Test-Time Training with Masked Autoencoders
What this work does
Adapts an image encoder to each test input with masked reconstruction while retaining a recognition head.
Connection to discovery
Illustrates both per-instance adaptation and the importance of selecting an auxiliary objective compatible with the downstream scientific task.
Evidence
Inspected training/probing alternatives, optimizer sensitivity and comparisons showing an auxiliary rotation objective can hurt rotation-invariant classes.
Scope and limitations
Image reconstruction and its linear-model analysis do not establish causal learning. Update costs and stopping rules remain part of the evaluation budget.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.