CutMix
Algorithms AIMech:CutMix DRAFT
A training augmentation procedure that replaces a rectangular region of one image with the corresponding region of another and mixes their class targets in proportion to the retained and pasted image areas.
Identity Notes & Concept Scope
CutMix is spatial patch mixing with area-weighted targets. It is distinct from whole-image intensity averaging, zero masking, learned synthesis and hidden-representation interpolation. The record describes input-image CutMix; feature-level variants are not silently treated as an identical architecture or layer. Its official implementation profile is pinned separately from the proceedings formula. No normalized identity match was found across all 442 pinned AIO labels/scoped synonyms and 253 preceding native IDs/names/aliases. Full-checkout fixed-string searches included ignored and hidden files in both worktree and main checkout, returning 27 lines in each checkout. These were neighboring-record and review references or prior candidate leads, not existing records for these identities. 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
Choose two training images and a binary spatial mask M. Construct M*x_A+(1−M)*x_B, preserving anchor pixels outside the rectangle and using donor pixels inside it. Train against lambda*y_A+(1−lambda)*y_B, with lambda representing the retained anchor area fraction. The paper draws a nominal coefficient from Beta(alpha,alpha), samples the rectangle center uniformly over the image, and sets nominal side lengths to the image dimensions times sqrt(1−lambda). The published experiments use alpha=1. The donor patch is taken from the same spatial coordinates as the overwritten anchor region. Unlike MixUp, individual pixels are selected from one source rather than blended throughout the image; unlike Cutout, the removed region is filled with another training image and target weights reflect both sources. On a discrete image, clipping and integer box construction can change the nominal area, so the inspected implementation recomputes the target coefficient from the actual rectangle.
Mechanism
#clipped-area-implementation
When beta is positive and the minibatch application draw passes cutmix_prob, the pinned implementation samples a Beta coefficient, permutes the batch, constructs a shared rectangle and replaces that region with the permuted donor pixels. rand_bbox converts each nominal side length to an integer, uses half-width floor division on each side of a random integer center and clips endpoints to the image dimensions. It then sets lambda=1−rectangle_area/image_area and sums lambda times the anchor cross-entropy with 1−lambda times the donor cross-entropy. With this shared coefficient, the objective equals cross-entropy against the correspondingly mixed targets. Training continues by backpropagation; ordinary unmodified inputs and labels are used when the application condition fails.
⚠️ Curated Limitations & Negative Results
- Patch area is a heuristic for class-target contribution; it need not equal the visible object area or semantic information retained. A pasted patch may contain background or remove the discriminative object region.
- Integer side lengths, floor division and boundary clipping make the actual area differ from the nominal Beta draw. The realized target coefficient is not generally distributed as the original Beta variable.
- The pinned training code uses one application decision, rectangle and coefficient for a minibatch and a random permutation for partners. Per-example variants and feature-level CutMix require separate version-specific descriptions.
- The script’s beta argument controls the Beta distribution, while alpha controls PyramidNet width; these names are not the paper’s mathematical alpha convention. Default beta and cutmix_prob are zero, so the augmentation must be enabled by configuration.
- The implementation was read, not executed. Old API usage and unspecified runtime dependencies preclude an execution guarantee. Logged training error against the anchor labels is distinct from the mixed-target training objective. No universal localization, robustness or performance improvement is asserted.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DataAugmentation
Data Augmentation
|
related_match | This procedure constructs training examples and specifies how a predictor is trained with them. The AIO target is augmentation in a preprocessing branch; this related association preserves the broader training and target-handling scope without asserting equivalence or a subclass relation. |
AIO:SupervisedLearning
Supervised Learning
|
related_match | The predictor learns from class labels associated with generated or combined training examples. This is an algorithm-to-task association, not identity with the general supervised-learning task. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
method-source
|
Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, Youngjoon Yoo
|
primary paper
full_text
|
ID: arXiv:1905.04899
sha256:5ba5b422109b…
|
cutmix-train
|
clovaai/CutMix-PyTorch contributors
|
official documentation
documentation
|
sha256:049026c02005…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T11:19:25Z
Updated: 2026-10-11T11:19:25Z
Source File:
data/curated/CutMix.yaml