This work analyzes neural scaling laws using power-law data spectra and derives analytical expressions for generalization error.
problem Understanding how neural network performance scales with key factors like data size and model complexity.
method Statistical mechanics techniques applied to one-pass stochastic gradient descent in a student-teacher framework.
result Derivation of analytical expressions for generalization error under power-law data spectra and identification of conditions for power-law scaling.
Large models follow power laws in performance with dataset size or parameters.
problem Understanding neural scaling laws in large language models.
method Joint generative data model and random feature model.
result Modeling and solving the dual limit reveals insights into scaling laws.
Non-Markovian point process shows power-law scaling, similar to nonlinear Markovian process.
problem Understanding the scaling behavior of non-Markovian point processes.
method Analyzed a confined fractional Brownian motion-driven point process and compared it to a nonlinear Markovian process.
result A nonlinear Markovian process can reproduce the power-law scaling behavior of a non-Markovian point process.
Unified theory for neural scaling laws in hierarchically compositional data.
problem Understanding neural scaling laws in hierarchically compositional data.
method Probabilistic context-free grammars and power-law distributed production rules.
result Unified learning curve behavior for classification and next-token prediction tasks.
Develops a simple model to understand learning curves for arbitrary power laws.
problem Lack of theoretical understanding of scaling laws in machine learning.
method Analyzes a toy model to determine if learning curves are universal or depend on data distribution.
result Determines that learning curves can exhibit n−β for arbitrary power β>0. New pruning method breaks power law scaling, potentially reducing error to exponential.
problem Improving neural network performance through scaling alone is costly.
method Developed a new data pruning metric to break power law scaling.
result Pruned datasets show better than power law scaling on various image datasets.
Study reveals neural scaling laws in random graphs and natural language models.
problem Understanding the origin of neural scaling laws in complex systems.
method Examined scaling laws in transformers trained on random walks and simplified natural language models.
result Neural scaling laws emerge in the absence of power law structure in data correlations.
This work explains scaling laws as redundancy laws in deep learning.
problem The mathematical origins of scaling laws in deep learning models remain unclear.
method Kernel regression and analysis of data covariance spectra.
result Scaling laws can be explained as redundancy laws, revealing the learning curve's slope depends on data redundancy.
LLMs learn peaked distributions slowly due to power-law losses.
problem Slow convergence of loss in training large language models.
method Systematic analysis of toy models and empirical evaluation of LLMs.
result Power-law time scaling with an exponent of 1/3 for learning peaked distributions.
This work extends the scaling law to multiple and kernel regression, challenging traditional machine learning principles.
problem Challenging traditional machine learning wisdom with scaling law in large practical models.
method Demonstrates the scaling law in multiple and kernel regression settings.
result The scaling law extends to multiple and kernel regression, providing deeper insights into LLMs.
Improved scaling laws in linear regression using data reuse.
problem Sustainability of neural scaling laws when running out of new data.
method Data reuse in multi-pass stochastic gradient descent (multi-pass SGD) for M-dimensional linear models trained on N data with sketched features. result Multi-pass SGD achieves a test error of Θ(M1−b+L(1−b)/a) with L>N, improving scaling laws in data-constrained regimes. Scaling laws found for reinforcement learning performance with model size and compute.
problem Challenges in extending generative modeling scaling laws to reinforcement learning.
method Introduced intrinsic performance as a monotonic function of mean episode return.
result Intrinsic performance scales as a power law in model size and environment interactions.
The study explains transformer scaling laws using statistical and approximation theories.
problem Understanding why transformer scaling laws exist for large models trained on low-dimensional data.
method Established statistical estimation and mathematical approximation theories for transformers on low-dimensional manifolds.
result Predicted a power law between generalization error and model and data sizes, with power depending on intrinsic data dimension.
A hierarchical model shows how scaling laws emerge from sequential feature recovery.
problem Emergence of scaling laws from feature learning in multi-layer networks.
method Layer-wise spectral algorithm adapted to compositional structure, sequential feature detection.
result Sequential detection of latent features, leading to explicit power-law decay of prediction error.
We extend our previous study of scaling range properties done for detrended fluctuation analysis (DFA) \cite{former_paper} to other techniques of fluctuation analysis (FA). The new technique called Modified Detrended Moving Average Analysis (MDMA) is introduced and its scaling range properties are examined and compared…
We study the growth dynamics of the size of manufacturing firms considering competition and normal distribution of competency. We start with the fact that all components of the system struggle with each other for growth as happened in real competitive bussiness world. The detailed quantitative agreement of the theory w…
Study shows how anisotropic data affects learning dynamics in phase retrieval.
problem Understanding learning dynamics in phase retrieval with anisotropic Gaussian inputs.
method Developed a tractable reduction to reveal a three-phase trajectory and derived scaling laws.
result Found that anisotropy leads to a three-phase trajectory: fast escape, slow convergence, and spectral-tail learning.
We introduce the stochastic multiplicative point process modelling trading activity of financial markets. Such a model system exhibits power-law spectral density S(f) ~ 1/f**beta, scaled as power of frequency for various values of beta between 0.5 and 2. Furthermore, we analyze the relation between the power-law autoco…
This study reveals statistical patterns in ERC20 token transactions on Ethereum blockchain.
problem Understanding transactional dynamics in decentralized systems.
method Examined over 44 million ERC20 token transfers, categorized by address type (EOA or SC), and analyzed using scaling laws.
result EOA-driven transactions exhibit consistent statistical behavior, while SC-driven activity displays sublinear scaling and bursty activity.
Scaling laws in linear regression explain model performance improvements with size and data.
problem Disagreement between empirical neural scaling laws and conventional wisdom on variance error.
method Infinite dimensional linear regression setup, one-pass SGD, Gaussian prior, power-law spectrum.
result Variance error is dominated by other errors, disappearing from the bound due to SGD's implicit regularization.
Power-law portfolios improve diversification by scaling weights sub-linearly.
problem Optimization methods struggle with unstable pair correlations and non-Gaussian risk measures.
method Construct portfolios with penalty proportional to arbitrary order moment of returns, leading to sub-linear weight scaling.
result Infinite order power-law portfolios are perfectly diversified, improving diversification over Kelly portfolios.
New mechanism found for power laws including Zipf's law.
problem Understanding the ubiquity of power law distributions.
method Introduced nonlinear self-excited Hawkes processes with fast-accelerating intensities.
result Wide class of nonlinear Hawkes processes have power law intensity PDFs.
We consider an ideal closed stock market, in which 100 traders have economic activities. The assets of the traders change through buying and selling stocks. We simulate the assets under conservation of both total currency and total number of stocks. If the traders are identical, then the assets are distributed as a sta…
Superposition accelerates training to a universal power-law exponent.
problem Training dynamics in neural networks.
method Teacher-student framework and analytic theory.
result Superposition leads to a universal power-law exponent of ~1, independent of data and channel statistics.
SignSGD outperforms SGD in linear regression with optimal scaling laws under PLRF model.
problem Improving linear regression performance with signSGD under power-law random features.
method Analysis of signSGD risk under PLRF model, comparison with SGD, identification of unique effects.
result SignSGD can have a steeper compute-optimal slope than SGD in noisy regimes, especially with WSD schedule.
We study the relaxation dynamics of a financial market just after the occurrence of a crash by investigating the number of times the absolute value of an index return is exceeding a given threshold value. We show that the empirical observation of a power law evolution of the number of events exceeding the selected thre…
We study distributions which have both fractal and non-fractal scale regions by introducing a typical scale into a scale invariant system. As one of models in which distributions follow power law in the large scale region and deviate further from the power law in the smaller scale region, we employ 2-dim quantum gravit…
We discover scaling laws for kernel regression loss under various learning rate schedules.
problem Understanding loss dynamics and learning rate schedules in kernel regression.
method Theoretical analysis of stochastic gradient descent on a power-law kernel regression model.
result Established a Functional Scaling Law (FSL) capturing the full loss trajectory under arbitrary learning rate schedules.
The paper predicts loss scaling across different datasets and compute scales.
problem Predicting loss scaling across different datasets and compute scales.
method Derive shifted power law relationships between train and test losses.
result Shifted power law relationships hold for various datasets and tasks, improving prediction accuracy.
Study how generalization scales with model size and data in quadratic neural networks.
problem Understanding how generalization scales with model size and data in quadratic neural networks.
method Analyzed ℓ2-regularized empirical test error minimization in a quadratic two-layer network with finite-sample setting and structured data. result Revealed a phase diagram with distinct scaling regimes as the number of parameters varies, showing data-dependent power laws controlled by spectral structure of the target.
We uncover scaling laws and statistical structure in complex datasets.
problem Understanding universal traits in complex datasets.
method Analogizing data to physical systems, using statistical physics and RMT.
result Real-world datasets and Gaussian data with long-range correlations share the same RMT universality class.
By using numerical simulation, we confirm that Takayasu--Sato--Takayasu (TST) model which leads Pareto's law satisfies the detailed balance under Gibrat's law. In the simulation, we take an exponential tent-shaped function as the growth rate distribution. We also numerically confirm the reflection law equivalent to the…
A power-law fit to the empirical inference-compute frontier in LOB prediction suggests a scaling-law-style frontier.
problem Limit order book prediction
method Using a suite of models ranging from small decision trees to neural LOB architectures
result A power-law fit to the low- and mid-compute non-MLPLOB frontier extrapolates across multiple orders of magnitude and attains R2=0.941 on the excluded high-compute MLPLOB target frontier. Analyzed a generalized voter model with power-law herding intensity, revealing anomalous diffusion and long-range memory.
problem Anomalous diffusion and long-range memory in a generalized voter model.
method Derived analytical expressions for moments and first passage time distribution, confirmed numerically.
result The model exhibits long-range memory indicators despite being a Markov model.
Detection of power-law behavior and studies of scaling exponents uncover the characteristics of complexity in many real world phenomena. The complexity of financial markets has always presented challenging issues and provided interesting findings, such as the inverse cubic law in the tails of stock price fluctuation di…
Our purpose is to relate the Fokker-Planck formalism proposed by [Friedrich et al., Phys. Rev. Lett. 84, 5224 (2000)] for the distribution of stock market returns to the empirically well-established power law distribution with an exponent in the range 3-5. We show how to use Friedrich et al.'s formalism to predict that…
Based on empirical financial time-series, we show that the "silence-breaking" probability follows a super-universal power law: the probability of observing a large movement is inversely proportional to the length of the on-going low-variability period. Such a scaling law has been previously predicted theoretically [R. …
We present a model for random simple graphs with a degree distribution that obeys a power law (i.e., is heavy-tailed). To attain this behavior, the edge probabilities in the graph are constructed from Bertoin-Fujita-Roynette-Yor (BFRY) random variables, which have been recently utilized in Bayesian statistics for the c…
We study the statistical properties of volatility---a measure of how much the market is likely to fluctuate. We estimate the volatility by the local average of the absolute price changes. We analyze (a) the S&P 500 stock index for the 13-year period Jan 1984 to Dec 1996 and (b) the market capitalizations of the largest…
New scaling laws explain deep learning performance growth.
problem Understanding neural network performance growth.
method Analyzed entire training dynamics of various architectures.
result Identified two dynamical scaling laws.
We analyze a database comprising quarterly sales of 55624 pharmaceutical products commercialized by 3939 pharmaceutical firms in the period 1992--2001. We study the probability density function (PDF) of growth in firms and product sales and find that the width of the PDF of growth decays with the sales as a power law w…
Deep learning (DL) creates impactful advances following a virtuous recipe: model architecture search, creating large training data sets, and scaling computation. It is widely believed that growing training sets and models should improve accuracy and result in better products. As DL application domains grow, we would li…
Study uncovers scaling laws and spectral properties of shallow neural networks.
problem Understanding scaling laws and spectral properties of shallow neural networks.
method Leveraging connections with matrix compressed sensing and LASSO, derived a phase diagram for excess risk.
result Uncovered crossovers between scaling regimes and plateau behaviors, validated empirical observations.
A simple model explains inference scaling in neural models.
problem Understanding how model performance improves with repeated inference attempts.
method A statistical ansatz based on memorization to study inference scaling laws.
result Inference loss exhibits a power law decay with increasing trials.
It is now well established empirically that financial price changes are distributed according to a power law, with cubic exponent. This is a fascinating regularity, as it holds for various classes of securities, on various markets, and on various time scales. The universality of this law suggests that there must be som…
Scaling laws for neural language models reveal optimal model size and compute allocation.
problem Understanding the optimal model size and compute allocation for neural language models.
method Empirical analysis of scaling laws for cross-entropy loss across model size, dataset size, and compute.
result Simple equations govern the dependence of overfitting and training speed on model/dataset size and model size, respectively.
We consider the scaling behaviors for fluctuations of the number of Korean firms bankrupted in the period from August 1 2002 to October 28 2003. We observe a power law for the distribution of the number of the bankrupted firms. The Pareto exponent is close to unity. We also consider the daily increments of the number o…
We present new findings in regard to data analysis in very high dimensional spaces. We use dimensionalities up to around one million. A particular benefit of Correspondence Analysis is its suitability for carrying out an orthonormal mapping, or scaling, of power law distributed data. Power law distributed data are foun…