AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Faster Activation Functions at the Edge for Post-Training Speedups

Anton Lydike, Jun Bi, Jackson Woodruff

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Efficient network compression and training

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 exact SiLU with FFCC FastSiLU post-trainingtrained models · 8 models (DenseNet, MobileNetV2, ResNet-18, ShuffleNetV2, SqueezeNet, MLPerf Tiny)accuracyNo clear effect
replace exact SiLU with FastSiLU post-trainingMLPerf Tiny Anomaly Detection autoencoder · MLPerf Tiny Anomaly Detectiontask metric: 91.94 -> nanHurt
train with Hardswish instead of SiLUMobileNetV2, DenseNet-121 · image classificationaccuracy: 74.0 -> 72.1 (DenseNet-121)Hurt
replace exact GELU with FastGELU in pretrained ConvNeXtConvNeXt-Tiny · ImageNetTop-1: 82.1 -> 82.0No clear effect
replace exact GELU with FastGELU in pretrained ConvNeXtConvNeXt-Femto · ImageNetTop-1: 77.4 -> 77.2Hurt
tune constants holistically over the whole expression instead of exponentiation alone (ablation)FFCC FastGELU · ConvNeXt ImageNetTop-1 accuracy: +0.1 (Tiny) to +0.3 pp (Femto)Improved
approximate SiLU with FFCC instead of exact SiLUArm Cortex-M7 · activation microbenchmarkcycles per activation: 138.2 -> 21.1Improved
approximate activations with FFCCSnapdragon and Intel chips · activation microbenchmarkwall time: 15-22x speedupImproved
replace exact GELU with FFCC approximation in ConvNeXt-FemtoConvNeXt-Femto · end-to-end per-stage latencyspeedup: 1.2x to 1.37xImproved
replace exact GELU with FFCC approximation in ConvNeXt-TinyConvNeXt-Tiny · end-to-end per-stage latencyspeedup: 1.19x to 1.37x in Stage 1Mixed
subtract small offset epsilon from exponent bias (ablation)bit-cast exp2 approximation · approximation errorerrorImproved

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