AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Revisiting Neural Scaling Laws in Language and Vision

Ibrahim Alabdulmohsin, Behnam Neyshabur, Xiaohua Zhai

NeurIPS 2022 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 22 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 M4 (with alpha) instead of M2 power law with offsetscaling law estimator · NMT 6 Enc 6 Decextrapolation RMSE: 3.9e-2 -> 1.0e-2Improved
use M4 instead of M3scaling law estimator · NMT 6 Enc 28 Decextrapolation RMSE: 1.6e-2 -> 3.0e-2Hurt
use M4 instead of M2scaling law estimator · NMT Transformer Enc/LSTM Decextrapolation RMSE: 1.3e-2 -> 1.2e-2No clear effect
use M4 instead of M2/M3scaling law estimator · language modelingextrapolation RMSE: e.g. 1.68e7: 6.0e-4 -> 3.1e-4; 1.34e8: 6.6e-4 (M3) vs 1.9e-3Mixed
use M4 instead of M1scaling law estimator · image classification (72 tasks)extrapolation RMSEImproved
use M4 instead of M2scaling law estimator · BIG-Bench date_understandingextrapolation RMSE: 3.2e-2 -> 1.5e-2 vs M3 4.7e-3Hurt
use M4 instead of M2scaling law estimator · BIG-Bench mult_data_wrangling 2-shotextrapolation RMSE: 1.6e-2 -> 6.2e-3Improved
add learnable alpha parameter to M4 (ablation)M4 · ImageNet 10-shot ViT/B/16extrapolation RMSE: 4.3e-2 -> 2.2e-2Improved
add learnable alpha parameter to M4 (ablation)M4 · ImageNet 10-shot BiT/101/3extrapolation RMSE: 5.1e-2 -> 1.7e-2Improved
select scaling exponent by best fit vs best extrapolationM2 on ViT/B/16 JFT-300M · ImageNet 10-shotextrapolation RMSE: c=-0.24 best fit vs c=-0.40 best extrapolationMixed

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