The success of deep neural networks largely depends on the st atistical structure of the training data. While learning dynamics and generalization on iso tropic data are well-established, the impact of pronounced anisotropy on these crucial aspect s is not yet fully understood. We examine the impact of data anisotropy, represented by a sp iked covariance structure, a canonical yet tractable model, on the learning dynamics and generalization error of a two-layer linear network in a linear regression setting. Our ana lysis reveals that the learning dynamics proceed in two distinct phases, governed initiall y by the input-output correlation and subsequently by other principal directions of the data s tructure. Furthermore, we derive an analytical expression for the generalization error, quantifying how the alignment of the spike structure of the data with the learning task improv es performance. Our findings offer deep theoretical insights into how data anisotropy sha pes the learning trajectory and final performance, providing a foundation for understandin g complex interactions in more advanced network architectures.

We study the multinomial logit (MNL) contextual bandit problem for sequential assortment selection. Although most existing research assumes utility functions to be linear in item features, this linearity assumption restricts the modeling of intricate interactions between items and user preferences. A recent work (Zhang & Luo, 2024) has investigated general utility function classes, yet its method faces fundamental trade-offs between computational tractability and statistical efficiency. To address this limitation, we propose a computationally efficient algorithm for MNL contextual bandits leveraging the upper confidence bound principle, specifically designed for non-linear parametric utility functions, including those modeled by neural networks. Under a realizability assumption and a mild geometric condition on the utility function class, our algorithm achieves a regret bound of $\tilde{O}(\sqrt{T})$, where $T$ denotes the total number of rounds. Our result establishes that sharp $\tilde{O}(\sqrt{T})$-regret is attainable even with neural network-based utilities, without relying on strong assumptions such as neural tangent kernel approximations. To the best of our knowledge, our proposed method is the first computationally tractable algorithm for MNL contextual bandits with non-linear utilities that provably attains $\tilde{O}(\sqrt{T})$ regret. Comprehensive numerical experiments validate the effectiveness of our approach, showing robust performance not only in realizable settings but also in scenarios with model misspecification.

Understanding the structure of complex, nonstationary, high-dimensional time-evolving signals is a central challenge in scientific data analysis. In many domains, such as speech and biomedical signal processing, the ability to learn disentangled and interpretable representations is critical for uncovering latent generative mechanisms. Traditional approaches to unsupervised representation learning, including variational autoencoders (VAEs), often struggle to capture the temporal and spectral diversity inherent in such data. Here we introduce variational decomposition autoencoding (VDA), a framework that extends VAEs by incorporating a strong structural bias toward signal decomposition. VDA is instantiated through variational decomposition autoencoders (DecVAEs), i.e., encoder-only neural networks that combine a signal decomposition model, a contrastive self-supervised task, and variational prior approximation to learn multiple latent subspaces aligned with time-frequency characteristics. We demonstrate the effectiveness of DecVAEs on simulated data and three publicly available scientific datasets, spanning speech recognition, dysarthria severity evaluation, and emotional speech classification. Our results demonstrate that DecVAEs surpass state-of-the-art VAE-based methods in terms of disentanglement quality, generalization across tasks, and the interpretability of latent encodings. These findings suggest that decomposition-aware architectures can serve as robust tools for extracting structured representations from dynamic signals, with potential applications in clinical diagnostics, human-computer interaction, and adaptive neurotechnologies.

Individualized treatment regimes (ITRs) aim to improve clinical outcomes by assigning treatment based on patient-specific characteristics. However, existing methods often struggle with high-dimensional covariates, limiting accuracy, interpretability, and real-world applicability. We propose a novel sufficient dimension reduction approach that directly targets the contrast between potential outcomes and identifies a low-dimensional subspace of the covariates capturing treatment effect heterogeneity. This reduced representation enables more accurate estimation of optimal ITRs through outcome-weighted learning. To accommodate observational data, our method incorporates kernel-based covariate balancing, allowing treatment assignment to depend on the full covariate set and avoiding the restrictive assumption that the subspace sufficient for modeling heterogeneous treatment effects is also sufficient for confounding adjustment. We show that the proposed method achieves universal consistency, i.e., its risk converges to the Bayes risk, under mild regularity conditions. We demonstrate its finite sample performance through simulations and an analysis of intensive care unit sepsis patient data to determine who should receive transthoracic echocardiography.

A central problem in machine learning theory is to characterize how learning dynamics select particular solutions among the many compatible with the training objective, a phenomenon, called implicit bias, which remains only partially characterized. In the present work, we identify a general mechanism, in terms of an explicit geometric correction of the learning dynamics, for the emergence of implicit biases, arising from the interaction between continuous symmetries in the model's parametrization and stochasticity in the optimization process. Our viewpoint is constructive in two complementary directions: given model symmetries, one can derive the implicit bias they induce; conversely, one can inverse-design a wide class of different implicit biases by computing specific redundant parameterizations. More precisely, we show that, when the dynamics is expressed in the quotient space obtained by factoring out the symmetry group of the parameterization, the resulting stochastic differential equation gains a closed form geometric correction in the stationary distribution of the optimizer dynamics favoring orbits with small local volume. We compute the resulting symmetry induced bias for a range of architectures, showing how several well known results fit into a single unified framework. The approach also provides a practical methodology for deriving implicit biases in new settings, and it yields concrete, testable predictions that we confirm by numerical simulations on toy models trained on synthetic data, leaving more complex scenarios for future work. Finally, we test the implicit bias inverse-design procedure in notable cases, including biases toward sparsity in linear features or in spectral properties of the model parameters.

Gradient descent on log-sum-exp (LSE) objectives performs implicit expectation--maximization (EM): the gradient with respect to each component output equals its responsibility. The same theory predicts collapse without volume control analogous to the log-determinant in Gaussian mixture models. We instantiate the theory in a single-layer encoder with an LSE objective and InfoMax regularization for volume control. Experiments confirm the theory's predictions. The gradient--responsibility identity holds exactly; LSE alone collapses; variance prevents dead components; decorrelation prevents redundancy. The model exhibits EM-like optimization dynamics in which lower loss does not correspond to better features and adaptive optimizers offer no advantage. The resulting decoder-free model learns interpretable mixture components, confirming that implicit EM theory can prescribe architectures.

Applying Physics-Informed Gaussian Process Regression to the eigenvalue problem $(\mathcal{L}-λ)u = 0$ poses a fundamental challenge, where the null source term results in a trivial predictive mean and a degenerate marginal likelihood. Drawing inspiration from system identification, we construct a transfer function-type indicator for the unknown eigenvalue/eigenfunction using the physics-informed Gaussian Process posterior. We demonstrate that the posterior covariance is only non-trivial when $λ$ corresponds to an eigenvalue of the partial differential operator $\mathcal{L}$, reflecting the existence of a non-trivial eigenspace, and any sample from the posterior lies in the eigenspace of the linear operator. We demonstrate the effectiveness of the proposed approach through several numerical examples with both linear and non-linear eigenvalue problems.

Driven by the right to be forgotten (RTBF), machine unlearning has become an essential requirement for privacy-preserving machine learning. However, its realization in decentralized federated learning (DFL) remains largely unexplored. In DFL, clients exchange local updates only with neighbors, causing model information to propagate and mix across the network. As a result, when a client requests data deletion, its influence is implicitly embedded throughout the system, making removal difficult without centralized coordination. We propose a novel certified unlearning framework for DFL based on Newton-style updates. Our approach first quantifies how a client's data influence propagates during training. Leveraging curvature information of the loss with respect to the target data, we then construct corrective updates using Newton-style approximations. To ensure scalability, we approximate second-order information via Fisher information matrices. The resulting updates are perturbed with calibrated noise and broadcast through the network to eliminate residual influence across clients. We theoretically prove that our approach satisfies the formal definition of certified unlearning, ensuring that the unlearned model is difficult to distinguish from a retrained model without the deleted data. We also establish utility bounds showing that the unlearned model remains close to retraining from scratch. Extensive experiments across diverse decentralized settings demonstrate the effectiveness and efficiency of our framework.

Cardiovascular disease (CVD) continues to be the major cause of death globally, calling for predictive models that not only handle diverse and high-dimensional biomedical signals but also maintain interpretability and privacy. We create a single multimodal learning framework that integrates cross modal transformers with graph neural networks and causal representation learning to measure personalized CVD risk. The model combines genomic variation, cardiac MRI, ECG waveforms, wearable streams, and structured EHR data to predict risk while also implementing causal invariance constraints across different clinical subpopulations. To maintain transparency, we employ SHAP based feature attribution, counterfactual explanations and causal latent alignment for understandable risk factors. Besides, we position the design in a federated, privacy, preserving optimization protocol and establish rules for convergence, calibration and uncertainty quantification under distributional shift. Experimental studies based on large-scale biobank and multi institutional datasets reveal state discrimination and robustness, exhibiting fair performance across demographic strata and clinically distinct cohorts. This study paves the way for a principled approach to clinically trustworthy, interpretable and privacy respecting CVD prediction at the population level.

Calibration Bands for Mean Estimates within the Exponential Dispersion Family null Lukasz Delong Selim Gatti Mario V. W uthrich Version of October 8, 2025 Abstract A statistical model is said to be calibrated if the resulting mean estimates perfectly match the true means of the underlying responses. Aiming for calibration is often not achievable in practice as one has to deal with finite samples of noisy observations. A weaker notion of calibration is auto-calibration. An auto-calibrated model satisfies that the expected value of the responses for a given mean estimate matches this estimate. Testing for auto-calibration has only been considered recently in the literature and we propose a new approach based on calibration bands. Calibration bands denote a set of lower and upper bounds such that the probability that the true means lie simultaneously inside those bounds exceeds some given confidence level. Such bands were constructed by Yang-Barber (2019) for sub-Gaussian distributions. Dimitriadis et al. (2023) then introduced narrower bands for the Bernoulli distribution. We use the same idea in order to extend the construction to the entire exponential dispersion family that contains for example the binomial, Poisson, negative binomial, gamma and normal distributions. Moreover, we show that the obtained calibration bands allow us to construct various tests for calibration and auto-calibration, respectively. As the construction of the bands does not rely on asymptotic results, we emphasize that our tests can be used for any sample size. Auto-calibration, calibration, calibration bands, exponential dispersion family, mean estimation, regression modeling, binomial distribution, Poisson distribution, negative binomial distribution, gamma distribution, normal distribution inverse Gaussian distribution. 1 Introduction Various statistical methods can be used to derive mean estimates from available observations, and it is important to understand whether these mean estimates are reliable for decision making. A statistical model is said to be calibrated if the resulting mean estimates perfectly match the true means of the underlying responses. In practice, calibration is often not achievable, as estimates are obtained from finite samples of noisy observations.