AdaMixUp
Algorithms AIMech:AdaMixUp DRAFT
Aliases: Adaptive MixUp
An adaptive interpolation-training procedure that learns example-dependent mixing-coefficient regions with a policy network and trains a classifier together with an intrusion discriminator intended to separate mixed examples from original data.
Identity Notes & Concept Scope
AdaMixUp learns an input-dependent mixing region and uses an intrusion objective, unlike original MixUp’s globally selected Beta distribution or Manifold Mixup’s hidden-layer interpolation. An intrusion discriminator does not make it a GAN: the generator and discriminator are trained toward separating mixtures from originals rather than adversarially making mixtures indistinguishable from real data. The mathematical distinct-class setup is relaxed in the paper’s implementation footnote. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 247 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 6 lines in the worktree and 6 in main. These were prior candidate leads, not concept records. 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
2 Claims
Mechanism
#method
For a pair of examples, a policy-region network outputs a three-component softmax (alpha,delta,remainder). Interpret the first two components as the interval(alpha,alpha+delta) for the mixing coefficient. Sample epsilon uniformly on(0,1) and set gamma=alpha+delta*epsilon so gradients can pass through the interval parameters. Form gamma*x1+(1−gamma)*x2 and the correspondingly mixed class targets. Train the classifier with observed and mixed targets, and train the policy generator together with an intrusion discriminator to distinguish generated mixtures from original examples. The discriminator shares the classifier’s feature network while using its own final classification layers. For higher-fold mixing, reuse the same pairwise generator recursively by mixing an original example with an original or previously mixed example; composing a previous coefficient vector lambda with coefficient gamma gives (gamma*lambda,1−gamma). The paper describes alternating minibatch optimization of classification/mixup loss and intrusion loss. Its intended intrusion supervision treats mixed examples as outside and observed examples as inside the data manifold; this is a learned surrogate, not direct access to the true manifold.
Mechanism
#implementation-loss-conventions
The pinned Fold-2 reimplementation generates interval parameters with softmax, samples gamma by reparameterization and uses the same coefficient to mix inputs and labels. Its classifier uses softmax cross-entropy over an equal-sized mixture of generated examples and one source group. Its binary intrusion head uses softmax cross-entropy, labels originals1 and mixtures0, and averages over two original groups and one mixed group. The returned cost sums the classification and intrusion means and a weight-decay term. This realizes the separation intention using negative log-probabilities, with source-specific label encoding and weighting rather than literal positive-log minimization or a demonstrated reproduction of the paper’s alternating update schedule.
⚠️ Curated Limitations & Negative Results
- The claimed avoidance of manifold intrusion depends on assumptions, generator/discriminator capacity and optimization. A learned discriminator cannot certify that every mixed point is outside the true manifold or that its soft target is correct.
- Equation5 prints positive log-probabilities despite calling the quantity cross-entropy and later minimizing it. The prose and pinned implementation support minimizing negative log-probability; the discrepancy is preserved rather than treating the printed sign as an executable loss or asserting an official erratum.
- The pinned public repository describes itself as a reimplementation. It uses binary labels opposite to the paper’s arbitrary0/1 convention and combines classification and intrusion batch means in one cost plus weight decay, while the paper describes alternating objectives.
- The code’s classification batch combines equal original/mixed counts, while its discriminator uses twice as many original as mixed examples. Those batch means are not asserted to equal the separately averaged terms of printed Equations5–6.
- The recursive mixture coefficients are nonnegative and sum to one, but this algebra alone proves neither semantic validity nor improved generalization. The record follows the verbal recursion and coefficient construction rather than the dimensionally inconsistent matrix-vector ordering in one argument of printed Equation7.
- No universal architecture, fold count, performance gain or execution of the downloaded author code is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataAugmentation
Data Augmentation
|
related_match | This algorithm learns transformations or synthetic examples for training. The AIO target denotes augmentation in the preprocessing branch; the local procedure includes policy learning and model training, so the association is related rather than equivalent or an asserted subclass. |
AIO:SupervisedLearning
Supervised Learning
|
related_match | The classifier fits observed and interpolated class targets. This relates the algorithm to a supervised learning task at a different abstraction level; it does not assert identity or a subclass relation. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Hongyu Guo, Yongyi Mao, Richong Zhang
|
primary paper
full_text
|
sha256:9d97175f9bf0…
|
ada-code
|
SITE5039/AdaMixUp contributors
|
official documentation
documentation
|
sha256:f8acb2389c7a…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T10:55:22Z
Updated: 2026-10-11T10:55:22Z
Source File:
data/curated/AdaMixUp.yaml