Breaking the Reclustering Barrier in Centroid-based Deep Clustering
Lukas Miklautz, Timo Klein, Kevin Sidak, Collin Leiber, Thomas Lang,, Andrii Shkabrii, Sebastian Tschiatschek, Claudia Plant
TL;DR
This paper introduces BRB, a novel algorithm that overcomes the reclustering barrier in centroid-based deep clustering, enabling continuous improvement and training from scratch, outperforming existing methods.
Contribution
The paper proposes BRB, a simple yet effective algorithm that breaks the reclustering barrier, allowing for sustained performance gains and training from scratch in centroid-based deep clustering.
Findings
BRB consistently improves clustering performance across benchmarks.
BRB enables training from scratch in deep clustering.
BRB performs competitively with state-of-the-art methods.
Abstract
This work investigates an important phenomenon in centroid-based deep clustering (DC) algorithms: Performance quickly saturates after a period of rapid early gains. Practitioners commonly address early saturation with periodic reclustering, which we demonstrate to be insufficient to address performance plateaus. We call this phenomenon the "reclustering barrier" and empirically show when the reclustering barrier occurs, what its underlying mechanisms are, and how it is possible to Break the Reclustering Barrier with our algorithm BRB. BRB avoids early over-commitment to initial clusterings and enables continuous adaptation to reinitialized clustering targets while remaining conceptually simple. Applying our algorithm to widely-used centroid-based DC algorithms, we show that (1) BRB consistently improves performance across a wide range of clustering benchmarks, (2) BRB enables training…
Peer Reviews
Decision·ICLR 2025 Poster
The problem is well presented and salient to the field. As far as I am aware, the proposed method is novel. The analysis is thorough, with ablations as well.
It would be better if the work studied some harder datasets/tasks, such as ImageNet-1k or some taxonomic image datasets (iNaturalist, BIOSCAN, etc.). A lot of the datasets considered are simpler digit classification datasets. The work could be further improved by evaluating on other modalities. **Minor points** Some references are badly formatted and need some attention to correct them. - Incorrect casing ("em algorithm", "hungarian method", "Spagcn", "rna") - DOIs shouldn't include the domain
### SIGNIFICANCE: This work is important because it seeks to overcome an important phenomenon commonly seen in centroid based deep clustering methods and this can be studied further over other recent deep clustering methods (also not centroid based as they may also face the similar issue). ### ORIGINALITY: This work is novel because it demonstrates the application of BRB over some existing methods that improves the performance further. ### CLARITY: This work is well-written and well-organized.
1. The biggest weakness is the baselines (on top of which BRB has been applied) are too old to compare, the latest one is from 2017. It would be good to see applying BRB on top of one or two recent DC method (like SCAN, GCC or SeCu mentioned in the paper) and report those results. This would provide a clearer comparison to the state-of-the-art. 2. Hyperparameter sensitivity. It would be good to report similar hyperparameters (T, α and learning rate) sensitivity (figure 8) to other one or two da
- The paper addresses an important issue in deep clustering, performance saturation. - The writing is smooth and easy to follow, with the primary contributions well-communicated. - The empirical evaluation is thorough, and the BRB algorithm is shown to enhance performance across benchmarks.
1. While the concept of reclustering is mentioned early in the paper, the term can carry specific meanings in the context of deep clustering. Typically, reclustering refers to periodically or iteratively reassigning data points to clusters. However, since DC algorithms inherently involve iterative steps of representation learning and clustering, the authors should better contextualize their use of the term. 2. Building on the previous point, many existing DC methods already incorporate some fo
Code & Models
Videos
Taxonomy
TopicsAdvanced Clustering Algorithms Research