Importance Weighting
A statistical method that corrects distribution differences by weighting samples according to density ratios. Under covariate shift it is used to restore the theoretical guarantees of conformal prediction.
Timeline (last 90 days)
The paper explores two implementation paths for data-free training: the importance-weighted version provides a consistent objective throughout training; the unweighted variant defines a self-consistent fixed-point training scheme.
The cost of the theoretical consistency of the importance-weighted version is that the weights may have large variance, and when the model-generated distribution differs greatly from the target distribution, the weights become extreme, leading to unstable training.
All Facts (2)
The paper explores two implementation paths for data-free training: the importance-weighted version provides a consistent objective throughout training; the unweighted variant defines a self-consistent fixed-point training scheme.
50%UnverifiedThe cost of the theoretical consistency of the importance-weighted version is that the weights may have large variance, and when the model-generated distribution differs greatly from the target distribution, the weights become extreme, leading to unstable training.
50%