SimSiam
Algorithms AIMech:SimSiam DRAFT
A self-supervised representation-learning procedure that trains a shared encoder and predictor to match the encoder output of another augmented view using symmetric negative cosine similarity with stopped gradients on each target output.
Identity Notes & Concept Scope
This is Chen and He’s SimSiam procedure, not generic Siamese-network architecture or a checkpoint. The two views share encoder weights directly; the target output is detached only for its directional loss. There is no separately averaged target encoder or explicit negative-pair term. 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
Generate two augmented views v1 and v2 of each image. A shared encoder f, including backbone and projection MLP, produces z1=f(v1) and z2=f(v2); a shared prediction MLP h produces p1=h(z1) and p2=h(z2). For nonzero vectors let D(p,z)=−(p/||p||)·(z/||z||). Average [D(p1,stopgrad(z2))+D(p2,stopgrad(z1))]/2 over images and update f and h by gradients. Each view’s encoder output is constant on its target side but receives gradients through its prediction side in the opposite term. The original procedure uses no momentum-averaged encoder, negative dictionary or clustering assignment step.
⚠️ Curated Limitations & Negative Results
- Stop-gradient changes the optimization dynamics but does not remove constant representations from the numerical loss’s minimizers. Section 5.3 explicitly treats noncollapse as empirical and says the alternating-optimization hypothesis does not explain its prevention.
- The predictor is trained; removing it or fixing random predictor weights is not the defined procedure. A symmetric target stop does not mean a branch never receives gradients.
- Batch normalization, augmentation, architecture, optimizer and scheduling affect the reported results. Absence of explicit negative pairs does not imply independent per-example learning when batch statistics are used.
- The record covers the original symmetric cosine formulation. Asymmetric, cross-entropy and multistep-alternation experiments are not silently made equivalent. Nonzero-norm handling is required; no universal performance or convergence guarantee 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. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Xinlei Chen, Kaiming He
|
primary paper
full_text
|
ID: https://openaccess.thecvf.com/content/CVPR2021/html/Chen_Exploring_Simple_Siamese_Representation_Learning_CVPR_2021_paper.html
sha256:89e52731a2da…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T08:04:24Z
Updated: 2026-10-11T08:04:24Z
Source File:
data/curated/SimSiam.yaml