DeepCluster
Algorithms AIMech:DeepCluster DRAFT
Aliases: Deep clustering for unsupervised learning of visual features
A visual representation-learning procedure that alternates clustering current neural features into pseudo-labels and training the feature extractor and a classifier to predict those labels.
Identity Notes & Concept Scope
This record covers the original k-means-based DeepCluster procedure of Caron et al. It is not generic deep clustering, standalone k-means, DeepCluster-v2, or a particular pretrained model. Cluster assignments provide training targets; the k-means centroid matrix is not used as the discriminative classifier. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 195 prior native IDs/names/aliases. Ignored-inclusive, word-delimited searches of both full checkouts found nine prior discovery/review references in each plus six unrelated language-stemmer tokens in the main environment. An initial substring search also matched unrelated layer-name suffixes. Neighboring k-means, SwAV, Barlow Twins and VICReg mechanisms and exact AIO definitions/edges were inspected; shared components do not establish identity.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
Extract features of the unlabeled training images with the current encoder while holding its parameters fixed. In the original reported configuration, PCA-reduce and whiten these features, then L2-normalize them before k-means. Cluster them by alternating nearest-centroid assignment and centroid updates toward the squared-distance objective; use the resulting hard cluster indices as pseudo-labels. The paper handles empty clusters by perturbing a selected nonempty centroid and reassigning its members between the two centers. Then hold pseudo-labels fixed and optimize the encoder and a classification head with cross-entropy on augmented images. Sample approximately uniformly across nonempty pseudo-label groups so large clusters do not dominate training. Repeat full-dataset feature extraction, clustering and discriminative training. The pinned author implementation reinitializes the final linear classifier for each reclustering round, uses 20 k-means iterations, and uses sampled pseudo-labels rather than differentiating through k-means.
⚠️ Curated Limitations & Negative Results
- K-means is nonconvex and the finite iteration implementation does not certify a global optimum. The alternating clustering and classification stages are not a proof of descent for one common objective or of general convergence.
- Empty-cluster repair and pseudo-label sampling are practical safeguards, not a universal guarantee against collapsed representations. The inspected wrapper delegates clustering to Faiss; its library internals were not audited.
- Uniform training exposure to pseudo-label groups does not force equal cluster sizes. Cluster IDs can be permuted or change between rounds and are not verified semantic class labels.
- PCA dimension 256, Sobel filtering, architecture, augmentations and epoch schedule describe the original experimental configuration. Whitening and L2 normalization require suitable numerical handling; no performance equivalence for arbitrary variants is asserted.
- Training discards ImageNet class labels, but the paper selects hyperparameters using labeled Pascal VOC validation. Full-dataset feature extraction has a cost; no training or benchmark reproduction was performed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:SelfsupervisedLearning
Self-supervised Learning
|
related_match | This algorithm learns from self-generated targets or relationships between unlabeled inputs. The pinned AIO target is a MachineLearningTask; this is a related algorithm-to-task mapping, not equivalence or an is-a assertion. |
AIO:UnsupervisedClustering
Unsupervised Clustering
|
related_match | This procedure uses or directly learns unlabeled image cluster assignments. The pinned target is an unlabeled clustering task; this is a related algorithm-to-task association, not equivalence, a semantic class-label guarantee or an is-a assertion. |
Evidence Sources & Primary Literature
4 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Mathilde Caron, Piotr Bojanowski, Armand Joulin, Matthijs Douze
|
primary paper
full_text
|
ID: https://openaccess.thecvf.com/content_ECCV_2018/html/Mathilde_Caron_Deep_Clustering_for_ECCV_2018_paper.html
sha256:cc527dbc1ccb…
|
additional-1
|
facebookresearch/deepcluster contributors
|
official documentation
documentation
|
ID: https://github.com/facebookresearch/deepcluster/blob/2d1927e8e3dd272329e879e510fbbdf1b1d02d17/main.py
sha256:2ca6ba2def53…
|
additional-2
|
facebookresearch/deepcluster contributors
|
official documentation
documentation
|
ID: https://github.com/facebookresearch/deepcluster/blob/2d1927e8e3dd272329e879e510fbbdf1b1d02d17/clustering.py
sha256:a7d07cb901f5…
|
additional-3
|
facebookresearch/deepcluster contributors
|
official documentation
documentation
|
ID: https://github.com/facebookresearch/deepcluster/blob/2d1927e8e3dd272329e879e510fbbdf1b1d02d17/util.py
sha256:f47aaf8fcb74…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T08:15:29Z
Updated: 2026-10-11T08:15:29Z
Source File:
data/curated/DeepCluster.yaml