AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

The Quantization Model of Neural Scaling

Eric J Michaud, Ziming Liu, Uzay Girit, Max Tegmark

NeurIPS 2023 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 8 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
Scale network parameters and training dataReLU MLP on multitask sparse parity · Multitask sparse paritymean test lossImproved
Learn more frequent subtasks before less frequent onesReLU MLP on multitask sparse parity · Multitask sparse paritysubtask lossImproved
Increase model size across the first six Pythia modelsPythia · The Pile test setmean cross-entropy loss: α_N = 0.083Improved
Increase model size for most language-model tokensPythia · The Pile test setper-token lossImproved
Use QDG gradient clustering to group prediction samplesPythia 19m · The Pile test setcluster coherenceImproved
Estimate quantum frequency from QDG cluster sizesPythia 19m · The Pile test set, QDG clustersrank-frequency slope: ≈ -1.24; -1.08Mixed
Use QDG to recover the known subtask frequency distributionReLU MLP on multitask sparse parity · Multitask sparse parity subtasksrank-frequency slope: -1.4; ≈ -1.1Hurt
Train on a larger dataset in the multi-epoch regimeReLU MLP on multitask sparse parity · Multitask sparse paritynumber of learned subtasks: n ∝ D^{1/(α+1)}Improved
Increase training stepsReLU MLP on multitask sparse parity · Multitask sparse parityloss scaling exponent: α_S ≈ 0.29; α_S ≈ 0.45Mixed
Use spectral clustering with different cluster counts and toy gradient dimensions/noise (ablation)QDG · QDG toy modelrank-frequency exponent estimate: error is on the order of 0.2Mixed
Use single-epoch step scaling instead of the theoretical convergence-time assumption (ablation)ReLU MLP on multitask sparse parity · Multitask sparse paritysubtask convergence time: S_k ∝ p_k^{-0.81}; S_k ∝ p_k^{-1}Mixed

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