AI Paper Atlas

Research map / Optimization / Nonconvex and smooth optimization

Black-Box Generalization: Stability of Zeroth-Order Learning

Konstantinos Nikolakakis, Farzin Haddadpour, Dionysios Kalogerias, Amin Karbasi

NeurIPS 2022 · Optimization › Nonconvex and smooth optimization · approach: Coordinate and Adam convergence analysis · 3 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Replace gradient updates with ZoSS loss-difference updates and decrease the step size by Γ_K^dSGD-style learning under Lipschitz, smooth losses · Bounded nonconvex lossgeneralization error bound: \dfrac{1+(CB)^{-1}}{n}\big(2CL^{2}\big)^{\frac{1}{C\beta+1}}(eT)^{\frac{C\beta}{C\beta+1}} -> \dfrac{1+(CB)^{-1}}{n}\bigMixed
Use α_t≤C/t instead of a step size scaled by Γ_K^dZoSS · Bounded nonconvex loss, single-example updatesgeneralization error bound: The bound is \dfrac{3e\big(1+(CB)^{-1}\big)^{2}}{2n}\big(1+(2+c)CL^{2}\big)T and is independent of d and K.Improved
Increase the number of loss evaluations KZoSS · High-dimensional black-box learninggeneralization errorMixed
Use mini-batch updates rather than single-example updatesZoSS · Unbounded loss, any batch size m≤ngeneralization error boundNo clear effect
Take K→∞ and c→0ZoSS · Generalization boundsgeneralization error boundImproved
Use a smaller smoothing parameter μ satisfying the stated stability conditionZoSS · Lipschitz, smooth lossesgeneralization error boundImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy