AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Mean Field Residual Networks: On the Edge of Chaos

Ge Yang, Samuel Schoenholz

NeurIPS 2017 · Architectures and efficiency › RNNs and normalization · approach: Theory of depth and width · 74 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
Increase Tanh network depth while keeping initialization variance fixedTanh residual networks · MNISTtest accuracyHurt
Scale Tanh initialization variance with depth, following the predicted gradient-explosion level curveTanh residual networks · MNISTtest accuracyImproved
Increase the skip-bias to residual-weight variance ratio while holding the leading gradient-growth term fixedTanh/FRN · MNISTtest accuracyMixed
Increase ReLU network depth toward the onset of numerical overflowReLU/FRN · MNISTtest accuracyImproved
Increase α and depth to maximize metric expressivity without numerical failureα-ReLU/FRN · MNISTtest accuracyImproved
Use gradient explosion as the initialization-selection criterion for ReLU or α-ReLUReLU and α-ReLU residual networks · MNISTtest accuracyNo clear effect
Use Xavier or He initialization without adapting variances to depthResidual networks · Initialization qualitytest performanceHurt
Use the mean-field theory to predict behavior from initialization hyperparametersResidual networks · MNISTrelative test performance predictionImproved

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