It is widely perceived that leveraging the success of modern machine learning techniques to mobile devices and wireless networks has the potential of enabling important new services. This, however, poses significant challenges, essentially due to that both data and processing power are highly distributed in a wireless network. In this paper, we develop a learning algorithm and an architecture that make use of multiple data streams and processing units, not only during the training phase but also during the inference phase. In particular, the analysis reveals how inference propagates and fuses across a network. We study the design criterion of our proposed method and its bandwidth requirements. Also, we discuss implementation aspects using neural networks in typical wireless radio access; and provide experiments that illustrate benefits over state-of-the-art techniques.

Due to their increasing spread, confidence in neural network predictions became more and more important. However, basic neural networks do not deliver certainty estimates or suffer from over or under confidence. Many researchers have been working on understanding and quantifying uncertainty in a neural network's prediction. As a result, different types and sources of uncertainty have been identified and a variety of approaches to measure and quantify uncertainty in neural networks have been proposed. This work gives a comprehensive overview of uncertainty estimation in neural networks, reviews recent advances in the field, highlights current challenges, and identifies potential research opportunities. It is intended to give anyone interested in uncertainty estimation in neural networks a broad overview and introduction, without presupposing prior knowledge in this field. A comprehensive introduction to the most crucial sources of uncertainty is given and their separation into reducible model uncertainty and not reducible data uncertainty is presented. The modeling of these uncertainties based on deterministic neural networks, Bayesian neural networks, ensemble of neural networks, and test-time data augmentation approaches is introduced and different branches of these fields as well as the latest developments are discussed. For a practical application, we discuss different measures of uncertainty, approaches for the calibration of neural networks and give an overview of existing baselines and implementations. Different examples from the wide spectrum of challenges in different fields give an idea of the needs and challenges regarding uncertainties in practical applications. Additionally, the practical limitations of current methods for mission- and safety-critical real world applications are discussed and an outlook on the next steps towards a broader usage of such methods is given.

Optimization is often cast as a deterministic problem, where the solution is found through some iterative procedure such as gradient descent. However, when training neural networks the loss function changes over (iteration) time due to the randomized selection of a subset of the samples. This randomization turns the optimization problem into a stochastic one. We propose to consider the loss as a noisy observation with respect to some reference optimum. This interpretation of the loss allows us to adopt Kalman filtering as an optimizer, as its recursive formulation is designed to estimate unknown parameters from noisy measurements. Moreover, we show that the Kalman Filter dynamical model for the evolution of the unknown parameters can be used to capture the gradient dynamics of advanced methods such as Momentum and Adam. We call this stochastic optimization method KaFiStO. KaFiStO is an easy to implement, scalable, and efficient method to train neural networks. We show that it also yields parameter estimates that are on par with or better than existing optimization algorithms across several neural network architectures and machine learning tasks, such as computer vision and language modeling.

Partition of unity networks (POU-Nets) have been shown capable of realizing algebraic convergence rates for regression and solution of PDEs, but require empirical tuning of training parameters. We enrich POU-Nets with a Gaussian noise model to obtain a probabilistic generalization amenable to gradient-based minimization of a maximum likelihood loss. The resulting architecture provides spatial representations of both noiseless and noisy data as Gaussian mixtures with closed form expressions for variance which provides an estimator of local error. The training process yields remarkably sharp partitions of input space based upon correlation of function values. This classification of training points is amenable to a hierarchical refinement strategy that significantly improves the localization of the regression, allowing for higher-order polynomial approximation to be utilized. The framework scales more favorably to large data sets as compared to Gaussian process regression and allows for spatially varying uncertainty, leveraging the expressive power of deep neural networks while bypassing expensive training associated with other probabilistic deep learning methods. Compared to standard deep neural networks, the framework demonstrates hp-convergence without the use of regularizers to tune the localization of partitions. We provide benchmarks quantifying performance in high/low-dimensions, demonstrating that convergence rates depend only on the latent dimension of data within high-dimensional space. Finally, we introduce a new open-source data set of PDE-based simulations of a semiconductor device and perform unsupervised extraction of a physically interpretable reduced-order basis.

This paper develops simple feed-forward neural networks that achieve the universal approximation property for all continuous functions with a fixed finite number of neurons. These neural networks are simple because they are designed with a simple and computable continuous activation function $\sigma$ leveraging a triangular-wave function and a softsign function. We prove that $\sigma$-activated networks with width $36d(2d+1)$ and depth $11$ can approximate any continuous function on a $d$-dimensioanl hypercube within an arbitrarily small error. Hence, for supervised learning and its related regression problems, the hypothesis space generated by these networks with a size not smaller than $36d(2d+1)\times 11$ is dense in the space of continuous functions. Furthermore, classification functions arising from image and signal classification are in the hypothesis space generated by $\sigma$-activated networks with width $36d(2d+1)$ and depth $12$, when there exist pairwise disjoint closed bounded subsets of $\mathbb{R}^d$ such that the samples of the same class are located in the same subset.

Deep learning models can enable accurate and efficient disease diagnosis, but have thus far been hampered by the data scarcity present in the medical world. Automated diagnosis studies have been constrained by underpowered single-center datasets, and although some results have shown promise, their generalizability to other institutions remains questionable as the data heterogeneity between institutions is not taken into account. By allowing models to be trained in a distributed manner that preserves patients' privacy, federated learning promises to alleviate these issues, by enabling diligent multi-center studies. We present the first federated learning study on the modality of cardiovascular magnetic resonance (CMR) and use four centers derived from subsets of the M\&M and ACDC datasets, focusing on the diagnosis of hypertrophic cardiomyopathy (HCM). We adapt a 3D-CNN network pretrained on action recognition and explore two different ways of incorporating shape prior information to the model, and four different data augmentation set-ups, systematically analyzing their impact on the different collaborative learning choices. We show that despite the small size of data (180 subjects derived from four centers), the privacy preserving federated learning achieves promising results that are competitive with traditional centralized learning. We further find that federatively trained models exhibit increased robustness and are more sensitive to domain shift effects.

This work investigates the problem of learning temporal interaction networks. A temporal interaction network consists of a series of chronological interactions between users and items. Previous methods tackle this problem by using different variants of recurrent neural networks to model sequential interactions, which fail to consider the structural information of temporal interaction networks and inevitably lead to sub-optimal results. To this end, we propose a novel Deep Structural Point Process termed as DSPP for learning temporal interaction networks. DSPP simultaneously incorporates the topological structure and long-range dependency structure into our intensity function to enhance model expressiveness. To be specific, by using the topological structure as a strong prior, we first design a topological fusion encoder to obtain node embeddings. An attentive shift encoder is then developed to learn the long-range dependency structure between users and items in continuous time. The proposed two modules enable our model to capture the user-item correlation and dynamic influence in temporal interaction networks. DSPP is evaluated on three real-world datasets for both tasks of item prediction and time prediction. Extensive experiments demonstrate that our model achieves consistent and significant improvements over state-of-the-art baselines.

Two-sided matching markets, such as Uber, Airbnb, stock markets, and dating apps, play a significant role in today's world. As a result, there is a tremendous and rising interest to design better mechanisms for two-sided matching. The seminal work of Gale and Shapley [14] introduced a simple mechanism for stable matching in two-sided markets--Deferred-acceptance (DA)--which has since has been applied in doctor-hospital matching [25], school choice [3, 22, 2], and the matching of cadets to their branches of military service [30, 29]. DA is stable, i.e., no pair of agents mutually prefer each other to their DA partners. On the other hand, DA is not strategy-proof (SP); that is, under fully general preferences, it is always possible that some agent can mis-report her preferences to obtain a better matching than she would receive under the DA mechanism.

The success of neural network embeddings has entailed a renewed interest in using knowledge graphs for a wide variety of machine learning and information retrieval tasks. In particular, recent recommendation methods based on graph embeddings have shown state-of-the-art performance. In general, these methods encode latent rating patterns and content features. Differently from previous work, in this paper, we propose to exploit embeddings extracted from graphs that combine information from ratings and aspect-based opinions expressed in textual reviews. We then adapt and evaluate state-of-the-art graph embedding techniques over graphs generated from Amazon and Yelp reviews on six domains, outperforming baseline recommenders. Additionally, our method has the advantage of providing explanations that involve the coverage of aspect-based opinions given by users about recommended items.

Attributed event sequences are commonly encountered in practice. A recent research line focuses on incorporating neural networks with the statistical model -- marked point processes, which is the conventional tool for dealing with attributed event sequences. Neural marked point processes possess good interpretability of probabilistic models as well as the representational power of neural networks. However, we find that performance of neural marked point processes is not always increasing as the network architecture becomes more complicated and larger, which is what we call the performance saturation phenomenon. This is due to the fact that the generalization error of neural marked point processes is determined by both the network representational ability and the model specification at the same time. Therefore we can draw two major conclusions: first, simple network structures can perform no worse than complicated ones for some cases; second, using a proper probabilistic assumption is as equally, if not more, important as improving the complexity of the network. Based on this observation, we propose a simple graph-based network structure called GCHP, which utilizes only graph convolutional layers, thus it can be easily accelerated by the parallel mechanism. We directly consider the distribution of interarrival times instead of imposing a specific assumption on the conditional intensity function, and propose to use a likelihood ratio loss with a moment matching mechanism for optimization and model selection. Experimental results show that GCHP can significantly reduce training time and the likelihood ratio loss with interarrival time probability assumptions can greatly improve the model performance.