Generative Flow Networks (GFlowNets) have emerged as a powe rful paradigm for generating composite structures, demonstrating consi derable promise across diverse applications. While substantial progress has been made in exploring their modeling validity and connections to other generative fram eworks, the theoretical understanding of their learning behavior remains large ly uncharted. In this work, we present a rigorous theoretical investigation of GF lowNets' learning behavior, focusing on four fundamental dimensions: converge nce, sample complexity, implicit regularization, and robustness. By analyzin g these aspects, we seek to elucidate the intricate mechanisms underlying GFlowNet's learning dynamics, shedding light on its strengths and limitations. Our finding s contribute to a deeper understanding of the factors influencing GFlowNet performa nce and provide insights into principled guidelines for their effective desi gn and deployment. This study not only bridges a critical gap in the theoretical land scape of GFlowNets but also lays the foundation for their evolution as a reliable an d interpretable framework for generative modeling. Through this, we aspire to adv ance the theoretical frontiers of GFlowNets and catalyze their broader adoption in the AI community.

Model averaging (MA) and ensembling play a crucial role in statistical and machine learning practice. When multiple candidate models are considered, MA techniques can be used to weight and combine them, often resulting in improved predictive accuracy and better estimation stability compared to model selection (MS) methods. In this paper, we address two challenges in combining least squares estimators from both theoretical and practical perspectives. We first establish several oracle inequalities for least squares MA via minimizing a Mallows' $C_p$ criterion under an arbitrary candidate model set. Compared to existing studies, these oracle inequalities yield faster excess risk and directly imply the asymptotic optimality of the resulting MA estimators under milder conditions. Moreover, we consider candidate model construction and investigate the problem of optimal all-subset combination for least squares estimators, which is an important yet rarely discussed topic in the existing literature. We show that there exists a fundamental limit to achieving the optimal all-subset MA risk. To attain this limit, we propose a novel Mallows-type MA procedure based on a dimension-adaptive $C_p$ criterion. The implicit ensembling effects of several MS procedures are also revealed and discussed. We conduct several numerical experiments to support our theoretical findings and demonstrate the effectiveness of the proposed Mallows-type MA estimator.

Lévy processes are widely used in financial modeling due to their ability to capture discontinuities and heavy tails, which are common in high-frequency asset return data. However, parameter estimation remains a challenge when associated likelihoods are unavailable or costly to compute. We propose a fast and accurate method for Lévy parameter estimation using the neural Bayes estimation (NBE) framework -- a simulation-based, likelihood-free approach that leverages permutation-invariant neural networks to approximate Bayes estimators. Through extensive simulations across several Lévy models, we show that NBE outperforms traditional methods in both accuracy and runtime, while also enabling rapid bootstrap-based uncertainty quantification. We illustrate our approach on a challenging high-frequency cryptocurrency return dataset, where the method captures evolving parameter dynamics and delivers reliable and interpretable inference at a fraction of the computational cost of traditional methods. NBE provides a scalable and practical solution for inference in complex financial models, enabling parameter estimation and uncertainty quantification over an entire year of data in just seconds. We additionally investigate nearly a decade of high-frequency Bitcoin returns, requiring less than one minute to estimate parameters under the proposed approach.

Despite the empirical success of foundation models, we do not have a systematic characterization of the representations that these models learn. In this paper, we establish the contexture theory. It shows that a large class of representation learning methods can be characterized as learning from the association between the input and a context variable. Specifically, we show that many popular methods aim to approximate the top-d singular functions of the expectation operator induced by the context, in which case we say that the representation learns the contexture. We demonstrate the generality of the contexture theory by proving that representation learning within various learning paradigms -- supervised, self-supervised, and manifold learning -- can all be studied from such a perspective. We also prove that the representations that learn the contexture are optimal on those tasks that are compatible with the context. One important implication of the contexture theory is that once the model is large enough to approximate the top singular functions, further scaling up the model size yields diminishing returns. Therefore, scaling is not all we need, and further improvement requires better contexts. To this end, we study how to evaluate the usefulness of a context without knowing the downstream tasks. We propose a metric and show by experiments that it correlates well with the actual performance of the encoder on many real datasets.

Extracting information about hadron resonances requires fitting theoretical models to experimental data. However, this data often comes from different experiments of different physics quantities in varying kinematic regions; studying coupled channels with different kinematic coverages and binning can make direct comparison challenging. The consistency of these datasets directly impacts the quality of the fit, thus making it difficult to accurately constrain the theoretical models. Sparse datasets in key kinematic regions further complicates the quantification of uncertainties, often requiring arbitrary weighting that may introduce bias. A robust approach to solving these problems involves utilising Gaussian Processes (GPs), a Bayesian inference machine learning technique that provides probabilistic predictions for unknown datapoints. Unlike traditional machine learning methods, GPs do not require any training; instead, they operate on three fundamental assumptions: 1. Some kernel function can be defined to measure the covariance between known datapoints; 2. This same kernel function can be used to predict the covariance between unknown datapoints; 3. Some idea of the form of the posterior distribution is known (e.g.

Multidimensional functional data streams arise in diverse scientific fields, yet their analysis poses significant challenges. We propose a novel online framework for functional principal component analysis that enables efficient and scalable modeling of such data. Our method represents functional principal components using tensor product splines, enforcing smoothness and orthonormality through a penalized framework on a Stiefel manifold. An efficient Riemannian stochastic gradient descent algorithm is developed, with extensions inspired by adaptive moment estimation and averaging techniques to accelerate convergence. Additionally, a dynamic tuning strategy for smoothing parameter selection is developed based on a rolling averaged block validation score that adapts to the streaming nature of the data. Extensive simulations and real-world applications demonstrate the flexibility and effectiveness of this framework for analyzing multidimensional functional data.

Real-world data contains aleatoric uncertainty - irreducible noise arising from imperfect measurements or from incomplete knowledge about the data generation process. Mean variance estimation (MVE) networks can learn this type of uncertainty but require ad-hoc regularization strategies to avoid overfitting and are unable to predict epistemic uncertainty (model uncertainty). Conversely, Bayesian neural networks predict epistemic uncertainty but are notoriously difficult to train due to the approximate nature of Bayesian inference. We propose to cooperatively train a variance network with a Bayesian neural network and demonstrate that the resulting model disentangles aleatoric and epistemic uncertainties while improving the mean estimation. We demonstrate the effectiveness and scalability of this method across a diverse range of datasets, including a time-dependent heteroscedastic regression dataset we created where the aleatoric uncertainty is known. The proposed method is straightforward to implement, robust, and adaptable to various model architectures.

Wide & Deep, a simple yet effective learning architecture for recommendation systems developed by Google, has had a significant impact in both academia and industry due to its combination of the memorization ability of generalized linear models and the generalization ability of deep models. Graph convolutional networks (GCNs) remain dominant in node classification tasks; however, recent studies have highlighted issues such as heterophily and expressiveness, which focus on graph structure while seemingly neglecting the potential role of node features. In this paper, we propose a flexible framework GCNIII, which leverages the Wide & Deep architecture and incorporates three techniques: Intersect memory, Initial residual and Identity mapping. We provide comprehensive empirical evidence showing that GCNIII can more effectively balance the trade-off between over-fitting and over-generalization on various semi- and full- supervised tasks. Additionally, we explore the use of large language models (LLMs) for node feature engineering to enhance the performance of GCNIII in cross-domain node classification tasks. Our implementation is available at https://github.com/CYCUCAS/GCNIII.

The goal of this paper is to show how different machine learning tools on the Riemannian manifold $\mathcal{P}_d$ of Symmetric Positive Definite (SPD) matrices can be united under a probabilistic framework. For this, we will need several Gaussian distributions defined on $\mathcal{P}_d$. We will show how popular classifiers on $\mathcal{P}_d$ can be reinterpreted as Bayes Classifiers using these Gaussian distributions. These distributions will also be used for outlier detection and dimension reduction. By showing that those distributions are pervasive in the tools used on $\mathcal{P}_d$, we allow for other machine learning tools to be extended to $\mathcal{P}_d$.

Estimating the causal effect of time-varying treatments on survival outcomes is a challenging task in many domains, particularly in medicine where treatment protocols adapt over time. While recent advances in representation learning have improved causal inference for static treatments, extending these methods to dynamic treatment regimes with survival outcomes remains under-explored. In this paper, we introduce TV-SurvCaus, a novel framework that extends representation balancing techniques to the time-varying treatment setting for survival analysis. We provide theoretical guarantees through (1) a generalized bound for time-varying precision in estimation of heterogeneous effects, (2) variance control via sequential balancing weights, (3) consistency results for dynamic treatment regimes, (4) convergence rates for representation learning with temporal dependencies, and (5) a formal bound on the bias due to treatment-confounder feedback. Our neural architecture incorporates sequence modeling to handle temporal dependencies while balancing time-dependent representations. Through extensive experiments on both synthetic and real-world datasets, we demonstrate that TV-SurvCaus outperforms existing methods in estimating individualized treatment effects with time-varying covariates and treatments. Our framework advances the field of causal inference by enabling more accurate estimation of treatment effects in dynamic, longitudinal settings with survival outcomes.