Architectures with stronger inductive biases attain higher in-distribution accuracy but degrade faster under temporal distribution shift, while frozen pretrained encoders trade accuracy for stability.
Deep residual learning for image recognition
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Drift Happens: An Empirical Study of Neural Architecture Robustness to Temporal Distribution Shift
Architectures with stronger inductive biases attain higher in-distribution accuracy but degrade faster under temporal distribution shift, while frozen pretrained encoders trade accuracy for stability.