M15 / LABORATORY

Augmentation. Free data,
from the data you already have.

Show a model the same image flipped, jittered, blurred, cut — and it learns that an "object" isn't a particular set of pixels but an invariant beneath transformations. Build a pipeline. Roll a batch. See it.

EXPERIMENT 08.01 · AUGMENTATION.PIPE STAGES · 5 BATCH · 16

PIPELINE

SOURCE

BATCH · 16 INDEPENDENT ROLLS apply pipeline w/ random params

Same source · 16 random samples · each result is what the model sees as a "different" training example.

STAGE-BY-STAGE · single roll, each stage applied in order

Discipline → Augmentation should be invariant to the label. Flipping a cat horizontally? Still a cat. Flipping a "6"? Now it's a "9" — wrong label. Pick augmentations that preserve the semantics you care about.