📑

학술

arXiv 등 학술 논문. CC-BY 라이선스로 자유 재사용 가능 — 출처표시 시 상업 사용 OK.

총 25,192건

Statistically Undetectable Backdoors in Deep Neural Networks

arXiv:2607.09532v1 Announce Type: cross Abstract: We show how an adversarial model trainer can plant backdoors in a large class of deep, feedforward neural networks. These backdoors are statistically undetectable in the white-box setting, meaning that the backdoored and honestly trained models are close in total variation distance, even given the full descriptions of the models (e.g., all of the weights). The backdoor provides access to invariance-based adversarial examples for every input, mapping distant inputs to unusually close outputs. However, without the backdoor, it is provably impossible (under standard cryptographic assumptions) to generate any such adversarial examples in polynomial time. Our theoretical and preliminary empirical findings demonstrate a fundamental power asymmetry between model trainers and model users.

arXiv Stat

Normalisation-Based Likelihood Ratio Estimation for Forensic Authorship Verification

arXiv:2607.09501v1 Announce Type: cross Abstract: Authorship verification (AV) is the task of determining whether two texts were written by the same author. In a forensic context, the strength of AV evidence can be quantified using likelihood ratios. Most AV methods are score-based and deriving well-calibrated likelihood ratios from these scores requires a separate calibration model. This, in turn, requires additional amounts of case-relevant data, which is often time-consuming to obtain and prepare. This study proposes two novel normalisation techniques, the Square Root Correction and the Hapax Correction, for deriving likelihood ratios from the AV method LambdaG without the need of a calibration model (Nini et al. 2026). These corrections are designed to mitigate the overestimation of evidential strength that may result from long or highly repetitive texts. Performance is evaluated against logistic regression calibration across fifteen corpora and a range of text lengths (100-9,500 tokens), using the log-likelihood ratio cost (Cllr). The proposed methods achieve performance comparable to logistic regression calibration, with the Hapax Correction outperforming it in approximately 45% of tests (weighted by corpora). Furthermore, performance was more frequently close (within 5%) when the Hapax Correction was outperformed by logistic regression calibration, compared with the reverse comparison. Eliminating the need to train a calibration model reduces data-requirements, time and complexity, thereby increasing the accessibility and transparency of forensic text comparison. This combination of empirical performance and practical advantages supports the adoption of the proposed methods in forensic settings.

arXiv Stat

Terminal Dimension Reduction for Time Series with Applications

arXiv:2607.09490v1 Announce Type: cross Abstract: Terminal embeddings have emerged as a powerful tool for dimension reduction. Given a set of points $P\subset \mathbb{R}^d$, a terminal embedding is a mapping $f:\mathbb{R}^d\rightarrow \mathbb{R}^t$ that preserves the pairwise distance between any pair of points $p\in P$ and $q\in \mathbb{R}^d$ up to small distortion under this mapping. Terminal embeddings have been particularly fruitful for constructing $k$-means and $k$-median coresets, where the objective is to find a typically weighted subset $\Omega$ of $P$ such that for any candidate solution, the cost of the clustering objective on $\Omega$ approximates the cost of the clustering objective on $P$ up to small distortion. Unfortunately, these techniques have not been extended to more complicated structures such as clustering time-series data under common straight-line interpolation between measurements. The main issue is that terminal embeddings, arguably the central technique in this line of research, cannot be linear and are thus not immediately suitable to preserve linear structures. In this work, we develop a generalization of terminal embeddings to affine line-segments that overcomes this issue. We showcase their applicability by using our lines-preserving terminal embeddings to obtain the first dimension-free coresets for clustering time-series under the Fr\'echet distance. The underlying dimension reduction uses Johnson-Lindenstrauss (JL) embeddings, and our experiments indicate that terminal embeddings perform similarly to JL and favorably against PCA for synthetic and real-world time-series, while only terminal embeddings extend pairwise distance preservation to the full ambient space.

arXiv Stat

Neural Collapse Is Forbidden: Information Floors in Language Models

arXiv:2607.09487v1 Announce Type: cross Abstract: Within-class variance in language-model representations is commonly read as incomplete neural collapse. We argue it is allocated information storage, and that the allocation obeys a law. A one-line centering identity voids a family of simplex equiangular-tight-frame claims, including our own earlier ones; in dimensionless variance shares across 14 models, macro-category structure carries only 4-12% of representational variance and within-token context carries 79-91%, stable across a 100x parameter range. On the theory side, token-level weight decay penalizes a category in proportion to its type count, not its occurrence mass, reducing next-token prediction to an imbalanced K-class problem whose optimum orders category norms by type count. A converse floor, proved for binary categories, forces within-category dispersion to be at least proportional to the conditional mutual information I(token; context | category). The law holds: identity dispersion, not total variance, tracks this information across every tested model and partition, under a model-free estimate and even across models, where one model's information predicts another's dispersion; and over pretraining the category share overshoots, decays, and partially recovers, because the information it must carry never left.

arXiv Stat

Similarity search generalisation in contrastive learning with InfoNCE loss

arXiv:2607.09405v1 Announce Type: cross Abstract: Similarity search is a primary application of embedding models trained by contrastive learning. For one of the most popular contrastive learning loss functions, InfoNCE, we show that the population risk with $k$ negative samples is $O(1/k)$ close to an expected cross-entropy which quantifies deviation between i) a softmax similarity search over unseen data using the learned embedding function, and ii) an idealised softmax search over the same data but using similarity implicitly represented in the positive sample generator. This complements existing interpretations of InfoNCE in the $k\to\infty$ limit which are phrased in terms of mutual information, and alignment versus uniformity in embeddings. To quantify generalisation performance, we introduce a new continuity bound for the InfoNCE loss, obtained via G\^{a}teaux differentiation. The bound preserves the structure of averaging over negative samples present in the loss function and features an ``inverse temperature'' parameter which can be tuned to account for the algorithmic temperature. For embedding functions which are Lipschitz in a parameter, this yields a simple demonstration that the averaging effect of $k$ negative samples in the InfoNCE loss carries over to stabilisation of the generalisation error as $k$ grows.

arXiv Stat

Optimal Top-$k$ Identification from Pairwise Comparisons

arXiv:2607.08979v1 Announce Type: cross Abstract: We study the active learning problem of fixed-confidence top-$k$ identification from noisy pairwise comparisons. In this problem, an algorithm sequentially chooses pairs of items to compare, observes the outcomes, and stops when it can return the set of top-$k$ items with error probability at most $\delta$. The objective is to design such a $\delta$-correct procedure that minimizes the expected number of comparisons (the sample complexity). This problem falls within the broader literature on fixed-confidence pure exploration in bandit models, where a common target is asymptotic optimality: the algorithm's expected sample complexity matches the information theoretic lower bound as $\delta \to 0$. Asymptotically optimal procedures have been developed for a range of fixed-confidence pure-exploration problems, however to the best of our knowledge, for top-$1$, or more generally top-$k$ identification from pairwise comparisons under latent utility models an asymptotically optimal algorithm has not been established. In this setting, we develop such an algorithm. We characterize the structure of the lower bound and formulate it as a saddle-point problem. This structure enables a computationally efficient primal-dual procedure that learns the asymptotically optimal comparison allocation online. We then construct an adaptive comparison-allocation algorithm that tracks the allocation learned by the primal-dual procedure and prove it is asymptotically optimal.

arXiv Stat

SPOT-IC: Improving prediction for interval-censored data via survival probability transfer

arXiv:2607.09640v1 Announce Type: new Abstract: Accurate prediction with interval-censored data is particularly challenging when censoring intervals are wide and follow-up is limited, as is common in studies of chronic diseases. Although auxiliary information from source studies may improve prediction in a target study, existing transfer learning methods typically impose restrictive assumptions on model or parameter similarity, or require access to individual-level source data. We propose a novel transfer learning method for interval-censored data that allows arbitrary source models and avoids sharing source data. Our approach transfers survival probability information from source studies through a carefully designed penalty and enables efficient computation via a simple EM algorithm. When multiple source studies are available and their informativeness is unknown, we further develop a data-adaptive aggregation procedure that is robust to negative transfer. Theoretical analysis shows that the proposed estimator attains a faster convergence rate than the target-only estimator whenever at least one source study is sufficiently informative. Extensive simulation studies and an application to data from the Alzheimer's Disease Neuroimaging Initiative demonstrate the effectiveness of our approach.

arXiv Stat

Dynamic Frechet Regression with Feature Selection for Distributional Data

arXiv:2607.09613v1 Announce Type: new Abstract: Many scientific and engineering applications generate responses that are not scalars or vectors, but statistical objects whose form evolves over an ordered index such as time, depth. Probability distributions are a prominent example, capturing variability and uncertainty that cannot be summarized by low-dimensional statistics. When such responses are observed sequentially, the resulting dynamic distributional trajectories pose significant challenges for regression, particularly in relating scalar predictors to both within-index variability and cross-index evolution. We propose Dynamic Fr\'echet Regression (DFR), a framework for modeling index-dependent trajectories of distribution-valued responses. DFR extends Global Fr\'echet Regression by introducing an index-aware weighting mechanism. At each index, predictions are defined as weighted Fr\'echet means in a metric space of distributions (e.g., Wasserstein space), preserving the intrinsic geometry of the response. The weights depend jointly on predictor similarity and index proximity, enabling index-specific prediction while borrowing strength across neighboring indices. To improve interpretability in high-dimensional settings, DFR incorporates a geometry-aware feature selection approach based on sparse metric learning, which identifies predictors driving distributional dynamics without relying on Euclidean coefficients. Simulation studies show improved predictive accuracy and feature recovery over existing methods. An application to additive manufacturing data demonstrates its ability to produce interpretable, index-specific distributional predictions.

arXiv Stat

A censoring-aware target interface for tabular foundation models in survival prediction

arXiv:2607.09577v1 Announce Type: new Abstract: Time-to-event prediction from tabular patient data is central to prognosis and biomedical decision support, but right-censored follow-up prevents direct use of ordinary regression labels. Tabular foundation models offer reusable prediction machinery for modest heterogeneous datasets, yet they generally assume fully observed outcomes. We introduce SurvFM-RMST, a censoring-aware target-interface framework that converts survival outcomes into jackknife pseudo-observation targets for restricted mean survival time, enabling multiple tabular backbones to perform horizon-specific RMST regression without survival-specific fine-tuning. In controlled simulations with known conditional RMST, SurvFM-RMST recovered restricted event-free time accurately, and pseudo-RMST targets outperformed naive restricted observed-time and event-only targets. Across 36 eligible static SurvSet datasets, SurvFM backbones were competitive with established survival and RMST-regression comparators, though relative performance varied by endpoint, horizon and practical constraints. Predicted RMST further stratified held-out patients into groups with ordered observed event-free time and event enrichment. Overall, the results support pseudo-RMST target construction as a portable interface between censored survival data and tabular foundation-model prediction.

arXiv Stat

An Explicit Link between Extreme Value Theory and Compositional Data Analysis

arXiv:2607.09567v1 Announce Type: new Abstract: Extreme value theory and compositional data analysis both study settings where relative information plays a central role. In multivariate extreme value theory, threshold exceedance limits satisfy homogeneity properties that separate the radial size of an extreme event from its relative profile. In compositional data analysis, positive vectors are analysed up to multiplicative scale, and inference is based on ratios or log-ratios between components. Consequently, both fields have developed several covariance and dependence representations of the underlying relative structure. In the H\"usler-Reiss model for extremes, these include variogram, covariance, and precision parametrizations. In compositional data analysis, analogous representations arise from pairwise log-ratios, centred log-ratios, and additive log-ratios. We establish an explicit link between the two fields that relates these different representations by a small set of simple transformations, including oblique projections, H\"usler-Reiss inverses, and the variogram map. From a methodological perspective, leveraging this algebraic connection enables the transfer of statistical approaches from one field to the other. For instance, we introduce intrinsic logistic-normal graphical models for compositional data, which are based on H\"usler-Reiss graphical models for extremes. Conversely, we explore how dimensionality reduction methods from compositional data analysis can be applied to the analysis of multivariate extremes.

arXiv Stat

Global testing of SNP-methylation interactions on binary phenotypes via a logistic functional regression model

arXiv:2607.09535v1 Announce Type: new Abstract: Understanding how genetic and epigenetic factors jointly influence binary health outcomes remains a major challenge in biomedical research. We propose a global test for the overall effect of interactions between DNA methylation and a set of single nucleotide polymorphisms (SNPs) on a binary phenotype. We propose a logistic functional regression model in which methylation measurements at CpG sites are transformed into smooth functional predictors interacting with discrete SNP genotypes through a localized kernel. This framework enables stable inference on region-level interactions while accounting for the spatial structure of methylation around SNPs. Extensive simulations show that the proposed test provides well-calibrated type I error and improved power over classical SNP-CpG pairwise analyses. The practical relevance of the method is illustrated using publicly available methylation and genotyping data from an obesity case-control study.

arXiv Stat

comprisk: A scikit-learn-compatible Python toolkit for competing-risks survival analysis

arXiv:2607.09431v1 Announce Type: new Abstract: Medical time-to-event data are frequently subject to competing risks, where the occurrence of one terminal event precludes the others and standard survival methods that treat competing events as censoring yield biased absolute-risk estimates. Correct analysis instead targets the cause-specific cumulative incidence function (CIF). This methodology has been available to applied researchers almost exclusively through R packages, forcing Python-based machine-learning workflows into a Python-to-R round trip. We present comprisk, a scikit-learn-compatible Python toolkit that consolidates the canonical competing-risks methods (a scalable competing-risks random survival forest together with Fine-Gray subdistribution-hazard regression including a penalized variant, cause-specific Cox regression, the Aalen-Johansen CIF estimator, and Gray's K-sample test) behind a single, consistent API, and adds competing-risks-aware model evaluation (inverse probability of censoring weighted time-dependent AUC and Brier score, cause-specific concordance indices with closed-form confidence intervals, and calibration curves). Every estimator is validated numerically against the established R reference implementations. The forest uses a histogram-based, numba-compiled split kernel that fits 10-22x faster than randomForestSRC at comparable discrimination on real electronic-health-record cohorts and scales to n = 10^6 on a consumer CPU. comprisk is distributed on PyPI and lets applied researchers perform correct and scalable competing-risks analysis entirely within the Python scientific stack.

arXiv Stat

Decision-analytical models as causal models

arXiv:2607.09397v1 Announce Type: new Abstract: Health economic evaluations are fundamentally concerned with answering causal questions by targeting estimands that contrast the costs and health consequences that would be observed under at least two different interventions. This requires the joint distribution of potential outcomes under each level of intervention, which, with appropriate causal assumptions, can in principle be identified from the joint distribution of observed health outcomes. Such data, however, are rarely available from a single source. This limitation has motivated the use of decision-analytical models to approximate the joint distribution of outcomes under each intervention directly, informed by causal parameters drawn and synthesized from multiple sources, so that the potential outcomes of interest can be approximated as an expectation over the model-implied outcome trajectories. The validity of this approach, however, depends on the credibility of the underlying assumptions. In this work, we formalize this procedure explicitly as a task of causal inference, thereby defining and decomposing decision-analytical model bias into components arising from model structure (model bias) and input parameters (target bias). Because decision-analytical models often rely on unconventional target parameters lacking straightforward observable analogues, and because bias in these parameters can propagate through the model, target bias may arise even in simple settings, a point of central focus in this work. More broadly, this work provides a unifying foundation for medical decision-analytical modelling and causal inference, making explicit the potential for decision-analytical model bias and the role of causal assumptions contributing to it. Ultimately, the resulting clinical decision is only as credible as the assumptions underlying it.

arXiv Stat

Spectrally Deconfounded Gradient Boosting

arXiv:2607.09371v1 Announce Type: new Abstract: Flexible machine-learning methods can be sensitive to hidden confounding: they may learn associations induced by unobserved confounders rather than stable signals. Spectral deconfounding mitigates this problem by shrinking high-variance directions of the covariate matrix that, under dense confounding, carry latent confounder information. Existing work has largely focused on linear models. We develop a nonlinear spectral deconfounding framework for gradient boosting. Our approach replaces the ordinary squared-error loss by a spectral loss, which alters the boosting dynamics by slowing down learning in confounding-aligned directions. We show that deconfounding is not achieved by the spectral loss alone, but by the interaction between spectral shrinkage and regularization, especially in terms of early stopping. Moreover, we provide a mixed-model interpretation that connects LAVA-type shrinkage to random-effects adjustment and yields an empirical-Bayes procedure for tuning the spectral loss. We also extend the method to general likelihoods and nonlinear confounding using Laplace approximations and kernel random effects. Across synthetic and real-world experiments, spectrally deconfounded boosting improves estimation of the target function under hidden confounding and is substantially more scalable than existing nonlinear spectral deconfounding baselines.

arXiv Stat

Inference of large scale relational state processes

arXiv:2607.09363v1 Announce Type: new Abstract: Relational states refer to concepts such as friendship or collaboration, in which a relationship persists over a certain amount of time. Study of relational states often involves figuring out what factors contribute to the creation or dissolution of these relationships. However, most methods available now restrict their attention to binary states, i.e., ties that are either present or absent, even though many real-world systems evolve through multiple relational states (e.g., acquaintance, friendship, close friendship). We propose a continuous-time framework for modelling and inferring relational state networks in which each edge evolves by transitioning between two or more states. In our model, transition intensities are driven by state-dependent covariates that might be decomposed into anchoring (current-state) and pulling (target-state) mechanisms, with both linear and smooth non-linear effects. We address two common sampling regimes. With full event histories, a Cox-type partial likelihood with nested case-control sampling enables efficient estimation of both parametric and smooth effects. Instead, for panel data we derive a general ODE formulation for the likelihood, which leads to a particularly efficient inference procedure for binary state model. Simulation studies confirm accurate recovery of model parameters, and an empirical application to adolescent friendship data reproduces the substantive conclusions of established modelling techniques while offering substantial computational gains. The framework preserves the interpretability of classical network effects, generalizes them to multi-state ties, and scales to larger, more complex designs under both full-history and panel sampling designs.

arXiv Stat

Rethinking the Scientific Method: An Introduction to Bayesian Epistemology

arXiv:2607.09281v1 Announce Type: new Abstract: Scientists routinely disagree not about data but about how to interpret evidence, because they implicitly operate from different epistemological frameworks without recognising it. The two dominant traditions, confirmationism and falsificationism, each capture genuine insights about scientific reasoning but face well-documented limitations. Confirmationism provides a natural account of how evidence supports hypotheses but cannot escape the problem of induction. Falsificationism provides logical rigour through deductive refutation but is undermined by the Duhem-Quine problem and offers no account of how scientists rationally accept theories and act on them. Here we argue that Bayesian epistemology provides a practical resolution to this impasse. By treating evidence probabilistically and operating over a finite, revisable set of hypotheses, the framework recovers the valid contributions of both traditions while addressing their core weaknesses. We show that confirmation and falsification emerge as special cases of Bayesian updating, that the subjectivity objection to priors is weaker than commonly supposed, and that the framework has direct practical consequences for study design, evidence synthesis, and publishing norms. Specifically, it replaces the falsifiability criterion with the more useful question of whether a hypothesis makes predictions that discriminate between competitors, and reframes the reproducibility crisis as an epistemological rather than a purely statistical problem. Adopting Bayesian epistemology, even informally as a mental model, can reduce friction between researchers, improve research efficiency, and help restore the cumulative character of scientific progress.

arXiv Stat

Influence Diagnostics in High-dimensional M-estimation: Precise Asymptotics

arXiv:2607.09250v1 Announce Type: new Abstract: The impact of a given training point on a statistical model is classically measured through its leave-one-out influence, which quantifies the effect of its removal from the training set on the model accuracy. While the statistics of leave-one-out influences are well understood in the low-dimensional, large sample limit $n\to \infty, d=O(1)$, they become more intricate in high dimensions, as the influence of a given sample develops non-trivial dependencies on all other training samples. For convex M-estimation under Gaussian design, in the high-dimensional limit $n\asymp d$, we show that the distribution of the influences across the training set converges to a limiting measure which we sharply characterize. Building on these results, we provide evidence that influential samples tend to lie close to the decision boundary, thereby making contact with a standard data selection heuristic in active learning.

arXiv Stat

Causal Perspectives on Network Meta-Analysis

arXiv:2607.09200v1 Announce Type: new Abstract: Pairwise and network meta-analyses occupy the highest tier of evidence-based medicine and routinely inform clinical guidelines and healthcare decision-making. Current approaches typically aggregate study-level treatment effects to obtain an overall estimate. We argue that the causal estimand should come first, with the aggregation derived only afterwards: the target population and the relevant sources of between-study heterogeneity should be explicitly defined before deriving the aggregation required for identification. This shift in perspective fundamentally changes both the estimands and the methodology. We develop a unified causal framework for pairwise and network meta-analysis based on aggregate data. By defining treatment effects with respect to a clinically meaningful target population, for example, the average population represented by the contributing trials, and accounting for heterogeneity induced by treatment-effect modifiers and center effects, we show that identification naturally leads to arm-level aggregation. In the network setting, this causal formulation departs fundamentally from the conventional contrast-based paradigm: arm-level aggregation emerges from the causal formulation rather than from a modeling choice, and treatment effects are identified without relying on the treatment network itself. This perspective provides an additional conceptual argument in the long-standing contrast-based versus arm-based debate. Numerical studies show that the proposed estimators target well-defined causal effects, whereas the causal interpretation of conventional approaches remains unclear. Although both approaches often produce similar estimates, we identify settings in which they diverge, with potentially important implications for the interpretation of meta-analytic evidence.

arXiv Stat

Orthogonalized Design Matrices Speed-ups of Bayesian Semiparametric Regression

arXiv:2607.09013v1 Announce Type: new Abstract: We explain how important classes of Bayesian semiparametric regression fitting and inference procedures can be sped up, significantly, via the use of orthogonalized design matrices. Typically, design matrices in semiparametric regression contain predictor observations and basis functions of such data. In Bayesian semiparametric regression, loop-type approaches such as Gibbs sampling and coordinate ascent variational inference typically are required. We show that pre-loop reformulation of Bayesian semiparametric regression models involving orthogonalized design matrices lead to two orders of magnitude, with respect to column dimension, computational reduction. Our computer experiments reveal that this simple paradigm results in approximately 5- to 60-fold speed-ups.

arXiv Stat

A Statistical Test for the Benefits of Personalizing Interventions

arXiv:2607.08951v1 Announce Type: new Abstract: From medicine to marketing to social sciences, the promise of tailoring interventions to individuals is undeniable. However, practical applications force weighing personalization's potential benefits with its possible increased cost and fragility. We introduce a statistical hypothesis test that evaluates, given historical data, evidence that a personalized intervention policy's performance will surpass deploying the best single intervention. The test maintains strict type-I error control while achieving asymptotic normality with the minimal possible variance under specified conditions. Results on diverse datasets from job training, depression treatment, education and recommendation systems demonstrate the test's versatility and its superior performance over alternatives. This test can support decision-makers throughout the intervention sciences by providing a simple and powerful quantification of the potential benefits of personalization.

← 이전2 / 1260다음 →