Last-layer retraining succeeds primarily because the held-out set has better group balance, supported by experiments that do not back the neural collapse hypothesis.
Note that Waterbirds is the only dataset that has a distribution shift and MultiNLI is the only dataset which is class-balanceda priori
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
On the Unreasonable Effectiveness of Last-layer Retraining
Last-layer retraining succeeds primarily because the held-out set has better group balance, supported by experiments that do not back the neural collapse hypothesis.