AI Paper Atlas

Research map / Learning theory and algorithms / Kernels and regression theory

Scaling Laws from the Data Manifold Dimension

Utkarsh Sharma, Jared Kaplan

JMLR 2022 · Learning theory and algorithms › Kernels and regression theory · approach: High-dimensional asymptotic generalization

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
measure intrinsic dimension on larger student networks (ablation)teacher/student CE · Teacher/studentintrinsic dimension: about 10%No clear effect
use CE loss instead of MSE for the ID vs 1/alpha relation (ablation)teacher/student · Teacher/studentslope of ID vs 1/alpha: 4.08 vs 3.9No clear effect
build the teacher as a product of independent sub-manifolds (T3+6) (ablation)teacher/student · Product manifold4/alpha: 4/alpha 5.76 with ID 7.86Mixed
repeat a 3-feature teacher three times (T3+3+3) (ablation)teacher/student · Product manifold4/alpha: 2.66 with ID 7.29 versus 3.3 for T3 aloneNo clear effect
replace the loss with generalized |y-y*|^p (ablation)teacher/student d about 7 · Teacher/studentalpha: alpha = 2p/dImproved
measure ID from one 1024-token passage rather than many sequences (ablation)GPT-2 117M · Language modelingintrinsic dimension: 90 -> 7Hurt
evaluate training loss instead of test loss for CIFAR10 (ablation)CNN · CIFAR104/alpha: 17.83 -> 14.19Mixed
use error rate instead of loss on CIFAR10 (ablation)CNN · CIFAR104/alpha: 17.83 -> 16.23No 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