Leverage Matrix
A geometric operator derived from the concept of leverage values in linear regression. It measures the influence of training samples on model predictions and, in machine unlearning, characterizes the geometric entanglement between forgotten data and retained data.
Timeline (last 90 days)
Two leverage matrices emerge in the framework: forget self-leverage and retain cross-leverage, capturing the entanglement of forget and retain data, respectively, under the geometry of the retrained model.
Existing gradient-based unlearning methods account only partially for the geometric relationships characterized by the leverage matrices, causing them to deviate significantly from the retrained model in output-space behavior.
The leverage matrix originates from the concept of leverage in linear regression, used to measure the degree of influence of each training sample on the model's predictions.
The precomputable leverage diagnostics compute the leverage matrices and related metrics using only statistics of the forget set and the retain set, without depending on the specific unlearning optimization process.
All Facts (4)
Two leverage matrices emerge in the framework: forget self-leverage and retain cross-leverage, capturing the entanglement of forget and retain data, respectively, under the geometry of the retrained model.
50%UnverifiedExisting gradient-based unlearning methods account only partially for the geometric relationships characterized by the leverage matrices, causing them to deviate significantly from the retrained model in output-space behavior.
50%UnverifiedThe leverage matrix originates from the concept of leverage in linear regression, used to measure the degree of influence of each training sample on the model's predictions.
50%UnverifiedThe precomputable leverage diagnostics compute the leverage matrices and related metrics using only statistics of the forget set and the retain set, without depending on the specific unlearning optimization process.
50%