Research map / Label-efficient and robust learning / Recognition and long-tail
From Logits to Hierarchies: Hierarchical Clustering made Simple
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Logits to Hierarchies (L2H), a lightweight training-free algorithm that builds a dendrogram from the logits of a pre-trained flat clustering model (TURTLE, TEMI) or supervised classifier, outperforming deep hierarchical clustering methods on CIFAR-10/100 and Food-101.
Problem
Deep specialized hierarchical clustering models scale poorly to large datasets and trail flat clustering models at leaf-level quality; a simple way to derive hierarchies from existing models is needed.
Approach
Iteratively merge groups of clusters using only logits: the lowest-scoring group is merged with the group that receives the most reassigned predicted probability from its data points when its own clusters are masked out. No fine-tuning or internal representations are needed.
Key results
- On CIFAR-100, L2H-TURTLE reaches DP 0.803 and LHD 0.235 versus 0.121 DP-related deep baseline DeepECT (LHD 0.121 is underlined as degenerate) and TreeVAE DP 0.103 / LHD 0.484.
- L2H-TURTLE on ImageNet-1K achieves NMI 0.882, ACC 0.726 and DP 0.560, with the full hierarchy computed in about 5 minutes including TURTLE training.
- L2H-TURTLE on CIFAR-10 obtains DP 0.988 and LHD 0.277, with leaf-level NMI 0.985, versus DP 0.341 for TreeVAE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply L2H to TURTLE instead of using TreeVAE | hierarchical clustering · CIFAR-10 | DP: 0.341 -> 0.988 | Helped |
| apply L2H to TURTLE instead of using TreeVAE | hierarchical clustering · CIFAR-100 | DP: 0.103 -> 0.803 | Helped |
| apply L2H to TEMI instead of using TreeVAE | hierarchical clustering · Food-101 | DP: 0.016 -> 0.801 | Helped |
| apply L2H to TURTLE instead of using TreeVAE | hierarchical clustering · CIFAR-100 | LHD: 0.484 -> 0.235 | Mixed |
| consider all group pairs for merging rather than selecting the lowest-scoring group first (ablation) | L2H-TURTLE · CIFAR-100 | DP: 0.803 -> 0.797 | No clear effect |
| consider all group pairs for merging rather than selecting the lowest-scoring group first (ablation) | L2H-TURTLE · CIFAR-100 | LHD: 0.235 -> 0.246 | Hurt |
| use the designated aggregation function, summing per-cluster mean probabilities, rather than averaging across groups and clusters (ablation) | L2H-TURTLE · CIFAR-100 | DP: 0.793 -> 0.803 | Helped |
| use the designated aggregation function, summing per-cluster mean probabilities, rather than averaging across groups and clusters (ablation) | L2H-TURTLE · CIFAR-100 | LHD: 0.283 -> 0.235 | Helped |
| set the leaf-cluster count K to the true class count rather than over- or underestimating it (ablation) | L2H-TURTLE · CIFAR-100 | hierarchical and flat clustering metrics | Helped |
| use L2H-TURTLE from the family-level model instead of training a separate TURTLE model at each coarser taxonomy level | TURTLE · INaturalist21, Order | ARI: 0.075 -> 0.097 | Helped |
| use L2H-TURTLE from the family-level model instead of training a separate TURTLE model at each coarser taxonomy level | TURTLE · INaturalist21, Kingdom | NMI: 0.479 -> 0.561 | Helped |
Surprising findings
- Deep hierarchical methods DeepECT and TreeVAE show much weaker leaf-level clustering than flat non-hierarchical models, and DeepECT often collapses to shallow trees on colored datasets.
- Training TURTLE at the finest iNaturalist21 family level and applying L2H to derive coarser levels outperformed separately trained TURTLE models at each coarse taxonomy level.
Limitations
- Relies on the assumption that logits from a pre-trained flat model are a good proxy for cluster similarity, which may fail if the model is poorly calibrated.
- Selecting meaningful hierarchy levels still requires manual inspection of the tree.
- Evaluated only on vision datasets.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 Food-101 ImageNet-1K iNaturalist21
Benchmarks: CIFAR-10 CIFAR-100 Food-101 ImageNet-1K iNaturalist21
Metrics: NMI ARI ACC Leaf Purity Dendrogram Purity Least Hierarchical Distance
Baselines: Agglomerative clustering HypHC DeepECT TreeVAE TCL Lowe et al (2024)
Builds on: TURTLE TEMI HypHC TreeVAE DeepECT
Closest papers by meaning
- Objective-Based Hierarchical Clustering of Deep Embedding Vectors (AAAI 2021 · similarity 0.91)
- Ultrametric Cluster Hierarchies: I Want ‘em All! (NeurIPS 2025 · similarity 0.90)
- Classification Fields: Arbitrarily Fine Recursive Hierarchical Clustering From Few Examples (NeurIPS 2026 · similarity 0.89)
- Expected Probabilistic Hierarchies (NeurIPS 2024 · similarity 0.88)
- Deep Taxonomic Networks for Unsupervised Hierarchical Prototype Discovery (NeurIPS 2025 · similarity 0.87)
- You Can Trust Your Clustering Model: A Parameter-free Self-Boosting Plug-in for Deep Clustering (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.