RandAugment
Algorithms AIMech:RandAugment DRAFT
An image-augmentation procedure that applies a uniformly sampled sequence of operations sharing one magnitude parameter, reducing augmentation selection to tuning the operation count and shared magnitude on the target task.
Identity Notes & Concept Scope
RandAugment is distinct from the learned operation probabilities and per-operation magnitudes of AutoAugment, the feedback-adapted parameter weights of CTAugment, and a particular library’s later RandAugment variant. Reducing the search to N and M does not eliminate hyperparameter tuning or make the procedure parameter-free. 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
For each training image, sample N operations uniformly with replacement from the selected transformation set and apply them in the sampled order, using the same magnitude parameter M with each operation’s own magnitude scale. The paper’s Figure3 uses 14 operations, including identity, geometric and appearance transformations, and permits repeated choices. The sequence therefore has K^N possible operation choices for a K-operation set, rather than sampling N distinct operations. Tune N and M using target-task validation, for example with the reported simple grid search, instead of a separate proxy-task controller search. The source also investigates different schedules for the shared magnitude. Its nominal magnitude scale and exploratory ranges are protocol choices; a shared M does not mean the same physical transformation strength across operation types.
⚠️ Curated Limitations & Negative Results
- Uniformly sampled transformations still require a suitable operation set and magnitudes for the data; label preservation is an assumption.
- N and M are hyperparameters, and the paper’s absence of a separate proxy search is not an absence of validation or model training.
- The source explores magnitude schedules and ranges beyond the nominal scale; no one schedule, numerical range or claimed optimal value is made universal.
- No benchmark reproduction or general accuracy/robustness guarantee 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 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, Jonathan Shlens, Quoc V. Le
|
primary paper
full_text
|
sha256:00796fa0e1ed…
|
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/RandAugment.yaml