Encoder Adaptation
In the two-stage training paradigm of long-tailed recognition, the process of further updating feature encoder parameters after classifier calibration to improve representation quality, whose incremental value depends on how well the representation learning stage covers rare classes.
Timeline (last 90 days)
In the controlled study, the same low-support-set regime shifted from a state of "training only the classifier is sufficient" to a state of "encoder adaptation brings practically meaningful gains" simply as representation allocation became more imbalanced.
This state transition persisted across multiple classifier correction methods, representation objectives, normalization, and data augmentation controls, and still held in ImageNet-scale evaluation while maintaining accuracy on frequent classes.
When designing long-tail recognition systems, one cannot judge whether encoder adaptation is needed based solely on support set size; when representation allocation is severely imbalanced, encoder adaptation may bring significant gains even if the support set is small.
If a benchmark fails to separate support set from representation allocation, then conclusions about "whether encoder adaptation is effective" drawn from that benchmark may be misleading.
All Facts (4)
In the controlled study, the same low-support-set regime shifted from a state of "training only the classifier is sufficient" to a state of "encoder adaptation brings practically meaningful gains" simply as representation allocation became more imbalanced.
50%UnverifiedThis state transition persisted across multiple classifier correction methods, representation objectives, normalization, and data augmentation controls, and still held in ImageNet-scale evaluation while maintaining accuracy on frequent classes.
50%UnverifiedWhen designing long-tail recognition systems, one cannot judge whether encoder adaptation is needed based solely on support set size; when representation allocation is severely imbalanced, encoder adaptation may bring significant gains even if the support set is small.
50%UnverifiedIf a benchmark fails to separate support set from representation allocation, then conclusions about "whether encoder adaptation is effective" drawn from that benchmark may be misleading.
50%