arXiv subjects

David Donoho

Publications and source records attributed to David Donoho.

At least 19 recordsLinked to original sources

arXiv2026

Internal Data Repetition Destroys Language Models

Language models are running out of high-quality training data, and even aggressively deduplicated corpora retain some amount of repetition. Earlier controlled studies predated Chinchilla-style scaling laws and could only measure the cost of repetition indirectly. We revisit repetition in the Chinchilla era, using a fitted no-repetition scaling law to report Compute-Equivalent Gain and Compute-Equivalent Loss. We show that under this modernized paradigm, repetition damage is systematic in three ways. First, holding compute allocated to repeated data constant, eval loss peaks at an intermediate repeat count $\Rep$; repeating a moderately sized subset a moderate number of times damages performance more than repeating a large subset a few times or a small subset many times. Second, the location of this peak is well-fit by a power law in model size; this scaling law reveals that the most damaging number of repeated data grows more quickly than compute. Finally, when repeated documents consume 10\% of the FLOPs budget in a controlled exact-document repetition setting, the compute-equivalent loss can be large: on FineWeb-Edu-Dedup, the most damaging repeat count for a Qwen3-style 344M-parameter model at $\OT=1$ matches the loss of a no-repetition run using 67% of the FLOPs. We demonstrate that these phenomena are not language-model-specific, and can be analytically understood in a simple statistical model: a misspecified linear regression with verbatim duplicates reproduces the same qualitative loss peak, quantifying how such peaks can arise from a statistical tradeoff between memorization and generalization. Our findings add precision to the study of duplication in language models, allowing practitioners to quantify the wasted compute incurred by the presence and repeat structure of duplicates in pretraining corpora.

cs.LG↗

arXiv2026

Scale Dependent Data Duplication

Data duplication during pretraining can degrade generalization and lead to memorization, motivating aggressive deduplication pipelines. However, at web scale, it is unclear what constitutes a ``duplicate'': beyond surface-form matches, semantically equivalent documents (e.g. translations) may induce redundant training signals once models become sufficiently capable. Practically, this means that semantic duplicates operate increasingly like exact duplicates during training. We present evidence that duplication is scale-dependent in two ways. First, as model capability increases, cross-entropy loss gradients for semantically equivalent documents become more aligned. Smaller models, by contrast, produce gradients that reflect surface similarity (e.g., shared tokens) rather than semantic similarity. Second, we embedded all 192 million FineWeb-Edu-Dedup documents using EmbeddingGemma-300m. For moderate corpus sizes, the cosine similarity between nearest-neighbors follows an isotropic power law baseline. However, as corpus size grows to hundreds of billions of tokens, the nearest-neighbor similarities deviate sharply, indicating accelerated semantic collisions. Finally, controlled pretraining on data sampled with replacement from pools of finite unique documents shows that limited uniqueness yields mild degradation for small models, but rapidly increasing loss penalties for larger models, breaking naive scaling extrapolation. We derive explicit scaling laws that allow practitioners to estimate deviation from expected scaling due to limited semantic uniqueness of the pretraining corpus. Our results identify and resolve an unstudied source of scale-dependence, allowing for more accurate prediction at scale.

cs.LG↗

arXiv2025

Position: Machine Learning Conferences Should Establish a "Refutations and Critiques" Track

Science progresses by iteratively advancing and correcting humanity's understanding of the world. In machine learning (ML) research, rapid advancements have led to an explosion of publications, but have also led to misleading, incorrect, flawed or perhaps even fraudulent studies being accepted and sometimes highlighted at ML conferences due to the fallibility of peer review. While such mistakes are understandable, ML conferences do not offer robust processes to help the field systematically correct when such errors are made. This position paper argues that ML conferences should establish a dedicated "Refutations and Critiques" (R&C) Track. This R&C Track would provide a high-profile, reputable platform to support vital research that critically challenges prior research, thereby fostering a dynamic self-correcting research ecosystem. We discuss key considerations including track design, review principles, potential pitfalls, and provide an illustrative example submission concerning a recent ICLR 2025 Oral. We conclude that ML conferences should create official, reputable mechanisms to help ML research self-correct.

cs.LG↗

arXiv2025

Optimal Vector Compressed Sensing Using James Stein Shrinkage

The trend in modern science and technology is to take vector measurements rather than scalars, ruthlessly scaling to ever higher dimensional vectors. For about two decades now, traditional scalar Compressed Sensing has been synonymous with a Convex Optimization based procedure called Basis Pursuit. In the vector recovery case, the natural tendency is to return to a straightforward vector extension of Basis Pursuit, also based on Convex Optimization. However, Convex Optimization is provably suboptimal, particularly when $B$ is large. In this paper, we propose SteinSense, a lightweight iterative algorithm, which is provably optimal when $B$ is large. It does not have any tuning parameter, does not need any training data, requires zero knowledge of sparsity, is embarrassingly simple to implement, and all of this makes it easily scalable to high vector dimensions. We conduct a massive volume of both real and synthetic experiments that confirm the efficacy of SteinSense, and also provide theoretical justification based on ideas from Approximate Message Passing. Fascinatingly, we discover that SteinSense is quite robust, delivering the same quality of performance on real data, and even under substantial departures from conditions under which existing theory holds.

cs.LG↗

arXiv2024

Universality of the $\pi^2/6$ Pathway in Avoiding Model Collapse

Researchers in empirical machine learning recently spotlighted their fears of so-called Model Collapse. They imagined a discard workflow, where an initial generative model is trained with real data, after which the real data are discarded, and subsequently, the model generates synthetic data on which a new model is trained. They came to the conclusion that models degenerate as model-fitting generations proceed. However, other researchers considered an augment workflow, where the original real data continue to be used in each generation of training, augmented by synthetic data from models fit in all earlier generations. Empirical results on canonical datasets and learning procedures confirmed the occurrence of model collapse under the discard workflow and avoidance of model collapse under the augment workflow. Under the augment workflow, theoretical evidence also confirmed avoidance in particular instances; specifically, Gerstgrasser et al. (2024) found that for classical Linear Regression, test risk at any later generation is bounded by a moderate multiple, viz. pi-squared-over-6 of the test risk of training with the original real data alone. Some commentators questioned the generality of theoretical conclusions based on the generative model assumed in Gerstgrasser et al. (2024): could similar conclusions be reached for other task/model pairings? In this work, we demonstrate the universality of the pi-squared-over-6 augment risk bound across a large family of canonical statistical models, offering key insights into exactly why collapse happens under the discard workflow and is avoided under the augment workflow. In the process, we provide a framework that is able to accommodate a large variety of workflows (beyond discard and augment), thereby enabling an experimenter to judge the comparative merits of multiple different workflows by simulating a simple Gaussian process.

cs.LG↗

arXiv2023

Data Science at the Singularity

A purported `AI Singularity' has been in the public eye recently. Mass media and US national political attention focused on `AI Doom' narratives hawked by social media influencers. The European Commission is announcing initiatives to forestall `AI Extinction'. In my opinion, `AI Singularity' is the wrong narrative for what's happening now; recent happenings signal something else entirely. Something fundamental to computation-based research really changed in the last ten years. In certain fields, progress is dramatically more rapid than previously, as the fields undergo a transition to frictionless reproducibility (FR). This transition markedly changes the rate of spread of ideas and practices, affects mindsets, and erases memories of much that came before. The emergence of frictionless reproducibility follows from the maturation of 3 data science principles in the last decade. Those principles involve data sharing, code sharing, and competitive challenges, however implemented in the particularly strong form of frictionless open services. Empirical Machine Learning (EML) is todays leading adherent field, and its consequent rapid changes are responsible for the AI progress we see. Still, other fields can and do benefit when they adhere to the same principles. Many rapid changes from this maturation are misidentified. The advent of FR in EML generates a steady flow of innovations; this flow stimulates outsider intuitions that there's an emergent superpower somewhere in AI. This opens the way for PR to push worrying narratives: not only `AI Extinction', but also the supposed monopoly of big tech on AI research. The helpful narrative observes that the superpower of EML is adherence to frictionless reproducibility practices; these practices are responsible for the striking progress in AI that we see everywhere.

stat.OT↗

arXiv2023

Is your data alignable? Principled and interpretable alignability testing and integration of single-cell data

Single-cell data integration can provide a comprehensive molecular view of cells, and many algorithms have been developed to remove unwanted technical or biological variations and integrate heterogeneous single-cell datasets. Despite their wide usage, existing methods suffer from several fundamental limitations. In particular, we lack a rigorous statistical test for whether two high-dimensional single-cell datasets are alignable (and therefore should even be aligned). Moreover, popular methods can substantially distort the data during alignment, making the aligned data and downstream analysis difficult to interpret. To overcome these limitations, we present a spectral manifold alignment and inference (SMAI) framework, which enables principled and interpretable alignability testing and structure-preserving integration of single-cell data with the same type of features. SMAI provides a statistical test to robustly assess the alignability between datasets to avoid misleading inference, and is justified by high-dimensional statistical theory. On a diverse range of real and simulated benchmark datasets, it outperforms commonly used alignment methods. Moreover, we show that SMAI improves various downstream analyses such as identification of differentially expressed genes and imputation of single-cell spatial transcriptomics, providing further biological insights. SMAI's interpretability also enables quantification and a deeper understanding of the sources of technical confounders in single-cell data.

q-bio.QM↗

arXiv2021

Convex Sparse Blind Deconvolution

In the blind deconvolution problem, we observe the convolution of an unknown filter and unknown signal and attempt to reconstruct the filter and signal. The problem seems impossible in general, since there are seemingly many more unknowns than knowns . Nevertheless, this problem arises in many application fields; and empirically, some of these fields have had success using heuristic methods -- even economically very important ones, in wireless communications and oil exploration. Today's fashionable heuristic formulations pose non-convex optimization problems which are then attacked heuristically as well. The fact that blind deconvolution can be solved under some repeatable and naturally-occurring circumstances poses a theoretical puzzle. To bridge the gulf between reported successes and theory's limited understanding, we exhibit a convex optimization problem that -- assuming signal sparsity -- can convert a crude approximation to the true filter into a high-accuracy recovery of the true filter. Our proposed formulation is based on L1 minimization of inverse filter outputs. We give sharp guarantees on performance of the minimizer assuming sparsity of signal, showing that our proposal precisely recovers the true inverse filter, up to shift and rescaling. There is a sparsity/initial accuracy tradeoff: the less accurate the initial approximation, the greater we rely on sparsity to enable exact recovery. To our knowledge this is the first reported tradeoff of this kind. We consider it surprising that this tradeoff is independent of dimension. We also develop finite-$N$ guarantees, for highly accurate reconstruction under $N\geq O(k \log(k) )$ with high probability. We further show stable approximation when the true inverse filter is infinitely long and extend our guarantees to the case where the observations are contaminated by stochastic or adversarial noise.

cs.IT↗

arXiv2019

Degrees of Freedom Analysis of Unrolled Neural Networks

Unrolled neural networks emerged recently as an effective model for learning inverse maps appearing in image restoration tasks. However, their generalization risk (i.e., test mean-squared-error) and its link to network design and train sample size remains mysterious. Leveraging the Stein's Unbiased Risk Estimator (SURE), this paper analyzes the generalization risk with its bias and variance components for recurrent unrolled networks. We particularly investigate the degrees-of-freedom (DOF) component of SURE, trace of the end-to-end network Jacobian, to quantify the prediction variance. We prove that DOF is well-approximated by the weighted \textit{path sparsity} of the network under incoherence conditions on the trained weights. Empirically, we examine the SURE components as a function of train sample size for both recurrent and non-recurrent (with many more parameters) unrolled networks. Our key observations indicate that: 1) DOF increases with train sample size and converges to the generalization risk for both recurrent and non-recurrent schemes; 2) recurrent network converges significantly faster (with less train samples) compared with non-recurrent scheme, hence recurrence serves as a regularization for low sample size regimes.

cs.LG↗

arXiv2018

Neural Proximal Gradient Descent for Compressive Imaging

Recovering high-resolution images from limited sensory data typically leads to a serious ill-posed inverse problem, demanding inversion algorithms that effectively capture the prior information. Learning a good inverse mapping from training data faces severe challenges, including: (i) scarcity of training data; (ii) need for plausible reconstructions that are physically feasible; (iii) need for fast reconstruction, especially in real-time applications. We develop a successful system solving all these challenges, using as basic architecture the recurrent application of proximal gradient algorithm. We learn a proximal map that works well with real images based on residual networks. Contraction of the resulting map is analyzed, and incoherence conditions are investigated that drive the convergence of the iterates. Extensive experiments are carried out under different settings: (a) reconstructing abdominal MRI of pediatric patients from highly undersampled Fourier-space data and (b) superresolving natural face images. Our key findings include: 1. a recurrent ResNet with a single residual block unrolled from an iterative algorithm yields an effective proximal which accurately reveals MR image details. 2. Our architecture significantly outperforms conventional non-recurrent deep ResNets by 2dB SNR; it is also trained much more rapidly. 3. It outperforms state-of-the-art compressed-sensing Wavelet-based methods by 4dB SNR, with 100x speedups in reconstruction time.

cs.CV↗

arXiv2017

Recurrent Generative Adversarial Networks for Proximal Learning and Automated Compressive Image Recovery

Recovering images from undersampled linear measurements typically leads to an ill-posed linear inverse problem, that asks for proper statistical priors. Building effective priors is however challenged by the low train and test overhead dictated by real-time tasks; and the need for retrieving visually "plausible" and physically "feasible" images with minimal hallucination. To cope with these challenges, we design a cascaded network architecture that unrolls the proximal gradient iterations by permeating benefits from generative residual networks (ResNet) to modeling the proximal operator. A mixture of pixel-wise and perceptual costs is then deployed to train proximals. The overall architecture resembles back-and-forth projection onto the intersection of feasible and plausible images. Extensive computational experiments are examined for a global task of reconstructing MR images of pediatric patients, and a more local task of superresolving CelebA faces, that are insightful to design efficient architectures. Our observations indicate that for MRI reconstruction, a recurrent ResNet with a single residual block effectively learns the proximal. This simple architecture appears to significantly outperform the alternative deep ResNet architecture by 2dB SNR, and the conventional compressed-sensing MRI by 4dB SNR with 100x faster inference. For image superresolution, our preliminary results indicate that modeling the denoising proximal demands deep ResNets.

cs.AI↗

arXiv2014

Minimax risk of matrix denoising by singular value thresholding

An unknown $m$ by $n$ matrix $X_0$ is to be estimated from noisy measurements $Y=X_0+Z$, where the noise matrix $Z$ has i.i.d. Gaussian entries. A popular matrix denoising scheme solves the nuclear norm penalization problem $\operatorname {min}_X\|Y-X\|_F^2/2+λ\|X\|_*$, where $\|X\|_*$ denotes the nuclear norm (sum of singular values). This is the analog, for matrices, of $\ell_1$ penalization in the vector case. It has been empirically observed that if $X_0$ has low rank, it may be recovered quite accurately from the noisy measurement $Y$. In a proportional growth framework where the rank $r_n$, number of rows $m_n$ and number of columns $n$ all tend to $\infty$ proportionally to each other ($r_n/m_n\rightarrow ρ$, $m_n/n\rightarrow β$), we evaluate the asymptotic minimax MSE $\mathcal {M}(ρ,β)=\lim_{m_n,n\rightarrow \infty}\inf_λ\sup_{\operatorname {rank}(X)\leq r_n}\operatorname {MSE}(X_0,\hat{X}_λ)$. Our formulas involve incomplete moments of the quarter- and semi-circle laws ($β=1$, square case) and the Marčenko-Pastur law ($β<1$, nonsquare case). For finite $m$ and $n$, we show that MSE increases as the nonzero singular values of $X_0$ grow larger. As a result, the finite-$n$ worst-case MSE, a quantity which can be evaluated numerically, is achieved when the signal $X_0$ is "infinitely strong." The nuclear norm penalization problem is solved by applying soft thresholding to the singular values of $Y$. We also derive the minimax threshold, namely the value $λ^*(ρ)$, which is the optimal place to threshold the singular values. All these results are obtained for general (nonsquare, nonsymmetric) real matrices. Comparable results are obtained for square symmetric nonnegative-definite matrices.

math.ST↗

arXiv2014

Higher Criticism for Large-Scale Inference, Especially for Rare and Weak Effects

In modern high-throughput data analysis, researchers perform a large number of statistical tests, expecting to find perhaps a small fraction of significant effects against a predominantly null background. Higher Criticism (HC) was introduced to determine whether there are any nonzero effects; more recently, it was applied to feature selection, where it provides a method for selecting useful predictive features from a large body of potentially useful features, among which only a rare few will prove truly useful. In this article, we review the basics of HC in both the testing and feature selection settings. HC is a flexible idea, which adapts easily to new situations; we point out simple adaptions to clique detection and bivariate outlier detection. HC, although still early in its development, is seeing increasing interest from practitioners; we illustrate this with worked examples. HC is computationally effective, which gives it a nice leverage in the increasingly more relevant "Big Data" settings we see today. We also review the underlying theoretical "ideology" behind HC. The Rare/Weak (RW) model is a theoretical framework simultaneously controlling the size and prevalence of useful/significant items among the useless/null bulk. The RW model shows that HC has important advantages over better known procedures such as False Discovery Rate (FDR) control and Family-wise Error control (FwER), in particular, certain optimality properties. We discuss the rare/weak phase diagram, a way to visualize clearly the class of RW settings where the true signals are so rare or so weak that detection and feature selection are simply impossible, and a way to understand the known optimality properties of HC.

math.ST↗

arXiv2013

High Dimensional Robust M-Estimation: Asymptotic Variance via Approximate Message Passing

In a recent article (Proc. Natl. Acad. Sci., 110(36), 14557-14562), El Karoui et al. study the distribution of robust regression estimators in the regime in which the number of parameters p is of the same order as the number of samples n. Using numerical simulations and `highly plausible' heuristic arguments, they unveil a striking new phenomenon. Namely, the regression coefficients contain an extra Gaussian noise component that is not explained by classical concepts such as the Fisher information matrix. We show here that that this phenomenon can be characterized rigorously techniques that were developed by the authors to analyze the Lasso estimator under high-dimensional asymptotics. We introduce an approximate message passing (AMP) algorithm to compute M-estimators and deploy state evolution to evaluate the operating characteristics of AMP and so also M-estimates. Our analysis clarifies that the `extra Gaussian noise' encountered in this problem is fundamentally similar to phenomena already studied for regularized least squares in the setting n<p.

math.ST↗

arXiv2013

Accurate Prediction of Phase Transitions in Compressed Sensing via a Connection to Minimax Denoising

Compressed sensing posits that, within limits, one can undersample a sparse signal and yet reconstruct it accurately. Knowing the precise limits to such undersampling is important both for theory and practice. We present a formula that characterizes the allowed undersampling of generalized sparse objects. The formula applies to Approximate Message Passing (AMP) algorithms for compressed sensing, which are here generalized to employ denoising operators besides the traditional scalar soft thresholding denoiser. This paper gives several examples including scalar denoisers not derived from convex penalization -- the firm shrinkage nonlinearity and the minimax nonlinearity -- and also nonscalar denoisers -- block thresholding, monotone regression, and total variation minimization. Let the variables eps = k/N and delta = n/N denote the generalized sparsity and undersampling fractions for sampling the k-generalized-sparse N-vector x_0 according to y=Ax_0. Here A is an n\times N measurement matrix whose entries are iid standard Gaussian. The formula states that the phase transition curve delta = delta(eps) separating successful from unsuccessful reconstruction of x_0 by AMP is given by: delta = M(eps| Denoiser), where M(eps| Denoiser) denotes the per-coordinate minimax mean squared error (MSE) of the specified, optimally-tuned denoiser in the directly observed problem y = x + z. In short, the phase transition of a noiseless undersampling problem is identical to the minimax MSE in a denoising problem.

cs.IT↗

arXiv2011

Compressed Sensing over $\ell_p$-balls: Minimax Mean Square Error

We consider the compressed sensing problem, where the object $x_0 \in \bR^N$ is to be recovered from incomplete measurements $y = Ax_0 + z$; here the sensing matrix $A$ is an $n \times N$ random matrix with iid Gaussian entries and $n < N$. A popular method of sparsity-promoting reconstruction is $\ell^1$-penalized least-squares reconstruction (aka LASSO, Basis Pursuit). It is currently popular to consider the strict sparsity model, where the object $x_0$ is nonzero in only a small fraction of entries. In this paper, we instead consider the much more broadly applicable $\ell_p$-sparsity model, where $x_0$ is sparse in the sense of having $\ell_p$ norm bounded by $ξ\cdot N^{1/p}$ for some fixed $0 < p \leq 1$ and $ξ> 0$. We study an asymptotic regime in which $n$ and $N$ both tend to infinity with limiting ratio $n/N = δ\in (0,1)$, both in the noisy ($z \neq 0$) and noiseless ($z=0$) cases. Under weak assumptions on $x_0$, we are able to precisely evaluate the worst-case asymptotic minimax mean-squared reconstruction error (AMSE) for $\ell^1$ penalized least-squares: min over penalization parameters, max over $\ell_p$-sparse objects $x_0$. We exhibit the asymptotically least-favorable object (hardest sparse signal to recover) and the maximin penalization. Our explicit formulas unexpectedly involve quantities appearing classically in statistical decision theory. Occurring in the present setting, they reflect a deeper connection between penalized $\ell^1$ minimization and scalar soft thresholding. This connection, which follows from earlier work of the authors and collaborators on the AMP iterative thresholding algorithm, is carefully explained. Our approach also gives precise results under weak-$\ell_p$ ball coefficient constraints, as we show here.

cs.IT↗

arXiv2008

Feature selection by Higher Criticism thresholding: optimal phase diagram

We consider two-class linear classification in a high-dimensional, low-sample size setting. Only a small fraction of the features are useful, the useful features are unknown to us, and each useful feature contributes weakly to the classification decision -- this setting was called the rare/weak model (RW Model). We select features by thresholding feature $z$-scores. The threshold is set by {\it higher criticism} (HC). Let $\pee_i$ denote the $P$-value associated to the $i$-th $z$-score and $\pee_{(i)}$ denote the $i$-th order statistic of the collection of $P$-values. The HC threshold (HCT) is the order statistic of the $z$-score corresponding to index $i$ maximizing $(i/n - \pee_{(i)})/\sqrt{\pee_{(i)}(1-\pee_{(i)})}$. The ideal threshold optimizes the classification error. In \cite{PNAS} we showed that HCT was numerically close to the ideal threshold. We formalize an asymptotic framework for studying the RW model, considering a sequence of problems with increasingly many features and relatively fewer observations. We show that along this sequence, the limiting performance of ideal HCT is essentially just as good as the limiting performance of ideal thresholding. Our results describe two-dimensional {\it phase space}, a two-dimensional diagram with coordinates quantifying "rare" and "weak" in the RW model. Phase space can be partitioned into two regions -- one where ideal threshold classification is successful, and one where the features are so weak and so rare that it must fail. Surprisingly, the regions where ideal HCT succeeds and fails make the exact same partition of the phase diagram. Other threshold methods, such as FDR threshold selection, are successful in a substantially smaller region of the phase space than either HCT or Ideal thresholding.

math.ST↗

arXiv2007

Asymptotic minimaxity of False Discovery Rate thresholding for sparse exponential data

We apply FDR thresholding to a non-Gaussian vector whose coordinates X_i, i=1,..., n, are independent exponential with individual means $μ_i$. The vector $μ=(μ_i)$ is thought to be sparse, with most coordinates 1 but a small fraction significantly larger than 1; roughly, most coordinates are simply `noise,' but a small fraction contain `signal.' We measure risk by per-coordinate mean-squared error in recovering $\log(μ_i)$, and study minimax estimation over parameter spaces defined by constraints on the per-coordinate p-norm of $\log(μ_i)$: $\frac{1}{n}\sum_{i=1}^n\log^p(μ_i)\leq η^p$. We show for large n and small $η$ that FDR thresholding can be nearly Minimax. The FDR control parameter 0 1/2 prevents near minimaxity. These conclusions mirror those found in the Gaussian case in Abramovich et al. [Ann. Statist. 34 (2006) 584--653]. The techniques developed here seem applicable to a wide range of other distributional assumptions, other loss measures and non-i.i.d. dependency structures.

math.ST↗