Recent developments in computer vision have enabled the availability of segmented images across various domains, such as medicine, where segmented radiography images play an important role in diagnosis-making. As prediction problems are common in medical image analysis, this work explores the use of segmented images (through the associated contours they highlight) as predictors in a supervised classification context. Consequently, we develop a new approach for image analysis that takes into account the shape of objects within images. For this aim, we introduce a new formalism that extends the study of single random planar curves to the joint analysis of multiple planar curves-referred to here as multivariate planar curves. In this framework, we propose a solution to the alignment issue in statistical shape analysis. The obtained multivariate shape variables are then used in functional classification methods through tangent projections. Detection of cardiomegaly in segmented X-rays and numerical experiments on synthetic data demonstrate the appeal and robustness of the proposed method.

In medical imaging, inter-observer variability among radiologists often introduces label uncertainty, particularly in modalities where visual interpretation is subjective. Lung ultrasound (LUS) is a prime example-it frequently presents a mixture of highly ambiguous regions and clearly discernible structures, making consistent annotation challenging even for experienced clinicians. In this work, we introduce a novel approach to both labeling and training AI models using expert-supplied, per-pixel confidence values. Rather than treating annotations as absolute ground truth, we design a data annotation protocol that captures the confidence that radiologists have in each labeled region, modeling the inherent aleatoric uncertainty present in real-world clinical data. We demonstrate that incorporating these confidence values during training leads to improved segmentation performance. More importantly, we show that this enhanced segmentation quality translates into better performance on downstream clinically-critical tasks-specifically, estimating S/F oxygenation ratio values, classifying S/F ratio change, and predicting 30-day patient readmission. While we empirically evaluate many methods for exposing the uncertainty to the learning model, we find that a simple approach that trains a model on binarized labels obtained with a (60%) confidence threshold works well. Importantly, high thresholds work far better than a naive approach of a 50% threshold, indicating that training on very confident pixels is far more effective. Our study systematically investigates the impact of training with varying confidence thresholds, comparing not only segmentation metrics but also downstream clinical outcomes. These results suggest that label confidence is a valuable signal that, when properly leveraged, can significantly enhance the reliability and clinical utility of AI in medical imaging.

Machine learning (ML) methods have been successfully employed in identifying variables that can predict the equity premium of individual stocks. In this paper, we investigate if ML can also be helpful in selecting variables relevant for optimal portfolio choice. To address this question, we parameterize minimum-variance portfolio weights as a function of a large pool of firm-level characteristics as well as their second-order and cross-product transformations, yielding a total of 4,610 predictors. We find that the gains from employing ML to select relevant predictors are substantial: minimum-variance portfolios achieve lower risk relative to sparse specifications commonly considered in the literature, especially when non-linear terms are added to the predictor space. Moreover, some of the selected predictors that help decreasing portfolio risk also increase returns, leading to minimum-variance portfolios with good performance in terms of Shape ratios in some situations. Our evidence suggests that ad-hoc sparsity can be detrimental to the performance of minimum-variance characteristics-based portfolios.

We study the problem of robust estimation under heterogeneous corruption rates, where each sample may be independently corrupted with a known but non-identical probability. This setting arises naturally in distributed and federated learning, crowdsourcing, and sensor networks, yet existing robust estimators typically assume uniform or worst-case corruption, ignoring structural heterogeneity. For mean estimation for multivariate bounded distributions and univariate gaussian distributions, we give tight minimax rates for all heterogeneous corruption patterns. For multivariate gaussian mean estimation and linear regression, we establish the minimax rate for squared error up to a factor of $\sqrt{d}$, where $d$ is the dimension. Roughly, our findings suggest that samples beyond a certain corruption threshold may be discarded by the optimal estimators -- this threshold is determined by the empirical distribution of the corruption rates given.

Modest statistical differences between the sampling performances of the D-Wave quantum annealer (QA) and the classical Markov Chain Monte Carlo (MCMC), when applied to Restricted Boltzmann Machines (RBMs), are explored to explain, and possibly address, the absence of significant and consistent improvements in RBM trainability when the D-Wave sampling was used in previous investigations. A novel hybrid sampling approach, combining the classical and the QA contributions, is investigated as a promising way to benefit from the modest differences between the two sampling methods. No improvements in the RBM training are achieved in this work, thereby suggesting that the differences between the QA-based and MCMC sampling, mainly found in the medium-to-low probability regions of the distribution, which are less important for the quality of the sample, are insufficient to benefit the training. Difficulties in achieving sufficiently high quality of embedding RBMs into the lattice of the newer generation of D-Wave hardware could be further complicating the task. On the other hand, the ability to generate samples of sufficient variety from lower-probability parts of the distribution has a potential to benefit other machine learning applications, such as the mitigation of catastrophic forgetting (CF) during incremental learning. The feasibility of using QA-generated patterns of desirable classes for CF mitigation by the generative replay is demonstrated in this work for the first time. While the efficiency of the CF mitigation using the D-Wave QA was comparable to that of the classical mitigation, both the speed of generating a large number of distinct desirable patterns and the potential for further improvement make this approach promising for a variety of challenging machine learning applications.

Model personalization, with broad applications in several fields, including precision medicine and healthcare (Johnson et al., 2021; Hilton et al., 2020; Abul-Husn and Kenny, 2019), advertising (Bilenko and Richardson, 2011), and interactive user inter faces (Ma et al., 2021), involves tailoring models to account for the unique characteristics and featur es of individuals (or subgroups) within a population. A key challenge in achieving model personaliza tion is addressing heterogeneity among individuals while leveraging their similarities. When eac h individual has access to a large amount of data, leveraging similarity is not essential, and one strai ghtforward method is to fit separate models to each individual, allowing for fully individualized m odeling. However, in most applications, including healthcare, access to a large sample size for each individual is difficult and expensive. An alternative approach to fully individualized modeling tra ins one model to fit all by combining data from all individuals. While this approach increases the sam ple size, it overlooks the unique traits of individuals and the variations between them. Therefore, middle-ground methods that can use shared information and capture data heterogeneity are nece ssary. An example of such a trade-off appears in telemonitoring appl ications for chronic disease management, such as remote assessment of Parkinson's disease s everity using smartphone-based sensor 1

Pre-trained machine learning (ML) predictions have been increasingly used to complement incomplete data to enable downstream scientific inquiries, but their naive integration risks biased inferences. Recently, multiple methods have been developed to provide valid inference with ML imputations regardless of prediction quality and to enhance efficiency relative to complete-case analyses. However, existing approaches are often limited to missing outcomes under a missing-completely-at-random (MCAR) assumption, failing to handle general missingness patterns under the more realistic missing-at-random (MAR) assumption. This paper develops a novel method which delivers valid statistical inference framework for general Z-estimation problems using ML imputations under the MAR assumption and for general missingness patterns. The core technical idea is to stratify observations by distinct missingness patterns and construct an estimator by appropriately weighting and aggregating pattern-specific information through a masking-and-imputation procedure on the complete cases. We provide theoretical guarantees of asymptotic normality of the proposed estimator and efficiency dominance over weighted complete-case analyses. Practically, the method affords simple implementations by leveraging existing weighted complete-case analysis software. Extensive simulations are carried out to validate theoretical results. The paper concludes with a brief discussion on practical implications, limitations, and potential future directions.

Standard gradient descent methods yield point estimates with no measure of confidence. This limitation is acute in overparameterized and low-data regimes, where models have many parameters relative to available data and can easily overfit. Bootstrapping is a classical statistical framework for uncertainty estimation based on resampling, but naively applying it to deep learning is impractical: it requires training many replicas, produces post-hoc estimates that cannot guide learning, and implicitly assumes comparable optima across runs - an assumption that fails in non-convex landscapes. We introduce Twin-Bootstrap Gradient Descent (Twin-Boot), a resampling-based training procedure that integrates uncertainty estimation into optimization. Two identical models are trained in parallel on independent bootstrap samples, and a periodic mean-reset keeps both trajectories in the same basin so that their divergence reflects local (within-basin) uncertainty. During training, we use this estimate to sample weights in an adaptive, data-driven way, providing regularization that favors flatter solutions. In deep neural networks and complex high-dimensional inverse problems, the approach improves calibration and generalization and yields interpretable uncertainty maps.

We propose a novel class of prior distributions for sequences of orthogonal functions, which are frequently required in various statistical models such as functional principal component analysis (FPCA). Our approach constructs priors sequentially by imposing adaptive orthogonality constraints through a hierarchical formulation of conditionally normal distributions. The orthogonality is controlled via hyperparameters, allowing for flexible trade-offs between exactness and smoothness, which can be learned from the observed data. We illustrate the properties of the proposed prior and show that it leads to nearly orthogonal posterior estimates. The proposed prior is employed in Bayesian FPCA, providing more interpretable principal functions and efficient low-rank representations. Through simulation studies and analysis of human mobility data in Tokyo, we demonstrate the superior performance of our approach in inducing orthogonality and improving functional component estimation.

Conformal Prediction (CP) has recently received a tremendous amount of interest, leading to a wide range of new theoretical and methodological results for predictive inference with formal theoretical guarantees. However, the vast majority of CP methods assume that all units in the training data have fully observed data on both the outcome and covariates of primary interest, an assumption that rarely holds in practice. In reality, training data are often missing the outcome, a subset of covariates, or both on some units. In addition, time-to-event outcomes in the training set may be censored due to dropout or administrative end-of-follow-up. Accurately accounting for such coarsened data in the training sample while fulfilling the primary objective of well-calibrated conformal predictive inference, requires robustness and efficiency considerations. In this paper, we consider the general problem of obtaining distribution-free valid prediction regions for an outcome given coarsened training data. Leveraging modern semiparametric theory, we achieve our goal by deriving the efficient influence function of the quantile of the outcome we aim to predict, under a given semiparametric model for the coarsened data, carefully combined with a novel conformal risk control procedure. Our principled use of semiparametric theory has the key advantage of facilitating flexible machine learning methods such as random forests to learn the underlying nuisance functions of the semiparametric model. A straightforward application of the proposed general framework produces prediction intervals with stronger coverage properties under covariate shift, as well as the construction of multiply robust prediction sets in monotone missingness scenarios. We further illustrate the performance of our methods through various simulation studies.