Apprentissage auto-supervisé

Self-supervised training constructs a prediction or comparison task from available data. This reduces the need for separate task annotations at that stage, while the data still requires collection, processing, and appropriate use. A derived training target is still supervision: self-supervised does not mean that learning has no objective or no target signal. BERT's masked-language-model objective selects token positions and predicts their original token identities using contextual information. The original procedure does not replace every selected position with a mask symbol. It can instead insert a random token or leave the token unchanged, while retaining the original token as the prediction target. That distinction separates the corrupted input from the training answer. Autoregressive language modeling uses another arrangement, predicting the next token from preceding text. SimCLR uses paired augmented views of an image. Two views from the same source image form a positive pair, and the contrastive objective compares their representations with other views under the defined setup. A learned nonlinear projection head transforms the encoder representation for that training objective. The original framework discards that projection head when using the encoder representation for downstream tasks. This is a particular architectural choice, not a required component of all self-supervised methods. Learned representations can be evaluated with a downstream classifier or adapted through fine-tuning, but transfer quality and label requirements depend on the task. Speech work such as wav2vec 2.0 provides another research example, with masked speech representations and subsequent supervised adaptation. Success on a pretraining objective does not itself establish reliable performance on every later application.