Papers

Out-of-distribution (OOD) detection plays a crucial role in ensuring the robustness of machine learning systems deployed in real-world applications. Recent approaches have explored the use of unlabeled data, showing potential for enhancing OOD detection capabilities. However, effectively utilizing unlabeled in-the-wild data remains challenging due to the mixed nature of both in-distribution (InD) and OOD samples. The lack of a distinct set of OOD samples complicates the task of training an optimal OOD classifier. In this work, we introduce Medix, a novel framework designed to identify potential outliers from unlabeled data using the median-based robust gradient statistics. We use the median because it provides a stable estimate of the central tendency, as an OOD detection mechanism, due to its robustness against noise and outliers. Using these identified outliers, along with labeled InD data, we train a robust OOD classifier. From a theoretical perspective, we derive error bounds that demonstrate Medix achieves a low error rate. Empirical results further substantiate our claims, as Medix outperforms existing methods across the board in open-world settings.

Gaussian mixture models are widely used to model data generated from multiple latent sources. Despite its popularity, most theoretical research assumes that the labels are either independent and identically distributed, or follows a Markov chain. It remains unclear how the fundamental limits of estimation change under more complex dependence. In this paper, we address this question for the spherical two-component Gaussian mixture model. We first show that for labels with an arbitrary dependence, a naive estimator based on the misspecified likelihood is $\sqrt{n}$-consistent. Additionally, under labels that follow an Ising model, we establish the information theoretic limitations for estimation, and discover an interesting phase transition as dependence becomes stronger. When the dependence is smaller than a threshold, the optimal estimator and its limiting variance exactly matches the independent case, for a wide class of Ising models. On the other hand, under stronger dependence, estimation becomes easier and the naive estimator is no longer optimal. Hence, we propose an alternative estimator based on the variational approximation of the likelihood, and argue its optimality under a specific Ising model.

Estimation in exploratory factor analysis often yields estimates on the boundary of the parameter space. Such occurrences, known as Heywood cases, are characterised by non-positive variance estimates and can cause issues in numerical optimisation procedures or convergence failures, which, in turn, can lead to misleading inferences, particularly regarding factor scores and model selection. We derive sufficient conditions on the model and a penalty to the log-likelihood function that i) guarantee the existence of maximum penalised likelihood estimates in the interior of the parameter space, and ii) ensure that the corresponding estimators possess the desirable asymptotic properties expected by the maximum likelihood estimator, namely consistency and asymptotic normality. Consistency and asymptotic normality are achieved when the penalisation is soft enough, in a way that adapts to the information accumulation about the model parameters. We formally show, for the first time, that the penalties of Akaike (1987) and Hirose et al. (2011) to the log-likelihood of the normal linear factor model satisfy the conditions for existence, and, hence, deal with Heywood cases. Their vanilla versions, though, can result in questionable finite-sample properties in estimation, inference, and model selection. The maximum softly-penalised likelihood framework we introduce enables the careful scaling of those penalties to ensure that the resulting estimation and inference procedures inherit the ML estimator's optimal properties. Through comprehensive simulation studies and the analysis of real data sets, we illustrate the desirable finite-sample properties of the maximum softly penalised likelihood estimators and associated procedures.

Hyperparameter tuning is a challenging problem especially when the system itself involves uncertainty. Due to noisy function evaluations, optimization under uncertainty can be computationally expensive. In this paper, we present a novel Bayesian optimization framework tailored for hyperparameter tuning under uncertainty, with a focus on optimizing a scale- or precision-type parameter in stochastic models. The proposed method employs a statistical surrogate for the underlying random variable, enabling analytical evaluation of the expectation operator. Moreover, we derive a closed-form expression for the optimizer of the random acquisition function, which significantly reduces computational cost per iteration. Numerical experiments demonstrate substantial gains in optimization efficiency: about 40-fold reduction in data usage relative to a Monte Carlo-based one-dimensional optimization scheme, and an order of magnitude improvement relative to a Gaussian process-based Bayesian optimization baseline.

Learning from temporally-correlated data is a core facet of modern machine learning. Yet our understanding of sequential learning remains incomplete, particularly in the multi-trajectory setting where data consists of many independent realizations of a time-indexed stochastic process. This important regime both reflects modern training pipelines such as for large foundation models, and offers the potential for learning without the typical mixing assumptions made in the single-trajectory case. However, instance-optimal bounds are known only for least-squares regression with dependent covariates; for more general models or loss functions, the only broadly applicable guarantees result from a reduction to either i.i.d. learning, with effective sample size scaling only in the number of trajectories, or an existing single-trajectory result when each individual trajectory mixes, with effective sample size scaling as the full data budget deflated by the mixing-time. In this work, we significantly broaden the scope of instance-optimal rates in multi-trajectory settings via the Hellinger localization framework, a general approach for maximum likelihood estimation. Our method proceeds by first controlling the squared Hellinger distance at the path-measure level via a reduction to i.i.d. learning, followed by localization as a quadratic form in parameter space weighted by the trajectory Fisher information. This yields instance-optimal bounds that scale with the full data budget under a broad set of conditions. We instantiate our framework across four diverse case studies: a simple mixture of Markov chains, dependent linear regression under non-Gaussian noise, generalized linear models with non-monotonic activations, and linear-attention sequence models. In all cases, our bounds nearly match the instance-optimal rates from asymptotic normality, substantially improving over standard reductions.

In supervised classification tasks, models are trained to predict a label for each data point. In real-world datasets, these labels are often noisy due to annotation errors. While the impact of label noise on the performance of deep learning models has been widely studied, its effects on the networks' hidden representations remain poorly understood. We address this gap by systematically comparing hidden representations using the Information Imbalance, a computationally efficient proxy of conditional mutual information. Through this analysis, we observe that the information content of the hidden representations follows a double descent as a function of the number of network parameters, akin to the behavior of the test error. We further demonstrate that in the underparameterized regime, representations learned with noisy labels are more informative than those learned with clean labels, while in the overparameterized regime, these representations are equally informative. Our results indicate that the representations of overparameterized networks are robust to label noise. We also found that the information imbalance between the penultimate and pre-softmax layers decreases with cross-entropy loss in the overparameterized regime. This offers a new perspective on understanding generalization in classification tasks. Extending our analysis to representations learned from random labels, we show that these perform worse than random features. This indicates that training on random labels drives networks much beyond lazy learning, as weights adapt to encode labels information.

Calibrated predictions are useful because their numerical values can be interpreted as probabilities. Calibration errors are therefore widely used to evaluate, compare, and tune probabilistic predictors. Recently, Haghtalab et al. (2024) introduced an additional requirement for such measures: truthfulness. A calibration measure is truthful if a predictor minimizes its expected measured error by reporting the true conditional label distribution. Many standard empirical calibration errors are non-truthful: a predictor may appear better calibrated by distorting its probabilities rather than reporting them truthfully. We study the practical role of truthfulness for calibration measurement in multiclass prediction. First, we introduce perfectly truthful calibration errors for multidimensional linear properties of the label distribution, generalizing the truthful calibration error for binary predictions in Hartline et al. (2025). This framework includes full multiclass calibration and classwise calibration. We also identify a truthful correction for confidence calibration. Second, we characterize the decision-theoretic implications of these truthful errors. For calibrated predictors, truthful calibration errors preserve the Blackwell dominance: a more informative calibrated predictor receives no larger expected error. Third, we show that this decision-theoretic interpretation explains and mitigates the well-observed ranking robustness problem of binned calibration errors. Empirically, non-truthful confidence-based errors can reverse model rankings when the number of bins changes, while our truthful errors give more stable rankings across binning choices.

We provide an upper bound on the number of neurons required in a shallow neural network to approximate a continuous function on a compact set with a given accuracy. This method, inspired by a specific proof of the Stone-Weierstrass theorem, is constructive and more general than previous bounds of this character, as it applies to any continuous function on any compact set.

Forest structural complexity metrics integrate multiple canopy attributes into a single value that reflects habitat quality and ecosystem function. Spaceborne lidar from the Global Ecosystem Dynamics Investigation (GEDI) has enabled mapping of structural complexity in temperate and tropical forests, but its sparse sampling limits continuous high-resolution mapping. We present a scalable, deep learning framework fusing GEDI observations with multimodal Synthetic Aperture Radar (SAR) datasets to produce global, high-resolution (25 m) wall-to-wall maps of forest structural complexity. Our adapted EfficientNetV2 architecture, trained on over 130 million GEDI footprints, achieves high performance (global R2 = 0.82) with fewer than 400,000 parameters, making it an accessible tool that enables researchers to process datasets at any scale without requiring specialized computing infrastructure. The model produces accurate predictions with calibrated uncertainty estimates across biomes and time periods, preserving fine-scale spatial patterns. It has been used to generate a global, multi-temporal dataset of forest structural complexity from 2015 to 2022. Through transfer learning, this framework can be extended to predict additional forest structural variables with minimal computational cost. This approach supports continuous, multi-temporal monitoring of global forest structural dynamics and provides tools for biodiversity conservation and ecosystem management efforts in a changing climate.

To reliably project future sea level rise, ice sheet models require inputs that respect physics. Embedding physical principles like mass conservation into models that interpolate Antarctic ice flow vector fields from sparse & noisy measurements not only promotes physical adherence but can also improve accuracy and robustness. While physics-informed neural networks (PINNs) impose physics as soft penalties, offering flexibility but no physical guarantees, we instead propose divergence-free neural networks (dfNNs), which enforce local mass conservation exactly via a vector calculus trick. Our comparison of dfNNs, PINNs, and unconstrained NNs on ice flux interpolation over Byrd Glacier suggests that "mass conservation on rails" yields more reliable estimates, and that directional guidance, a learning strategy leveraging continent-wide satellite velocity data, boosts performance across models.

Financial fraud has been growing exponentially in recent years. The rise of cryptocurrencies as an investment asset has simultaneously seen a parallel growth in cryptocurrency scams. To detect possible cryptocurrency fraud, and in particular market manipulation, previous research focused on the detection of changes in the network of trades; however, market manipulators are now trading across multiple cryptocurrency platforms, making their detection more difficult. Hence, it is important to consider the identification of changes across several trading networks or a `network of networks' over time. To this end, in this article, we propose a new change-point detection method in the network structure of tensor-variate data. This new method, labeled TenSeg, first employs a tensor decomposition, and second detects multiple change-points in the second-order (cross-covariance or network) structure of the decomposed data. It allows for change-point detection in the presence of frequent changes of possibly small magnitudes and is computationally fast. We apply our method to several simulated datasets and to a cryptocurrency dataset, which consists of network tensor-variate data from the Ethereum blockchain. We demonstrate that our approach substantially outperforms other state-of-the-art change-point techniques, and the detected change-points in the Ethereum data set coincide with changes across several trading networks or a `network of networks' over time. Finally, all the relevant R code implementing the method in the article are available on https://github.com/Anastasiou-Andreas/TenSeg.

Italy reports some of the lowest levels of mortality in the developed world. Recent evidence, however, suggests that even in low mortality countries improvements may be slowing and regional inequalities widening. This study contributes new empirical evidence to the debate by analysing mortality data by single year of age for males and females across 107 provinces in Italy from 2002 to 2019. We extend the widely used Lee Carter model to include spatially varying age specific effects, and further specify it to capture space age time interactions. The model is estimated in a Bayesian framework using the inlabru package, which builds on INLA (Integrated Nested Laplace Approximation) for non linear models and facilitates the use of smoothing priors. This approach borrows strength across provinces and years, mitigating random fluctuations in small area death counts. Results demonstrate the value of such a granular approach, highlighting the existence of an uneven geography of mortality despite overall national improvements. Mortality disadvantage is concentrated in parts of the Centre South and North West, while the Centre North and North East fare relatively better. These geographical differences have widened since 2010, with clear age and gender specific patterns, being more pronounced at younger adult ages for men and at older adult ages for women. Future work may involve refining the analysis to mortality by cause of death or socioeconomic status, informing more targeted public health policies to address mortality disparities across Italy's provinces.

Atrial fibrillation (AF) is a common cardiac arrhythmia characterised by disordered electrical activity in the atria. The standard treatment is catheter ablation, which is invasive and irreversible. Recent advances in computational electrophysiology offer the potential for patient-specific models, often referred to as digital twins, that can be used to guide clinical decisions. To be of practical value, we must be able to rapidly calibrate physics-based models using routine clinical measurements. We pose this calibration task as a static inverse problem, where the goal is to infer tissue-level electrophysiological parameters from the available observations. To make this tractable, we replace the expensive forward model with Gaussian process emulators (GPEs), and propose a novel adaptation of the ensemble Kalman filter (EnKF) for static non-linear inverse problems. The approach yields parameter samples that can be interpreted as coming from the best Gaussian approximation of the posterior distribution. We compare our results with those obtained using Markov chain Monte Carlo (MCMC) sampling and demonstrate the potential of the approach to enable near-real-time patient-specific calibration, a key step towards predicting outcomes of AF treatment within clinical timescales. The approach is readily applicable to a wide range of static inverse problems in science and engineering.

Uncertainty quantification (UQ) over graphs arises in a number of safety-critical applications in network science. The Gaussian process (GP), as a classical Bayesian framework for UQ, has been developed to handle graph-structured data by devising topology-aware kernel functions. However, such GP-based approaches are limited not only by the prohibitive computational complexity, but also the strict modeling assumptions that might yield poor coverage, especially with labels arriving on the fly. To effect scalability, we devise a novel graph-aware parametric GP model by leveraging the random feature (RF)-based kernel approximation, which is amenable to efficient recursive Bayesian model updates. To further allow for adaptivity, an ensemble of graph-aware RF-based scalable GPs have been leveraged, with per-GP weight adapted to data arriving incrementally. To ensure valid coverage with robustness to model mis-specification, we wed the GP-based set predictors with the online conformal prediction framework, which post-processes the prediction sets using adaptive thresholds. Experimental results the proposed method yields improved coverage and efficient prediction sets over existing baselines by adaptively ensembling the GP models and setting the key threshold parameters in CP.

This paper demonstrates that, under a particular convention, the convex functions that characterise the phi divergences also generate Archimedean copulas in at least two dimensions. As a special case, we develop the family of Archimedean copulas associated with the important family of power divergences, which we call the power-divergence copulas. The properties of the family are extensively studied, including the subfamilies that are absolutely continuous or have a singular component, the ordering of the family, limiting cases (i.e., the Frechet-Hoeffding lower bound and Frechet-Hoeffding upper bound), the Kendall's tau and tail-dependence coefficients, and cases that extend to three or more dimensions. In an illustrative application, the power-divergence copulas are used to model a Danish fire insurance dataset. It is shown that the power-divergence copulas provide an adequate fit to the bivariate distribution of two kinds of fire-related losses claimed by businesses, while several benchmarks (a suite of well known Archimedean, extreme-value, and elliptical copulas) do not.

Feature attributions are post-training analysis methods that assess how various input features of a machine learning model contribute to an output prediction. Their interpretation is straightforward when features act independently, but it becomes less clear when the predictive model involves interactions, such as multiplicative relationships or joint feature contributions. In this work, we propose a general theory of higher-order feature attribution, which we develop on the foundation of Integrated Gradients (IG). This work extends existing frameworks in the literature on explainable AI. When using IG as the method of feature attribution, we discover natural connections to statistics and topological signal processing. We provide several theoretical results that establish the theory, and we validate our theory on a few examples.

Financial spillovers in interconnected systems, such as global banking networks, require tools that capture temporal and frequency dynamics, while incorporating the underlying network topology. While current network time series models are developed in the time-domain, frequency-domain approaches, which reveal how cross-nodal dependencies vary across different cycles, remain under-explored. This paper develops a spectral analysis framework that accommodates flexible forms of network dependence, including interactions mediated through intermediate nodes. This ensures that inter-nodal relationships are not restricted to direct connections, a feature crucial for capturing indirect financial spillovers. We define the network time series spectral density, alongside coherence and partial coherence, and propose both parametric and network-constrained nonparametric methods for their estimation. Simulations and theoretical results demonstrate the strong performance of the parametric approach when the data-generating process aligns with the model structure, whereas the nonparametric alternative provides robustness against model misspecification. An application to global bank connectedness shows that the proposed spectral measures capture inter-bank frequency-specific spillover effects, yielding results consistent with existing measures while additionally uncovering richer patterns of volatility transmission that are intimately connected to the network topology.

Temporal difference (TD) learning is a cornerstone of reinforcement learning. In the average-reward setting, standard TD($\lambda$) is highly sensitive to the choice of step-size and thus requires careful tuning to maintain numerical stability. We introduce average-reward implicit TD($\lambda$), which employs an implicit fixed point update to provide data-adaptive stabilization while preserving the per iteration computational complexity of standard average-reward TD($\lambda$). In contrast to prior finite-time analyses of average-reward TD($\lambda$), which impose restrictive step-size conditions, we establish finite-time error bounds for the implicit variant under substantially weaker step-size requirements. Empirically, average-reward implicit TD($\lambda$) operates reliably over a much broader range of step-sizes and exhibits markedly improved numerical stability. This enables more efficient policy evaluation and policy learning, highlighting its effectiveness as a robust alternative to average-reward TD($\lambda$).

Latent space models assume that network ties are more likely between nodes that are closer together in an underlying latent space. Euclidean space is a popular choice for the underlying geometry, but hyperbolic geometry can mimic more realistic patterns of ties in complex networks. To identify the underlying geometry, past research has applied non-Euclidean extensions of multidimensional scaling (MDS) to the observed geodesic distances: the shortest path lengths between nodes. The difference in stress, a standard goodness-of-fit metric for MDS, across the geometries is then used to select a latent geometry with superior model fit (lower stress). The effectiveness of this method is assessed through simulations of latent space networks in Euclidean and hyperbolic geometries. To better account for uncertainty, we extend permutation-based hypothesis tests for MDS to the latent network setting. However, these tests do not incorporate any network structure. We propose a parametric bootstrap distribution of networks, conditioned on observed geodesic distances and the Gaussian Latent Position Model (GLPM). Our method extends the Davidson-MacKinnon J-test to latent space network models with differing latent geometries. We pay particular attention to large and sparse networks, and both the permutation test and the bootstrapping methods show an improvement in detecting the underlying geometry.

We investigate the problem of characterizing the optimal variance proxy for sub-Gaussian random variables,whose moment-generating function exhibits bounded growth at infinity. We apply a general characterization method to discrete random variables with equally spaced atoms. We thoroughly study 3-mass distributions, thereby generalizing the well-studied Bernoulli case. We also prove that the discrete uniform distribution over $N$ points is strictly sub-Gaussian. Finally, we provide an open-source Python package that combines analytical and numerical approaches to compute optimal sub-Gaussian variance proxies across a wide range of distributions.

Existing methods for evaluating graph generative models primarily rely on Maximum Mean Discrepancy (MMD) metrics based on graph descriptors. While these metrics can rank generative models, they do not provide an absolute measure of performance. Their values are also highly sensitive to extrinsic parameters, namely kernel and descriptor parametrization, making them incomparable across different graph descriptors. We introduce PolyGraph Discrepancy (PGD), a new evaluation framework that addresses these limitations. It approximates the Jensen-Shannon distance of graph distributions by fitting binary classifiers to distinguish between real and generated graphs, featurized by these descriptors. The data log-likelihood of these classifiers approximates a variational lower bound on the JS distance between the two distributions. Resulting metrics are constrained to the unit interval [0,1] and are comparable across different graph descriptors. We further derive a theoretically grounded summary metric that combines these individual metrics to provide a maximally tight lower bound on the distance for the given descriptors. Thorough experiments demonstrate that PGD provides a more robust and insightful evaluation compared to MMD metrics. The PolyGraph framework for benchmarking graph generative models is made publicly available at https://github.com/BorgwardtLab/polygraph-benchmark.

In this paper, we first situate the challenges for measuring data quality under Project Lighthouse in the broader academic context. We then discuss in detail the three core data quality metrics we use for measurement–two of which extend prior academic work. Using those data quality metrics as examples, we propose a framework, based on machine learning classification, for empirically justifying the choice of data quality metrics and their associated minimum thresholds. Finally we outline how these methods enable us to rigorously meet the principle of data minimization when analyzing potential experience gaps under Project Lighthouse, which we term quantitative data minimization.

Deep neural networks have achieved remarkable success, yet our understanding of how they learn remains limited. These models can learn high-dimensional tasks, which is generally statistically intractable due to the curse of dimensionality. This apparent paradox suggests that learnable data must have an underlying latent structure. What is the nature of this structure? How do neural networks encode and exploit it, and how does it quantitatively impact performance - for instance, how does generalization improve with the number of training examples? This thesis addresses these questions by studying the roles of locality and compositionality in data, tasks, and deep learning representations.

Mixtures of linear dynamical systems (MoLDS) provide a path to model time-series data that exhibit diverse temporal dynamics across trajectories. However, its application remains challenging in complex and noisy settings, limiting its effectiveness for neural data analysis. Tensor-based moment methods can provide global identifiability guarantees for MoLDS, but their performance degrades under noise and complexity. Commonly used expectation-maximization (EM) methods offer flexibility in fitting latent models but are highly sensitive to initialization and prone to poor local minima. Here, we propose a tensor-based method that provides identifiability guarantees for learning MoLDS, which is followed by EM updates to combine the strengths of both approaches. The novelty in our approach lies in the construction of moment tensors using the input-output data to recover globally consistent estimates of mixture weights and system parameters. These estimates can then be refined through a Kalman EM algorithm, with closed-form updates for all LDS parameters. We validate our framework on synthetic benchmarks and real-world datasets. On synthetic data, the proposed Tensor-EM method achieves more reliable recovery and improved robustness compared to either pure tensor or randomly initialized EM methods. We then analyze neural recordings from the primate somatosensory cortex while a non-human primate performs reaches in different directions. Our method successfully models and clusters different conditions as separate subsystems, consistent with supervised single-LDS fits for each condition. Finally, we apply this approach to another neural dataset where monkeys perform a sequential reaching task. These results demonstrate that MoLDS provides an effective framework for modeling complex neural data, and that Tensor-EM is a reliable approach to MoLDS learning for these applications.

In this paper, we investigate the general problem of assessing symmetry in data points on the hyper-dimensional torus, a question that originally emerged in applications from bioinformatics and directional statistics. We develop optimal tests for symmetry for both scenarios where the center of symmetry is known and where it is unknown. Our new tests are not only valid under a given parametric hypothesis but also under a very broad class of symmetric distributions. The asymptotic behavior of the proposed tests is studied both under the null hypothesis and local alternatives. A key contribution of our paper is that we accompany our asymptotic results with error guarantees by deriving quantitative bounds on the distributional distance between the exact (unknown) distribution of the test statistic and its asymptotic counterpart by leveraging Stein's method. The finite-sample performance of the tests is evaluated through simulation studies, and their practical utility in bioinformatics is demonstrated via an application to protein folding data.

Phytoplankton are microscopic algae responsible for roughly half of the world's photosynthesis that play a critical role in global carbon cycles and oxygen production, and measuring the abundance of their subtypes across a wide range of spatiotemporal scales is of great relevance to oceanography. High-frequency flow cytometry is a powerful technique in which oceanographers at sea can rapidly record the optical properties of tens of thousands of individual phytoplankton cells every few minutes. Identifying distinct subpopulations within these vast datasets (a process known as "gating") remains a major challenge and has largely been performed manually so far. In this paper, we introduce a fast, automated gating method, which accurately identifies phytoplankton populations by fitting a time-evolving mixture of Gaussians model using an expectation-maximization-like algorithm with kernel smoothing. We use simulated data to demonstrate the validity and robustness of this approach, and use oceanographic cruise data to highlight the method's ability to not only replicate but surpass expert manual gating. Finally, we provide the flowkernel R package, written in literate programming, that implements the algorithm efficiently.

Measuring dependence between random variables is a fundamental problem in Statistics, with applications across diverse fields. While classical measures such as Pearson's correlation have been widely used for over a century, they have notable limitations, particularly in capturing nonlinear relationships and extending to general metric spaces. In recent years, the theory of Optimal Transport and Wasserstein distances has provided new tools to define measures of dependence that generalize beyond Euclidean settings. This survey explores recent proposals, outlining two main approaches: one based on the distance between the joint distribution and the product of marginals, and another leveraging conditional distributions. We discuss key properties, including characterization of independence, normalization, invariances, robustness, sample, and computational complexity. Additionally, we propose an alternative perspective that measures deviation from maximal dependence rather than independence, leading to new insights and potential extensions. Our work highlights recent advances in the field and suggests directions for further research in the measurement of dependence using Optimal Transport.

This paper provides data-dependent bounds on the expected error of the Gibbs algorithm in the overparameterized interpolation regime, where low training errors are also obtained for impossible data, such as random labels in classification. The results show that generalization in the low-temperature regime is already signaled by small training errors in the noisier high-temperature regime. The bounds are stable under approximation with Langevin Monte Carlo algorithms. The analysis motivates the design of an algorithm to compute bounds, which on the MNIST, CIFAR-10, and SVHN datasets yield nontrivial, close predictions on the test error for true labeled data, while maintaining a correct upper bound on the test error for random labels.

Out-of-Distribution (OOD) detection is critical to AI reliability and safety, yet in many practical settings, only a limited amount of training data is available. Bayesian Neural Networks (BNNs) are a promising class of model on which to base OOD detection, because they explicitly represent epistemic (i.e. model) uncertainty. In the small training data regime, BNNs are especially valuable because they can incorporate prior model information. We introduce a new family of Bayesian posthoc OOD scores based on expected logit vectors, and compare 5 Bayesian and 4 deterministic posthoc OOD scores. Experiments on MNIST and CIFAR-10 In-Distributions, with 5000 training samples or less, show that the Bayesian methods outperform corresponding deterministic methods.

This paper develops distribution theory and bootstrap-based inference methods for a broad class of convex pairwise difference estimators. These estimators minimize a kernel-weighted convex-in-parameter function over observation pairs with similar covariates, where the similarity is governed by a localization (bandwidth) parameter. While classical results establish asymptotic normality under restrictive bandwidth conditions, we show that valid Gaussian and bootstrap-based inference remains possible under substantially weaker assumptions. First, we extend the theory of small bandwidth asymptotics to convex pairwise difference estimation settings, deriving robust Gaussian approximations even when a smaller than standard bandwidth is used. Second, we employ a debiasing procedure based on generalized jackknifing to enable inference with larger bandwidths, while preserving convexity of the objective function. Third, we construct a novel bootstrap method that adjusts for bandwidth-induced variance distortions, yielding valid inference across a wide range of bandwidth choices. Our proposed inference method enjoys demonstrably greater robustness, while retaining the practical appeal of convex pairwise difference estimators.