linear probing evaluation

Linear probing is the standard yardstick for asking how good a learned representation is on its own terms. You freeze the pretrained backbone so its weights cannot change, take the features it produces, and train only a single linear layer (a logistic regression) on top to predict labels. The accuracy of that thin classifier is a proxy for how linearly separable — how readily usable — the frozen features already are, before any task-specific adaptation.

The discipline of freezing is the whole point. If you fine-tuned the backbone, a strong final accuracy could come from the backbone reshaping itself during fine-tuning rather than from the quality of the pretrained features, confounding what you are trying to measure. By allowing only a linear head, the protocol isolates the representation: good self-supervised methods are expected to place semantically similar inputs in linearly separable regions, so a simple boundary suffices. It is cheap, reproducible, and comparable across methods, which is why nearly every SSL paper reports it.

Its limitations are worth naming. Linear probe accuracy and fine-tuning accuracy can rank methods differently — MAE, for instance, probes modestly but fine-tunes excellently, because its features are rich but not linearly organized. Probing also rewards features tuned to the specific probe dataset's classes and can miss qualities (like local spatial structure) that matter for detection or segmentation.