linear evaluation protocol

The linear evaluation protocol is the standard yardstick for comparing self-supervised methods. The procedure is deliberately constrained: take the pretrained encoder, freeze all its weights so they cannot change, and train only a single linear layer (one fully-connected layer plus softmax) on top of its frozen features using a labelled dataset such as ImageNet. The reported number is the accuracy of that linear classifier. Because the backbone is frozen, the score measures how good the representation already is, not how well the network can adapt.

The logic is that a single linear layer can only draw flat decision boundaries, so it can only succeed if the pretrained features have already organized the data into linearly separable clusters. A high linear-probe accuracy therefore certifies that the representation has done the hard semantic work itself. It is also cheap and reproducible, only the linear layer is trained, which is why it became the field's default leaderboard metric and why contrastive methods (which explicitly optimize for separable embeddings) tend to look strong on it.

The crucial caveat is that linear probing measures one specific property, linear separability, and can badly mis-rank methods whose strengths lie elsewhere. The canonical example is MAE: it scores modestly under linear probing yet outperforms contrastive methods after full fine-tuning, because its features are rich but not arranged for a single linear cut. Practitioners therefore complement the linear probe with k-nearest-neighbour evaluation (also frozen, but nonlinear), end-to-end fine-tuning, and transfer to detection and segmentation, so that no method is judged by a single, potentially misleading number.