Research map / Label-efficient and robust learning / Recognition and long-tail

You Can Trust Your Clustering Model: A Parameter-free Self-Boosting Plug-in for Deep Clustering

NeurIPS 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 0 citations (OpenAlex)

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 DCBoost, a parameter-free plug-in that uses adaptive per-batch k-NN consistency filtering to select high-confidence samples and a discriminative loss (positive, negative prototype, and instance consistency terms) to improve global structure of existing deep clustering models.

Problem

Existing deep clustering models learn reliable local neighborhood structure but poor global structure (weak intra-class compactness and inter-class separation), and heuristic sample filtering typically requires manually tuned hyperparameters.

Approach

Select high-confidence samples by checking whether all k nearest neighbors within a mini-batch share the sample's pseudo-label, choosing k adaptively by maximizing k times the selected fraction, then apply a BYOL-style online/target discriminative loss (class-weighted positive pull, prototype repulsion, instance consistency) on those samples as a plug-in to existing models.

Key results

• Applied to ProPos, DCBoost raises average performance across five datasets from 79.2 to 82.2 (+3.0) and ProPos CIFAR-10 ACC from 94.4% to 96.0%.
• Applied to BYOL, average improves from 72.7 to 77.3 (+4.6); applied to CC, from 68.2 to 71.3 (+3.1).
• On CIFAR-10 with ProPos, silhouette coefficient rises from 0.10 to 0.74 per the introduction; paired t-tests over five runs show significant gains (p<0.05) for all combinations.

Measured effects reported in the paper

Surprising findings

• Removing the negative (prototype repulsion) loss without sample selection causes partial clustering collapse, dropping performance below baseline.
• Removing the instance consistency loss L_ins collapses STL-10 ACC from 93.6% to 20.5%.

Limitations

• Assumes the number of clusters is known.
• Lacks mechanisms for highly imbalanced or non-uniform data distributions (gains on long-tailed data are shown but not tailored).
• Requires within-batch same-class samples, so large-scale datasets with many classes need a queue and fixed k.

Experimental setup

Datasets: CIFAR-10 CIFAR-20 STL-10 ImageNet-10 ImageNet-Dogs Tiny-ImageNet CIFAR-100 ImageNet-1K CIFAR10-LT CIFAR20-LT

Benchmarks: CIFAR-10 CIFAR-20 STL-10 ImageNet-10 ImageNet-Dogs Tiny-ImageNet CIFAR-100 ImageNet-1K

Metrics: NMI ACC ARI silhouette coefficient k-NN accuracy

Baselines: CC SCAN CDC BYOL CoNR ProPos NNM GCC IDFD TCL TCC SPICE

Builds on: BYOL ProPos CC SCAN CoNR

Closest papers by meaning

• From Logits to Hierarchies: Hierarchical Clustering made Simple (ICML 2025 · similarity 0.87)
• Breaking the Reclustering Barrier in Centroid-based Deep Clustering (ICLR 2025 · similarity 0.86)
• Towards Calibrated Deep Clustering Network (ICLR 2025 · similarity 0.86)
• Learning from Sample Stability for Deep Clustering (ICML 2025 · similarity 0.85)
• Hard Regularization to Prevent Deep Online Clustering Collapse without Data Augmentation (AAAI 2024 · similarity 0.85)
• Exploring a Principled Framework for Deep Subspace Clustering (ICLR 2025 · similarity 0.85)