Home / Curated / Algorithms / SwAV

SwAV

Algorithms AIMech:SwAV DRAFT

Aliases: Swapping assignments between views

An online self-supervised visual representation-learning procedure that predicts soft prototype assignments of one augmented image view from other views, using entropy-regularized balanced assignment and jointly learned features and prototypes.

Identity Notes & Concept Scope SwAV means swapping assignments between views. It uses online soft prototype codes rather than offline full-dataset k-means labels, direct pairwise feature negatives, or an exponential-moving-average teacher. Its prototype assignments are learned training targets, not verified semantic class labels. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 191 prior native IDs/names/aliases. Ignored-inclusive searches of both full checkouts found only prior discovery leads (four hits per checkout) and one unrelated main-environment DFSWAVE token. Neighboring SimCLR, MoCo, BYOL and DINO procedures and exact AIO definitions/edges were inspected; shared components do not establish identity.

Claims & Grounded Evidence

1 Claim

Mechanism #method

Encode augmented views into nonzero features and L2-normalize them; compare them with K trainable prototype columns C. For a group of B features Z with positive integer B and K and epsilon>0, approximately maximize Tr(Q^T C^T Z)+epsilon H(Q) over nonnegative Q with row sums 1/K and column sums 1/B, where H(Q)=−sum_(k,b) Q_kb log Q_kb. Compute the soft transport assignment by Sinkhorn row/column rescaling of exp(C^T Z/epsilon), then normalize each column into a probability code q_b (q_b=B Q_:b at exact column balance). Stop gradients through codes. For each image, predict a global view’s code from the other views using p(v)=softmax(C^T z(v)/tau), tau>0, and cross-entropy −sum_k q_k log p_k, averaged over images and valid cross-view pairs. In multicrop training codes are computed from the two global crops; predictions use global and smaller crops while excluding the same crop as its code target. Update encoder/projection parameters and prototypes by gradients and normalize prototype columns, as in the supplement. The small-batch variant augments assignment features with a recent-feature queue but uses only current-image codes in the loss.

supports Source: Unsupervised Learning of Visual Features by Contrasting Cluster Assignments · Locator: Sections 3.1–3.2, Equations (1)–(6), main PDF pp. 3–5, with Section 3.2 continuation on p. 6
supports Source: Supplementary Material for Unsupervised Learning of Visual Features by Contrasting Cluster Assignments · Locator: Appendix A.1 training loop, stopped codes, prototype normalization and Sinkhorn pseudocode, supplement PDF pp. 1–2; Appendix A.2 crop distributions, p. 2

⚠️ Curated Limitations & Negative Results

Finite Sinkhorn iterations approximate the balanced transport solution. Probability codes must sum to one per image; the mass-1 transport matrix with column sums 1/B is not already that probability matrix. Exponential scores need numerically stable evaluation.
The source describes equipartition as preventing identical codes, but also reports uniform collapse with strong entropy regularization. Uniform soft assignments satisfy the balance constraints, so balance alone is not a general noncollapse theorem.
The source motivates a feature queue when B is small relative to K. This is an empirical assignment-quality issue; the continuous transport polytope is not mathematically empty merely because B<K. Queued features can be stale and are not a momentum encoder.
The main equations and supplement pseudocode use different compact sum/mean conventions. This record states averaged cross-entropy explicitly and does not transfer optimizer coefficients between differently scaled losses.
Prototype number, assignment regularization, crop distribution, initial prototype freezing and optimization schedule are experimental choices. No semantic clustering guarantee, universal batch-size independence or reproduced benchmark is asserted.

Artificial Intelligence Ontology (AIO) Grounding

Mapped AIO Classes

AIO Target Class	Mapping Relation	Mapping Rationale
`AIO:SelfsupervisedLearning` Self-supervised Learning	related_match	This is an algorithm for learning representations from relationships between unlabeled views. 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	SwAV uses balanced online soft assignments to learned prototypes in its representation-learning procedure. 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

2 Sources

Source ID	Publication / Document	Type / Access	Reference & Checksum
`method-source`	Unsupervised Learning of Visual Features by Contrasting Cluster Assignments Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, Armand Joulin	primary paper full_text	ID: https://proceedings.neurips.cc/paper/2020/hash/70feb62b69f16e0238f741fab228fec2-Abstract.html sha256:a18124b18b01…
`additional-1`	Supplementary Material for Unsupervised Learning of Visual Features by Contrasting Cluster Assignments Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, Armand Joulin	primary paper full_text	ID: https://proceedings.neurips.cc/paper_files/paper/2020/file/70feb62b69f16e0238f741fab228fec2-Supplemental.pdf sha256:8cd2a74ab798…

Curation Provenance

Curator: Codex

Method: ai_assisted

Model: GPT-6

Created: 2026-10-11T08:04:24Z

Updated: 2026-10-11T08:05:31Z

Source File: data/curated/SwAV.yaml