Ramin Akbari
Blog
Categories
All
(1)
Kernel Methods
(1)
Representation Learning
(1)
Statistical Independence
(1)
Blog
Deep Dive into Obliviator: Nonlinear Guardedness in Concept Erasure
NeurIPS 2025
Representation Learning
Statistical Independence
Kernel Methods
This post discusses
Obliviator
, a method for capturing and removing specific concept dependencies within the learned representations of pre-trained language models. Concept erasure presents two central challenges: first, the need to target all types of dependencies for complete erasure; and second, the requirement to leave unrelated information untouched. We elaborate on how Obliviator navigates this balance to achieve a strong benchmark in utility-erasure trade-off.
arXiv
Code
Slides
Poster
No matching items
