Deep Taxonomic Networks for Unsupervised Hierarchical Prototype Discovery

2025-09-27

AI Summary

Existing deep hierarchical clustering methods suffer from two key limitations: (1) strong coupling between the learned hierarchy and a pre-specified number of categories, and (2) insufficient exploitation of prototype information at intermediate hierarchical levels. To address these, we propose a deep latent variable model based on variational autoencoders, incorporating a differentiable full binary-tree Gaussian mixture prior over latent variables. This enables joint learning of hierarchical latent representations and prototype clusters. We theoretically prove that ELBO optimization inherently induces hierarchical relationships among prototypes without supervision. The method requires neither class labels nor prior knowledge of the number of categories, and supports end-to-end multi-granularity feature learning. Extensive experiments on multiple image benchmarks demonstrate significant improvements over state-of-the-art baselines in both hierarchical clustering accuracy and interpretability of the induced classification structure. To our knowledge, this is the first approach to automatically discover, from unlabeled data, a semantically coherent, fine-grained, and prototype-based hierarchical taxonomy.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Representation Learning for Vision
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods, algorithms and applications for the development of semantic models, knowledge graphs and other forms of structured data models with machine-interpretable semantics
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Inspired by the human ability to learn and organize knowledge into hierarchical taxonomies with prototypes, this paper addresses key limitations in current deep hierarchical clustering methods. Existing methods often tie the structure to the number of classes and underutilize the rich prototype information available at intermediate hierarchical levels. We introduce deep taxonomic networks, a novel deep latent variable approach designed to bridge these gaps. Our method optimizes a large latent taxonomic hierarchy, specifically a complete binary tree structured mixture-of-Gaussian prior within a variational inference framework, to automatically discover taxonomic structures and associated prototype clusters directly from unlabeled data without assuming true label sizes. We analytically show that optimizing the ELBO of our method encourages the discovery of hierarchical relationships among prototypes. Empirically, our learned models demonstrate strong hierarchical clustering performance, outperforming baselines across diverse image classification datasets using our novel evaluation mechanism that leverages prototype clusters discovered at all hierarchical levels. Qualitative results further reveal that deep taxonomic networks discover rich and interpretable hierarchical taxonomies, capturing both coarse-grained semantic categories and fine-grained visual distinctions.

Problem

• Automatically discovers hierarchical taxonomic structures from unlabeled data
• Addresses underutilization of prototype information in intermediate levels
• Learns interpretable hierarchies capturing coarse and fine-grained distinctions

Innovation

• Deep taxonomic networks with binary tree latent hierarchy
• Variational inference optimizes mixture-of-Gaussian prior structure
• Unsupervised prototype discovery across all hierarchical levels