augmentation invariance

Augmentation invariance is the property that the representation barely changes when you apply a label-preserving transformation to the input. In joint-embedding self-supervised learning it is not a side effect but the entire training objective: by forcing two differently augmented views of the same image to produce nearly identical embeddings, you are explicitly teaching the encoder to ignore whatever the augmentations changed. The encoder learns: these differences do not matter, so do not encode them.

This makes the choice of augmentations the single most important inductive bias in the whole method, you literally specify what the features should be blind to. Random resized cropping teaches invariance to scale and position and is the workhorse that creates most of the difficulty. Colour jitter teaches the model not to rely on exact colour and hue. Gaussian blur reduces dependence on fine texture. Horizontal flip teaches left-right invariance. The features therefore encode what survives all these transformations, the stable object identity, while discarding what they perturb.

The flip side is that invariance can be harmful when the property you destroyed actually matters downstream. Aggressive colour augmentation builds colour-invariant features, which is great for object classification but bad for a task like flower-species recognition where colour is the signal; cropping invariance can erase the spatial information needed for localization in detection. This tension drives current research into equivariance (where the representation transforms in a known, predictable way with the input rather than staying constant) and into task-aware augmentation policies, because a single fixed invariance cannot be optimal for every downstream task at once.