1 paper · 1 filter
Zice Wang, Zhenyu Zhang
Frozen pretrained features are often treated as a safe interface for downstream learning: only a small linear readout is trained, while the backbone is fixed. We show that this rea…