SamplePairing
Algorithms AIMech:SamplePairing DRAFT
An image-classification augmentation procedure that averages an anchor image with a randomly selected training image using equal pixel weights and trains against the anchor image’s label, with a training schedule that includes phases without pairing.
Identity Notes & Concept Scope
SamplePairing is the equal-image-weight, anchor-label procedure of Inoue, distinct as a training protocol from the standard Beta-distributed input/target MixUp construction. This distinction does not imply that every expected objective differs: under exchangeable pairing and a loss linear in targets, averaging the anchor-label losses over both orders matches the half-and-half target loss for the same averaged image. The paper itself discusses this relationship. It is also distinct from learned Smart Augmentation and from spatial patch replacement in CutMix. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 253 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 27 lines in each checkout. These were neighboring-record and review references or prior candidate leads, not existing records for these identities. Twelve exact AIO CLASS nodes with synonyms and outgoing edges and six neighboring native mechanisms were inspected. Absence is bounded to this snapshot and semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
For a training anchor, randomly select a partner from the training set without requiring the same class. Apply ordinary image augmentations to the inputs when used, then average corresponding pixels and channels with equal weights. Train on the mixed image using only the anchor label; the partner label is not required by this rule. The published training schedule starts without SamplePairing, enables it after a warm-up, intermittently disables it during training, and finishes with a no-pairing fine-tuning phase. The source uses one warm-up epoch for ILSVRC and 100 for the other datasets, then alternates 300,000 paired / 100,000 unpaired images for ILSVRC or eight paired / two unpaired epochs otherwise. These numerical intervals are experimental choices, not the identity of every use. A learned augmenter, a random mixing ratio, same-class pairing and label interpolation are not required by the original procedure.
⚠️ Curated Limitations & Negative Results
- Averaged images can contain incompatible classes; the assigned anchor label is a training convention, not a claim that the mixture has a unique correct semantic label.
- The paper’s N-squared count describes ordered pairing choices. Commutativity and duplicate inputs can make different pairings produce the same averaged image, so it is not a guarantee of N-squared distinct visual samples.
- The expected-loss relation to half-and-half soft labels requires the stated symmetry and loss assumptions. It does not equate individual sampled gradients, arbitrary pair distributions, the full training schedules or the original randomly weighted MixUp procedure.
- The source’s training-accuracy ceiling discussion is idealized and does not establish a universal bound for every finite training run. No benchmark advantage, training-cost equivalence or medical-task result is imported.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataAugmentation
Data Augmentation
|
related_match | This procedure constructs training examples and specifies how a predictor is trained with them. The AIO target is augmentation in a preprocessing branch; this related association preserves the broader training and target-handling scope without asserting equivalence or a subclass relation. |
AIO:SupervisedLearning
Supervised Learning
|
related_match | The predictor learns from class labels associated with generated or combined training examples. This is an algorithm-to-task association, not identity with the general supervised-learning task. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Hiroshi Inoue
|
primary paper
full_text
|
ID: arXiv:1801.02929
sha256:af0e6c947d83…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T11:19:25Z
Updated: 2026-10-11T11:19:25Z
Source File:
data/curated/SamplePairing.yaml