Contrastive learning frameworks typically create multiple views of the same unlabeled data by applying augmentations, such as jittering or permutation for time-series data. The training objective is to learn representations that are invariant to those augmentations, which strengthens the encoder against perturbations and helps extract robust and discriminative features.

According to the source, contrastive learning frameworks generally generate multiple views of unlabeled data using augmentations. They then encourage the model to learn representations that remain invariant to these changes. This training objective improves the encoder's ability to resist perturbations and to extract robust and discriminative features. In the context of multivariate time-series data, common augmentations include time-domain-based operations such as jittering or permutation, which create different views of the same sequence. The encoder is trained to keep temporal patterns stable under these perturbations.

Key points

Multiple views are created through data augmentations applied to unlabeled samples.
The typical training objective is invariance: representations should not change when the data view changes.
This objective makes the encoder more resistant to perturbations.
It also supports learning robust and discriminative features.
For multivariate time-series, augmentations often include jittering and permutation to preserve temporal consistency.

How do contrastive learning frameworks typically create multiple views of unlabeled data, and what is the training objective?

Related questions