Interpretable AI with Local Distillation
A black-box teacher is used to produce sparse local linear models that keep most of its predictive power while exposing which features matter at each query point.
The paper proposes “local distillation,” where the teacher defines nearby cases by similar predicted outcomes and anchors each local fit with its own prediction.
The authors add Gaussian randomization and refits to measure which selected features are stable.
They report near-teacher accuracy across 17 benchmark datasets, with a sparse linear model produced for each test point.
In a cancer gene-expression example, the method surfaces patient subgroups using different genes, which a global linear model misses.
ArXiv · AI/CL/LG's note
The paper proposes “local distillation,” where the teacher defines nearby cases by similar predicted outcomes and anchors each local fit with its own prediction.
The authors add Gaussian randomization and refits to measure which selected features are stable.
They report near-teacher accuracy across 17 benchmark datasets, with a sparse linear model produced for each test point.
In a cancer gene-expression example, the method surfaces patient subgroups using different genes, which a global linear model misses.
ArXiv · AI/CL/LG's note
score 4