This paper shows that dropout training in Generalized Linear Models is the minimax solution of a two-player, zero-sum game where an adversarial nature corrupts a statistician's covariates using a multiplicative nonparametric errors-in-variables model. In this game---known as a Distributionally Robust Optimization problem---nature's least favorable distribution is dropout noise, where nature independently deletes entries of the covariate vector with some fixed probability $\delta$. Our decision-theoretic analysis shows that dropout training---the statistician's minimax strategy in the game---indeed provides out-of-sample expected loss guarantees for distributions that arise from multiplicative perturbations of in-sample data. This paper also provides a novel, parallelizable, Unbiased Multi-Level Monte Carlo algorithm to speed-up the implementation of dropout training. Our algorithm has a much smaller computational cost compared to the naive implementation of dropout, provided the number of data points is much smaller than the dimension of the covariate vector.

The induction of additional randomness in parallel and sequential ensemble methods has proven to be worthwhile in many aspects. In this manuscript, we propose and examine a novel random tree depth injection approach suitable for sequential and parallel tree-based approaches including Boosting and Random Forests. The resulting methods are called \emph{Random Boost} and \emph{Random$^2$ Forest}. Both approaches serve as valuable extensions to the existing literature on the gradient boosting framework and random forests. A Monte Carlo simulation, in which tree-shaped data sets with different numbers of final partitions are built, suggests that there are several scenarios where \emph{Random Boost} and \emph{Random$^2$ Forest} can improve the prediction performance of conventional hierarchical boosting and random forest approaches. The new algorithms appear to be especially successful in cases where there are merely a few high-order interactions in the generated data. In addition, our simulations suggest that our random tree depth injection approach can improve computation time by up to 40%, while at the same time the performance losses in terms of prediction accuracy turn out to be minor or even negligible in most cases.

Federated learning (FL) is a distributed learning process where the model (weights and checkpoints) is transferred to the devices that posses data rather than the classical way of transferring and aggregating the data centrally. In this way, sensitive data does not leave the user devices. FL uses the FedAvg algorithm, which is trained in the iterative model averaging way, on the non-iid and unbalanced distributed data, without depending on the data quantity. Some issues with the FL are, 1) no scalability, as the model is iteratively trained over all the devices, which amplifies with device drops; 2) security and privacy trade-off of the learning process still not robust enough and 3) overall communication efficiency and the cost are higher. To mitigate these challenges we present Federated Learning and Privately Scaling (FLaPS) architecture, which improves scalability as well as the security and privacy of the system. The devices are grouped into clusters which further gives better privacy scaled turn around time to finish a round of training. Therefore, even if a device gets dropped in the middle of training, the whole process can be started again after a definite amount of time. The data and model both are communicated using differentially private reports with iterative shuffling which provides a better privacy-utility trade-off. We evaluated FLaPS on MNIST, CIFAR10, and TINY-IMAGENET-200 dataset using various CNN models. Experimental results prove FLaPS to be an improved, time and privacy scaled environment having better and comparable after-learning-parameters with respect to the central and FL models.

Finite mixture models, typically Gaussian mixtures, are well known and widely used as model-based clustering. In practical situations, there are many non-Gaussian data that are heavy-tailed and/or asymmetric. Normal inverse Gaussian (NIG) distributions are normal-variance mean which mixing densities are inverse Gaussian distributions and can be used for both haavy-tail and asymmetry. For NIG mixture models, both expectation-maximization method and variational Bayesian (VB) algorithms have been proposed. However, the existing VB algorithm for NIG mixture have a disadvantage that the shape of the mixing density is limited. In this paper, we propose another VB algorithm for NIG mixture that improves on the shortcomings. We also propose an extension of Dirichlet process mixture models to overcome the difficulty in determining the number of clusters in finite mixture models. We evaluated the performance with artificial data and found that it outperformed Gaussian mixtures and existing implementations for NIG mixtures, especially for highly non-normative data.

Machine Learning in general and Deep Learning in particular has gained much interest in the recent decade and has shown significant performance improvements for many Computer Vision or Natural Language Processing tasks. In order to deal with databases which have just a small amount of training samples or to deal with models which have large amount of parameters, the regularization is indispensable. In this paper, we enforce the manifold preservation (manifold learning) from the original data into latent presentation by using "manifold attack". The later is inspired in a fashion of adversarial learning : finding virtual points that distort mostly the manifold preservation then using these points as supplementary samples to train the model. We show that our approach of regularization provides improvements for the accuracy rate and for the robustness to adversarial examples.

Graph classification is a widely studied problem and has broad applications. In many real-world problems, the number of labeled graphs available for training classification models is limited, which renders these models prone to overfitting. To address this problem, we propose two approaches based on contrastive self-supervised learning (CSSL) to alleviate overfitting. In the first approach, we use CSSL to pretrain graph encoders on widely-available unlabeled graphs without relying on human-provided labels, then finetune the pretrained encoders on labeled graphs. In the second approach, we develop a regularizer based on CSSL, and solve the supervised classification task and the unsupervised CSSL task simultaneously. To perform CSSL on graphs, given a collection of original graphs, we perform data augmentation to create augmented graphs out of the original graphs. An augmented graph is created by consecutively applying a sequence of graph alteration operations. A contrastive loss is defined to learn graph encoders by judging whether two augmented graphs are from the same original graph. Experiments on various graph classification datasets demonstrate the effectiveness of our proposed methods.

The Wasserstein distance provides a notion of dissimilarities between probability measures, which has recent applications in learning of structured data with varying size such as images and text documents. In this work, we analyze the $k$-nearest neighbor classifier ($k$-NN) under the Wasserstein distance and establish the universal consistency on families of distributions. Using previous known results on the consistency of the $k$-NN classifier on infinite dimensional metric spaces, it suffices to show that the families is a countable union of finite dimensional components. As a result, we are able to prove universal consistency of $k$-NN on spaces of finitely supported measures, the space of finite wavelet series and the spaces of Gaussian measures with commuting covariance matrices.

Successful application of machine learning models to real-world prediction problems, e.g. financial forecasting and personalized medicine, has proved to be challenging, because such settings require limiting and quantifying the uncertainty in the model predictions, i.e. providing valid and accurate prediction intervals. Conformal Prediction is a distribution-free approach to construct valid prediction intervals in finite samples. However, the prediction intervals constructed by Conformal Prediction are often (because of over-fitting, inappropriate measures of nonconformity, or other issues) overly conservative and hence inadequate for the application(s) at hand. This paper proposes an AutoML framework called Automatic Machine Learning for Conformal Prediction (AutoCP). Unlike the familiar AutoML frameworks that attempt to select the best prediction model, AutoCP constructs prediction intervals that achieve the user-specified target coverage rate while optimizing the interval length to be accurate and less conservative. We tested AutoCP on a variety of datasets and found that it significantly outperforms benchmark algorithms.

The research on and application of artificial intelligence (AI) has triggered a comprehensive scientific, economic, social and political discussion. Here we argue that statistics, as an interdisciplinary scientific field, plays a substantial role both for the theoretical and practical understanding of AI and for its future development. Statistics might even be considered a core element of AI. With its specialist knowledge of data evaluation, starting with the precise formulation of the research question and passing through a study design stage on to analysis and interpretation of the results, statistics is a natural partner for other disciplines in teaching, research and practice. This paper aims at contributing to the current discussion by highlighting the relevance of statistical methodology in the context of AI development. In particular, we discuss contributions of statistics to the field of artificial intelligence concerning methodological development, planning and design of studies, assessment of data quality and data collection, differentiation of causality and associations and assessment of uncertainty in results. Moreover, the paper also deals with the equally necessary and meaningful extension of curricula in schools and universities.

We derive a new margin-based regularization formulation, termed multi-margin regularization (MMR), for deep neural networks (DNNs). The MMR is inspired by principles that were applied in margin analysis of shallow linear classifiers, e.g., support vector machine (SVM). Unlike SVM, MMR is continuously scaled by the radius of the bounding sphere (i.e., the maximal norm of the feature vector in the data), which is constantly changing during training. We empirically demonstrate that by a simple supplement to the loss function, our method achieves better results on various classification tasks across domains. Using the same concept, we also derive a selective sampling scheme and demonstrate accelerated training of DNNs by selecting samples according to a minimal margin score (MMS). This score measures the minimal amount of displacement an input should undergo until its predicted classification is switched. We evaluate our proposed methods on three image classification tasks and six language text classification tasks. Specifically, we show improved empirical results on CIFAR10, CIFAR100 and ImageNet using state-of-the-art convolutional neural networks (CNNs) and BERT-BASE architecture for the MNLI, QQP, QNLI, MRPC, SST-2 and RTE benchmarks.