Test Time Training with Self Supervision for Generalization under Distribution Shifts
What this work does
Jointly trains supervised and self-supervised tasks, then adapts shared parameters on an unlabeled test input using the auxiliary loss before prediction.
Connection to discovery
Suggests learning an update rule whose auxiliary signal can be obtained from observations when the desired test label is unavailable.
Evidence
Corruption benchmarks compare frozen joint training, per-input adaptation and persistent online adaptation.
Scope and limitations
Auxiliary improvement must align with the actual task. Persistence across inputs changes the information budget, and the visual rotation task does not directly supply an LLM discovery objective.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.