MixUp
Algorithms AIMech:MixUp DRAFT
A training-data interpolation procedure that fits a predictor to convex combinations of example inputs and their corresponding targets using the same random mixing coefficient for both.
Identity Notes & Concept Scope
This record denotes the original input-and-target interpolation training rule, not MixMatch’s full semi-supervised algorithm, hidden-representation Manifold Mixup, patch replacement, or a trained model. Case variants of mixup are the same name. AIO DataAugmentation is relevant to the virtual-example construction, while the entire training rule also specifies a target and loss; a related mapping avoids forcing the complete procedure into AIO’s preprocessing branch. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 234 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 21 lines in each from prior discovery and the related MixMatch record. Twelve exact AIO CLASS nodes, their scoped synonyms and outgoing edges, plus six neighboring native mechanisms were inspected. This absence assessment is bounded to the pinned snapshot and inspected semantic neighborhood.
Claims & Grounded Evidence
1 Claim
Mechanism
#method
Draw pairs (x_i,y_i) and (x_j,y_j) from the training data and lambda from Beta(a,a), a>0. Construct x_tilde=lambda*x_i+(1−lambda)*x_j and y_tilde=lambda*y_i+(1−lambda)*y_j, then minimize the prediction loss on these mixed input-target pairs as an empirical vicinal risk. The classification example mixes one-hot labels into soft targets; the formulation is expressed for feature-target vectors. The source presents paired minibatches and also describes mixing a minibatch with its random shuffle. Its Figure1 uses one sampled coefficient for a paired minibatch; this is not a requirement that every variant share one coefficient. As a tends toward zero from above, the interpolation distribution concentrates at the endpoints, recovering empirical-risk training in that limit. The original rule does not replace lambda by max(lambda,1−lambda); that first-example-majority convention belongs to the later MixMatch family.
⚠️ Curated Limitations & Negative Results
- Inputs and targets need meaningful convex combinations in compatible vector spaces. Interpolations need not be natural samples or remain on a data manifold.
- Linearity between examples is an inductive bias, not a guarantee that the ground-truth conditional labels interpolate linearly or that accuracy and robustness improve.
- Beta(a,a) requires a>0; the empirical-risk limit a→0 is not a valid Beta(0,0) sampling instruction.
- No performance result, particular network architecture, fixed alpha or sampling implementation is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataAugmentation
Data Augmentation
|
related_match | The procedure constructs virtual training pairs by interpolating existing examples. The AIO target is data augmentation in a preprocessing branch, whereas this local record also specifies target interpolation and predictor training; this is a related association, not equivalence or a subclass claim. |
AIO:SupervisedLearning
Supervised Learning
|
related_match | The original method trains against interpolated observed targets. The target is the supervised learning task; this is an algorithm-to-task association. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz
|
primary paper
full_text
|
ID: arXiv:1710.09412v2
sha256:d477fc9b3c52…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T10:22:43Z
Updated: 2026-10-11T10:22:43Z
Source File:
data/curated/MixUp.yaml