Adapting Without Gradients: Affine Statistics Transport and What Its Certificate Can Tell You
CASTER adapts frozen models at test time without backprop, then uses a certificate to decide when that transport is unsafe.
The method stores source class statistics, estimates a shared affine shift from target-batch moments, and transports the source distributions before classification. In the reported tests, it beats k-NN on the same frozen features in 27 of 28 backbone-dataset settings while using a median 18x less state. The paper also shows the risk: on ImageNet-C, unconditional transport drops 21.2 top-1 points with small batches across 1000 classes. Its residual-to-margin certificate flags all cases losing more than 10 points above 3.9, and gating turns an average -3.35-point transport effect into a +1.69-point gain. HF Daily Papers' note
The method stores source class statistics, estimates a shared affine shift from target-batch moments, and transports the source distributions before classification. In the reported tests, it beats k-NN on the same frozen features in 27 of 28 backbone-dataset settings while using a median 18x less state. The paper also shows the risk: on ImageNet-C, unconditional transport drops 21.2 top-1 points with small batches across 1000 classes. Its residual-to-margin certificate flags all cases losing more than 10 points above 3.9, and gating turns an average -3.35-point transport effect into a +1.69-point gain. HF Daily Papers' note
score 4