Unverified50% confidenceFactExact time
This study reveals a fundamental disconnect between loss and behavior in gradient-based machine unlearning and provides a unified output-space analysis framework.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGradient-based machine unlearning methods typically achieve unlearning by optimizing a forget loss and a retain loss, but similar loss values do not necessarily imply similar output-space behavior.78% similarUnverifiedThe core question posed by this study is: when a gradient-based unlearner differs from a model retrained only on the retained data, what structure does this output-space gap have.78% similarUnverifiedThe two core metrics for measuring machine unlearning quality are Forget Rate and Retain Rate; these two metrics are in natural tension, as excessive forgetting can lead to concept collapse.75% similarUnverifiedThese two leverage matrices are shown both theoretically and experimentally to drive the influence of the unlearning residual on forget outputs and retain outputs.75% similarUnverifiedThe paper proves fully matching approximate upper and lower bounds for exponential-mode and finite-state-mode cases of algebraically decaying forecast memory.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1113851API
curl https://kongchang.com/api/v1/knowledge/claims/1113851MCP
get_claim(id=1113851)