisions, allowing users to see exactly which semantic attributes (e.g., “edge,” “texture,” “object part”) drive classification. By providing transparent, tree-based explanations, ViNCE addresses the critical need for interpretability in regulated environments like finance and healthcare, eliminating the mystery of black-box models and enabling rapid auditability, and informed decision-making across diverse use cases.