An objective shapes hidden vectors so they retain information useful for reconstruction, prediction, contrast, or downstream tasks. Pretrained representations can then be transferred, probed, or fine-tuned.
Representation learning automatically discovers features that make relevant structure in raw data easier for a model to use.
An objective shapes hidden vectors so they retain information useful for reconstruction, prediction, contrast, or downstream tasks. Pretrained representations can then be transferred, probed, or fine-tuned.