Home / Curated / Algorithms / SeLa

SeLa

Algorithms AIMech:SeLa DRAFT

Aliases: Self-labelling via simultaneous clustering and representation learning, Self-labeling via simultaneous clustering and representation learning

A self-supervised representation-learning procedure that alternates neural pseudo-label prediction and dataset-wide self-labelling through an entropy-regularized balanced optimal-transport assignment.

Identity Notes & Concept Scope SeLa denotes Asano et al.’s simultaneous clustering and representation-learning procedure. It differs from k-means-based DeepCluster and online cross-view SwAV. Its global transport formulation and the author implementation’s hardening of soft assignments are distinguished explicitly. 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

Let p_ki be the network softmax probability for class k on image i, for N images and K classes. At a self-labelling step hold model parameters fixed and form P_ki=p_ki/N. Approximately minimize <Q,−log P>+(1/lambda) KL(Q||r c^T), lambda>0, over nonnegative K-by-N Q with row masses r_k=1/K and column masses c_i=1/N. With A=P^lambda, obtain Q=diag(alpha) A diag(beta) by alternating alpha=r/(A beta) and beta=c/(A^T alpha), using elementwise division. A conditional soft code would be N Q_:i; the pinned author implementation instead takes y_i=argmax_k Q_ki, without gradients through assignment, and trains the encoder and head on these hard pseudo-labels using cross-entropy and image augmentation until the next global reassignment. This practical hardening need not preserve exact equal class counts. Multiple heads can supply separate self-labellings sharing one feature extractor.

supports Source: Self-labelling via simultaneous clustering and representation learning · Locator: Sections 3–3.5, Equations (1)–(8), PDF pp. 3–6, especially marginal constraints on p. 3 and alternating optimization on p. 4
supports Source: yukimasano/self-label author implementation: sinkhornknopp.py · Locator: sinkhornknopp.py lines 16–45, 112–138 and 150–191: full-dataset predictions, prescribed marginal factors and hard argmax labels
supports Source: yukimasano/self-label author implementation: main.py · Locator: main.py lines 79–102: no-gradient reassignment and cross-entropy training of one or several heads

⚠️ Curated Limitations & Negative Results

The paper’s displayed reciprocal Sinkhorn updates omit marginal factors; this record states the 1/K and 1/N factors required by its constraints and confirmed in the pinned CPU and GPU implementations. Probabilities, log probabilities and mass-normalized joint matrices must not be conflated.
The original hard equal-size partition assumes K divides N. For other N≥K it needs integer count capacities differing by at most one. With integer count capacities, the unregularized transport problem admits an extreme optimum for which NQ is an integral one-hot assignment matrix: the mass-normalized Q has entries 0 or 1/N. Arbitrary tied optimal mixtures and entropy-regularized plans need not be hard assignments.
Finite regularization and finite iteration approximate the original combinatorial assignment. Columnwise argmax can destroy balance; uniform soft plans are feasible. The paper’s ideal alternating-objective convergence claim is not a guarantee for this approximate hard-label/SGD implementation or against every collapsed state.
The mutual-information interpretation concerns labels and data indices under the stated marginal and fitted-model assumptions, not a measured mutual information between continuous embeddings.
Global prediction storage and reassignment scale with images and classes. Cluster count, augmentation, optimization and the balanced-class prior affect the learned grouping; semantic labels and universal benchmark improvements are not guaranteed.

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

3 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`method-source`	Self-labelling via simultaneous clustering and representation learning Yuki Markus Asano, Christian Rupprecht, Andrea Vedaldi	primary paper full_text	ID: https://arxiv.org/abs/1911.05371v3 sha256:784ac85aaf16…
`additional-1`	yukimasano/self-label author implementation: sinkhornknopp.py yukimasano/self-label contributors	official documentation documentation	ID: https://github.com/yukimasano/self-label/blob/3891e376a82d1e4ac6de6d39505b692b8885c971/sinkhornknopp.py sha256:cbf89a32a5c6…
`additional-2`	yukimasano/self-label author implementation: main.py yukimasano/self-label contributors	official documentation documentation	ID: https://github.com/yukimasano/self-label/blob/3891e376a82d1e4ac6de6d39505b692b8885c971/main.py sha256:9fe71e682454…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T08:15:29Z

Updated: 2026-10-11T08:17:24Z

Source File: data/curated/SeLa.yaml