MixMatch
Algorithms AIMech:MixMatch DRAFT
A semi-supervised classifier-training procedure that averages predictions over augmented unlabeled inputs, sharpens the resulting soft targets, mixes augmented labeled and unlabeled input-target pairs, and optimizes labeled cross-entropy plus an unlabeled squared-probability loss.
Identity Notes & Concept Scope
MixMatch combines averaged augmented-view label guessing, probability sharpening and the modified MixUp pairing rule. These components define a distinct training algorithm rather than an alias for MixUp, Mean Teacher or a hard pseudo-label confidence filter. Its classification soft targets differ from representation matching in BYOL, DINO and SimSiam. This record describes a training algorithm, not a trained checkpoint or fixed network architecture. The inspected AIO SemiSupervisedLLM class specifically combines LLM pretraining with supervised fine-tuning; its related synonym self-training is not an exact identity for this method. AIO SelfsupervisedLearning excludes human-annotated labels, Distillation describes large-to-small compression, and RegularizationLayer is a layer: none supplies an equivalent class for this algorithm. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 230 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout; matches concern prior discovery leads and the Pi-model normalization regression. Ten exact AIO CLASS nodes, their scoped synonyms and outgoing edges, plus five neighboring native records were inspected. The corrected Unicode-aware normalizer preserves Π; the previous Pi-model/Model collision was a lexical defect, not semantic identity (review issue #58).
Claims & Grounded Evidence
1 Claim
Mechanism
#method
Given B labeled examples with one-hot labels and B unlabeled examples, augment each labeled input once and each unlabeled input K times. Average the current classifier probabilities across the K views of each unlabeled input, then sharpen the average q_bar using q_c=q_bar_c^(1/T)/sum_j q_bar_j^(1/T), for T>0; treat these guessed labels as constants during differentiation. Create augmented labeled pairs X_hat of size B and unlabeled pairs U_hat of size KB, then shuffle their concatenation into W. Mix each X_hat pair with the corresponding first B pairs in W, and each U_hat pair with the corresponding remaining KB pairs. For each MixUp operation draw lambda from Beta(alpha,alpha), alpha>0, set lambda_prime=max(lambda,1−lambda), and combine both inputs and targets with weights lambda_prime and 1−lambda_prime, keeping at least half of the first pair. For the resulting X_prime and U_prime, minimize average soft-target cross-entropy over X_prime plus lambda_u/(C|U_prime|) times sum over U_prime of squared target/prediction differences, where C is the number of classes. The reported implementation ramps lambda_u linearly. Guessed labels use current-model predictions; the evaluation-only weight EMA in Section 4.1 is not a Mean Teacher target generator.
⚠️ Curated Limitations & Negative Results
- The original setup pairs B labeled examples with B unlabeled examples and K augmentations per unlabeled input. Alternative batching or interpolation rules require an explicit variant scope.
- For 0<T<1 sharpening concentrates probability. As a curator consequence of the formula, the T→0 limit is one-hot only with a unique maximum; tied maxima retain equal limiting mass. T>1 smooths rather than sharpens.
- Correct label-preserving augmentation and useful interpolation are assumptions; guessed labels can be wrong and the algorithm has no defining hard confidence threshold.
- The record does not assert published error rates or universal superiority. Reported T, K, alpha, lambda_u and evaluation-EMA settings are experimental choices, not fixed constants of every implementation.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:Classification
Classification
|
related_match | This procedure learns class predictions from labeled and unlabeled examples. The pinned AIO target describes a classification task under SupervisedLearning, not this semi-supervised training algorithm; this is an algorithm-to-task association, not equivalence or an is-a assertion. |
AIO:MachineLearningTask
Machine Learning Task
|
related_match | The algorithm specifies how to train a predictor from data. The AIO target is the general learning task; this association does not identify an algorithm with that task or assert an ontology subclass edge. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
David Berthelot, Nicholas Carlini, Ian Goodfellow, Avital Oliver, Nicolas Papernot, Colin Raffel
|
primary paper
full_text
|
sha256:096c017ec11f…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T10:12:04Z
Updated: 2026-10-11T10:12:04Z
Source File:
data/curated/MixMatch.yaml