AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

Implicit Bias of Per-sample Adam on Separable Data: Departure from the Full-batch Regime

Beomhan Baek, Minhak Song, Chulhee Yun

ICLR 2026 · Optimization › Implicit bias and generalization · approach: Implicit bias in linear networks

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
use batch size 1 (incremental) instead of full batch for Adam (ablation)Adam · Gaussian data (N=10, d=50)direction of convergence (cosine similarity to max-margin solutions)Mixed
use incremental Adam on scaled Rademacher data instead of full-batch Adam (ablation)Adam · Scaled Rademacher data (4 points, d=4)direction of convergenceMixed
increase batch size toward full batch for incremental Adam (ablation)Adam · Gaussian data (N=10, d=50)cosine similarity to l_inf-max-marginImproved
vary beta1 in {0.1, 0.5, 0.9} with beta2=0.95 (ablation)incremental Adam, batch size 1 · Gaussian data (N=10, d=50)limit directionNo clear effect
lower beta2 to 0.1 or 0.5 instead of 0.95 (ablation)incremental Adam, batch size 1 · Gaussian data (N=10, d=50)agreement with fixed-point solutionHurt
use Signum with beta=0.99 and smaller batch sizes (5, 2, 1) (ablation)Signum · Gaussian data (N=10, d=50)convergence to l_inf-max-marginNo clear effect

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