Pith. sign in

REVIEW 7 cited by

Rotational Equilibrium: How Weight Decay Balances Learning Across Neural Networks

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2305.17212 v4 pith:2TJ74S6D submitted 2023-05-26 cs.LG

classification cs.LG
keywords weightdecaylearningacrossrotationadamdeepequilibrium
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

This study investigates how weight decay affects the update behavior of individual neurons in deep neural networks through a combination of applied analysis and experimentation. Weight decay can cause the expected magnitude and angular updates of a neuron's weight vector to converge to a steady state we call rotational equilibrium. These states can be highly homogeneous, effectively balancing the average rotation -- a proxy for the effective learning rate -- across different layers and neurons. Our work analyzes these dynamics across optimizers like Adam, Lion, and SGD with momentum, offering a new simple perspective on training that elucidates the efficacy of widely used but poorly understood methods in deep learning. We demonstrate how balanced rotation plays a key role in the effectiveness of normalization like Weight Standardization, as well as that of AdamW over Adam with L2-regularization. Finally, we show that explicitly controlling the rotation provides the benefits of weight decay while substantially reducing the need for learning rate warmup.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. SOAP, Muon, and Beyond: Pushing LLM Pretraining Scales

    cs.LG 2026-07 conditional novelty 6.0 of 10

    SOAP and Muon, stabilized by per-step QR eigenbasis updates and KL-Shampoo covariance accumulation, beat AdamW on large-batch LLM pretraining up to 100M-token batches.

  2. Improving Neural Network Training by Decoupling the Magnitude and Direction of Weight Vectors

    cs.LG 2026-06 unverdicted novelty 6.0 of 10

    Splitting weight matrices into a fixed-norm direction and learnable per-row/column magnitudes improves LLM training over AdamW/Muon, removes weight decay and warmup, and transfers the optimal LR across width.

  3. Training Dynamics of the Cooldown Stage in Warmup-Stable-Decay Learning Rate Scheduler

    cs.LG 2025-08 conditional novelty 6.0 of 10

    Cooldown shapes in WSD training trade off bias and variance, with sqrt and lowered-linear-0.7 shapes best for single runs, mirror-cosine best for averaging, and higher beta_2 during cooldown matching shape-selection gains.

  4. Lions and Muons: Optimization via Stochastic Frank-Wolfe under Heavy-Tailed Noise

    math.OC 2025-06 reject novelty 6.0 of 10

    Lion and Muon with weight decay are shown to be instances of one stochastic Frank-Wolfe algorithm, and clipped and variance-reduced variants get the first high-probability convergence rates for nonconvex Frank-Wolfe u...

  5. Scale Weight Decay and Train Better

    cs.LG 2026-07 conditional novelty 5.0 of 10

    Muon with weight decay scaled by η/η_max reaches the same MoE validation loss ~30% faster than constant-decay Muon while preserving asymptotic stationarity of the unregularized objective.

  6. Taming Transformer Without Using Learning Rate Warmup

    cs.LG 2025-05 conditional novelty 5.0 of 10

    A spectral-norm ratio rule on the optimizer shrinks per-layer learning rates when weight updates are too large, letting Transformers train without warmup.

  7. Energy Considerations for Large Pretrained Neural Networks

    cs.LG 2025-06 reject novelty 4.0 of 10

    Fine-tuning pretrained CNNs on ImageNet used far less electricity when weights were quantized to about 8 to 13 bits, while pruning and low-rank factorization did not reliably reduce energy.

Pith tools