Superposition Yields Robust Neural Scaling

• URL: http://arxiv.org/abs/2505.10465v2
• Date: Sun, 18 May 2025 15:54:08 GMT
• Title: Superposition Yields Robust Neural Scaling
• Authors: Yizhou Liu, Ziming Liu, Jeff Gore,
• Abstract summary: We study the origin of the neural scaling law -- the finding that loss decreases as a power law with model size.<n>We found that when superposition is weak, meaning only the most frequent features are represented without interference, the scaling of loss with model size depends on the underlying feature frequency.<n>We conclude that representation superposition is an important mechanism underlying the observed neural scaling laws.
• Score: 9.278468089636547
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The success of today's large language models (LLMs) depends on the observation that larger models perform better. However, the origin of this neural scaling law -- the finding that loss decreases as a power law with model size -- remains unclear. Starting from two empirical principles -- that LLMs represent more things than the model dimensions (widths) they have (i.e., representations are superposed), and that words or concepts in language occur with varying frequencies -- we constructed a toy model to study the loss scaling with model size. We found that when superposition is weak, meaning only the most frequent features are represented without interference, the scaling of loss with model size depends on the underlying feature frequency; if feature frequencies follow a power law, so does the loss. In contrast, under strong superposition, where all features are represented but overlap with each other, the loss becomes inversely proportional to the model dimension across a wide range of feature frequency distributions. This robust scaling behavior is explained geometrically: when many more vectors are packed into a lower dimensional space, the interference (squared overlaps) between vectors scales inversely with that dimension. We then analyzed four families of open-sourced LLMs and found that they exhibit strong superposition and quantitatively match the predictions of our toy model. The Chinchilla scaling law turned out to also agree with our results. We conclude that representation superposition is an important mechanism underlying the observed neural scaling laws. We anticipate that these insights will inspire new training strategies and model architectures to achieve better performance with less computation and fewer parameters.

Related papers

• Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks [59.552873049024775]
  We show that compute-optimally trained models exhibit a remarkably precise universality.<n>With learning rate decay, the collapse becomes so tight that differences in the normalized curves across models fall below the noise floor.<n>We explain these phenomena by connecting collapse to the power-law structure in typical neural scaling laws.
  arXiv  Detail & Related papers  (2025-07-02T20:03:34Z)
• Compute-Optimal LLMs Provably Generalize Better With Scale [102.29926217670926]
  We develop generalization bounds on the pretraining objective of large language models (LLMs) in the compute-optimal regime.<n>We introduce a novel, fully empirical Freedman-type martingale concentration that tightens existing bounds by accounting for the variance of the loss function.<n>We produce a scaling law for the generalization gap, with bounds that become predictably stronger with scale.
  arXiv  Detail & Related papers  (2025-04-21T16:26:56Z)
• Scaling Law Phenomena Across Regression Paradigms: Multiple and Kernel Approaches [28.569601803576845]
  We show that for models with Transformer architecture, the test loss exhibits a power-law relationship with model size, dataset size, and the amount of computation used in training.<n>Our analysis provides deeper insights into the scaling law, potentially enhancing our understanding of Large Language Models.
  arXiv  Detail & Related papers  (2025-03-03T08:57:49Z)
• Scaling Laws in Linear Regression: Compute, Parameters, and Data [86.48154162485712]
  We study the theory of scaling laws in an infinite dimensional linear regression setup.<n>We show that the reducible part of the test error is $Theta(-(a-1) + N-(a-1)/a)$.<n>Our theory is consistent with the empirical neural scaling laws and verified by numerical simulation.
  arXiv  Detail & Related papers  (2024-06-12T17:53:29Z)
• Scaling and renormalization in high-dimensional regression [72.59731158970894]
  We present a unifying perspective on recent results on ridge regression.<n>We use the basic tools of random matrix theory and free probability, aimed at readers with backgrounds in physics and deep learning.<n>Our results extend and provide a unifying perspective on earlier models of scaling laws.
  arXiv  Detail & Related papers  (2024-05-01T15:59:00Z)
• A Dynamical Model of Neural Scaling Laws [79.59705237659547]
  We analyze a random feature model trained with gradient descent as a solvable model of network training and generalization. Our theory shows how the gap between training and test loss can gradually build up over time due to repeated reuse of data.
  arXiv  Detail & Related papers  (2024-02-02T01:41:38Z)
• The Quantization Model of Neural Scaling [19.057931064238584]
  We propose the Quantization Model of neural scaling laws, explaining both the observed power law dropoff of loss with model and data size, and the sudden emergence of new capabilities with scale. We show that when quanta are learned in order of decreasing use frequency, then a power law in use explains observed power law scaling of loss.
  arXiv  Detail & Related papers  (2023-03-23T17:58:43Z)
• Training Trajectories of Language Models Across Scales [99.38721327771208]
  Scaling up language models has led to unprecedented performance gains. How do language models of different sizes learn during pre-training? Why do larger language models demonstrate more desirable behaviors?
  arXiv  Detail & Related papers  (2022-12-19T19:16:29Z)
• Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning [35.25854322376364]
  We show that different data modalities are embedded at arm's length in their shared representation in multi-modal models such as CLIP. In contrastive learning keeps the different modalities separate by a certain distance, which is influenced by the temperature parameter in the loss function. Our experiments further demonstrate that varying the modality gap distance has a significant impact in improving the model's downstream zero-shot classification performance and fairness.
  arXiv  Detail & Related papers  (2022-03-03T22:53:54Z)
• On the Role of Optimization in Double Descent: A Least Squares Study [30.44215064390409]
  We show an excess risk bound for the descent gradient solution of the least squares objective. We find that in case of noiseless regression, double descent is explained solely by optimization-related quantities. We empirically explore if our predictions hold for neural networks.
  arXiv  Detail & Related papers  (2021-07-27T09:13:11Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.