AutoAugment
Algorithms AIMech:AutoAugment DRAFT
An automated augmentation-policy search procedure that trains a controller to propose image-transform subpolicies and rewards those proposals using validation accuracy of a child model trained with the proposed augmentation.
Identity Notes & Concept Scope
AutoAugment denotes a learned policy-search process and its resulting policies, not a single image transform or a network checkpoint. It differs from RandAugment’s uniform operation sampling and compact target-task hyperparameter search, and from CTAugment’s online prediction-feedback adaptation. Reinforcement learning is used to choose augmentations here; the child’s classification task is not thereby reclassified as reinforcement learning. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 238 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 34 lines in each from prior discovery, saved reviews and related native records. Ten exact AIO CLASS nodes, their scoped synonyms and outgoing edges, plus five 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
In the published search space, a policy contains five subpolicies, each with two ordered operations. An operation has a type, application probability and magnitude; the search uses 16 operation types, 11 probability values and 10 magnitude values, with magnitude ignored when an operation does not need it. During child-model training, select a subpolicy uniformly for each image and apply its operations according to their respective probabilities. An autoregressive recurrent controller samples the policy decisions. Train the child model on the training split with that policy, measure held-out validation accuracy, and use that reward to update the controller with proximal policy optimization. The described controller emits 30 decisions for the five two-operation subpolicies. After search, concatenate the subpolicies from the five best policies into a 25-subpolicy policy for final model training. Particular child architectures, search budgets and augmentation baselines are experimental settings rather than the identity of the procedure.
⚠️ Curated Limitations & Negative Results
- Searching by training child models requires computation, and a policy selected on a smaller proxy model or dataset may not transfer optimally to the final training setting.
- Operation probabilities and magnitudes are task-sensitive; an operation can destroy class information.
- The record follows the augmentation decision space and does not import the paper’s unrelated convolutional-cell/block wording in the controller hyperparameter paragraph.
- No claim of globally optimal policies, fixed controller architecture, universal policy transfer or benchmark improvement is made.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataAugmentation
Data Augmentation
|
related_match | This algorithm selects or learns transformations used for augmented training examples. The AIO target denotes data augmentation in the preprocessing branch, while the local identity includes a selection or training procedure; no equivalence or subclass membership is asserted. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Ekin D. Cubuk, Barret Zoph, Dandelion Mané, Vijay Vasudevan, Quoc V. Le
|
primary paper
full_text
|
sha256:9a8a7cb244b0…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T10:34:59Z
Updated: 2026-10-11T10:34:59Z
Source File:
data/curated/AutoAugment.yaml