시험 시간 훈련

A deployed classifier can encounter inputs that differ from its training distribution. Test-Time Training with Self-Supervision for Generalization under Distribution Shifts studies adapting model parameters on an unlabeled test sample before predicting its main label. It uses a self-supervised task whose target can be constructed without knowing that label, such as recognizing a rotation applied to an image. This is different from using the unknown correct class as training data. The reported classification improvements apply to the tested distribution shifts and setup rather than every possible deployment. Learning to Learn at Test Time develops another use of this idea for sequence modeling. A TTT layer’s hidden state is the parameter set of an inner model, updated through a self-supervised learning rule along the sequence. The paper studies a linear inner model, TTT-Linear, and a two-layer multilayer perceptron, TTT-MLP. This fixed-size recurrent state provides linear sequence-length complexity under the stated fixed model dimensions and update procedure, whereas ordinary dense attention has a quadratic pairwise term. Linear complexity does not guarantee lower latency on every device, because implementation and memory access also matter. The outer training loop optimizes the surrounding network and learned components of the inner self-supervised task for the language-model objective. The inner loop instead updates the state model’s weights on the current sequence. A sequence-state update should not be described as necessarily rewriting every deployed language-model weight or retaining personal information across every future request.