AI Paper Atlas

Research map / Optimization / Optimizers

Escaping the Gravitational Pull of Softmax

Jincheng Mei, Chenjun Xiao, Bo Dai, Lihong Li, Csaba Szepesvari, Dale Schuurmans

NeurIPS 2020 · Optimization › Optimizers · approach: Randomized gradient estimation · 12 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 softmax with escort transform (EPG) for policy gradient on banditspolicy gradient · one-state MDPs, K=10/50/100convergence to optimal policy / suboptimalityImproved
replace softmax with escort transform (p=2) in policy gradientpolicy gradient with ReLU network · Four-room with sub-goalsvalue / optimal goal reachingImproved
replace softmax with escort transform (p=4) for misclassification-probability minimizationone-hidden-layer MLP · MNIST (policy gradient objective)training objective and test errorImproved
replace softmax cross entropy with escort cross entropy (p=2)one-hidden-layer MLP · MNISTtraining objective speed and test errorImproved
use small escort exponent p instead of large p in EPG (ablation)EPG · single-state MDP K=3, r=(0.2,0.9,1.0)resistance to suboptimal corner attractionImproved
use p=1 escort instead of p>=2 (ablation)EPG · single-state MDP K=3smoothness of optimization curvesHurt
mirror descent with KL regularization compared with EPG p=2EPG · single-state MDP K=3suboptimality trajectoryNo 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