AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Numerical influence of ReLU’(0) on backpropagation

David Bertoin, Jerome Bolte, Sébastien Gerchinovitz, Edouard Pauwels

NeurIPS 2021 · Architectures and efficiency › RNNs and normalization · approach: Theory of depth and width · 1 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
switch from 32 to 64 bit precisionfully connected net MNIST · MNIST weight samplingproportion of weights in bifurcation zone: 40% -> 0%Improved
switch from 32 to 16 bit precisionfully connected net MNIST · MNIST weight samplingproportion of weights in bifurcation zone: 40% -> 100%Hurt
set ReLU'(0)=1 instead of 0ResNet50 SGD · ImageNetaccuracy: 75% to 63% or 55%Hurt
add batch normalizationVGG11 · CIFAR10test accuracy sensitivity to ReLU'(0)Improved
use Adam instead of SGDVGG11 · CIFAR10, SVHNsensitivity to ReLU'(0)Improved
use 64 bit precision with ReLU'(0)=1VGG11 · CIFAR10test performanceImproved
train with ReLU'(0)=1 in 16 bitsVGG11 · CIFAR10trainabilityHurt
vary ReLU'(0) magnitude beyond 0.1 without batch normVGG11 · CIFAR10test accuracyHurt
add dropout with varying probabilityfully connected net · MNISTinteraction with ReLU'(0)No clear effect
increase sample sizefully connected net · MNIST weight samplingprobability of bifurcation zoneHurt
increase network width and depthfully connected net · MNIST weight samplingprobability of bifurcation zoneHurt
vary mini-batch sizefully connected net · MNIST weight samplingprobability of bifurcation zoneNo 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