AI Paper Atlas

Research map / Architectures and efficiency / Quantization

You Already Have It: A Generator-Free Low-Precision DNN Training Framework Using Stochastic Rounding

Geng Yuan, Sung-En Chang, Qing Jin, Alec Lu, Yanyu Li, Yushu Wu, Zhenglun Kong, Yanyue Xie, Peiyan Dong, Minghai Qin, Xiaolong Ma, Xulong Tang and 2 more

ECCV 2022 · Architectures and efficiency › Quantization · approach: Low-precision training formats · 9 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
use gradients as the random-number source instead of errors (ablation)generator-free stochastic rounding · CIFAR-100 (ResNet-32) 6-bitaccuracy (%): 8.34 -> 74.06Improved
use gradients instead of activations as source (ablation)generator-free stochastic rounding · CIFAR-100 (ResNet-32) 6-bitaccuracy (%): 70.19 -> 74.06Improved
use gradients instead of weights as source (ablation)generator-free stochastic rounding · CIFAR-100 (ResNet-32) 6-bitaccuracy (%): 73.32 -> 74.06Improved
use gradients instead of errors as source (ablation)generator-free stochastic rounding · CIFAR-10 (ResNet-20) 6-bitaccuracy (%): 71.11 -> 91.88Improved
use gradients instead of activations as source (ablation)generator-free stochastic rounding · CIFAR-10 (ResNet-20) 6-bitaccuracy (%): 90.36 -> 91.88Improved
extract from later layers of errors (layer-31 vs layer-2) (ablation)generator-free stochastic rounding · CIFAR-100 (ResNet-32) 6-bitaccuracy (%): 8.34 -> 54.22Improved
synthesize random numbers from LSBs of multiple source data instead of direct bit extraction (ablation)generator-free stochastic rounding · ResNet-20 CIFAR-10random number qualityImproved
add number-mapping strategy to biased sources (ablation)generator-free stochastic rounding · biased sources (activations, errors)accuracyImproved
source with higher information entropy (ablation)generator-free stochastic rounding · CIFAR-100 (ResNet-32) 6-bitinformation entropy (bits): 2.8017 -> 5.8774Improved
replace LFSR-based RNGs with generator-free extractionlow-precision training · ImageNet ResNet-18accuracy (%): 70.72 -> 71.10Improved
replace LFSR RNGs with extractionlow-precision training · CIFAR-10/100 6-bit and 8-bitaccuracy (%): 91.83 -> 91.88No clear effect
replace stochastic rounding with nearest roundinglow-precision training · ImageNet ResNet-18accuracy (%): 71.10 -> 68.13Hurt

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