[KongchangAI]
Conceptmatched perturbation

Matched Perturbation

An experimental design method in causal inference that constructs control perturbations comparable in magnitude to the target variable but different in structure, in order to test whether an effect is specific to the particular structure of the target variable.

Timeline (last 90 days)

Nov 18

In the matched perturbation experiments, the errors of the shuffled and Gaussian versions averaged close to zero, while the real error behaved nearly linearly when mirrored/scaled, demonstrating that behavioral changes are related to error structure rather than bit count.

Unverified50%
Nov 18

Matched perturbation is an experimental design idea in causal inference that tests effect specificity by constructing control perturbations of comparable magnitude but different structure.

Unverified50%

All Facts (2)

Source Articles