Proves hardness of hyper-gradient stationarity for zero-respecting algorithms in nonconvex-convex bilevel optimization and establishes improved optimal complexity bounds under PL condition for nonconvex-nonconvex cases.
By Lemma G.1, it suffices to set the parameters in the inner loop as Equation ( 8)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
math.OC 1years
2023 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
On Finding Small Hyper-Gradients in Bilevel Optimization: Hardness Results and Improved Analysis
Proves hardness of hyper-gradient stationarity for zero-respecting algorithms in nonconvex-convex bilevel optimization and establishes improved optimal complexity bounds under PL condition for nonconvex-nonconvex cases.