Wessel Bruinsma ‡ wpb23@cam.ac.uk 1. Introduction A principled approach to prediction tasks is to choose a statistical model that explains the data. The choice of the model class is crucial and has to observe the bias-variance tradeoff, which motivates the need for principled approaches to selecting the best model class from a set of options. Whilst model selection can be done manually by trial and error, the process tends to consume considerable time and resources and be prone to human biases. Bayesian model selection (MacKay, 1992; Rasmussen and Ghahramani, 2001), treats the model class as a random variable and computes its posterior distribution. It offers a built-in complexity regulariser, commonly known as Bayesian Occams razor, which penalises models whose complexity is excessive or too modest.

In the last years, crowdsourcing is transforming the way classification training sets are obtained. Instead of relying on a single expert annotator, crowdsourcing shares the labelling effort among a large number of collaborators. For instance, this is being applied to the data acquired by the laureate Laser Interferometer Gravitational Waves Observatory (LIGO), in order to detect glitches which might hinder the identification of true gravitational-waves. The crowdsourcing scenario poses new challenging difficulties, as it deals with different opinions from a heterogeneous group of annotators with unknown degrees of expertise. Probabilistic methods, such as Gaussian Processes (GP), have proven successful in modeling this setting. However, GPs do not scale well to large data sets, which hampers their broad adoption in real practice (in particular at LIGO). This has led to the recent introduction of deep learning based crowdsourcing methods, which have become the state-of-the-art. However, the accurate uncertainty quantification of GPs has been partially sacrificed. This is an important aspect for astrophysicists in LIGO, since a glitch detection system should provide very accurate probability distributions of its predictions. In this work, we leverage the most popular sparse GP approximation to develop a novel GP based crowdsourcing method that factorizes into mini-batches. This makes it able to cope with previously-prohibitive data sets. The approach, which we refer to as Scalable Variational Gaussian Processes for Crowdsourcing (SVGPCR), brings back GP-based methods to the state-of-the-art, and excels at uncertainty quantification. SVGPCR is shown to outperform deep learning based methods and previous probabilistic approaches when applied to the LIGO data. Moreover, its behavior and main properties are carefully analyzed in a controlled experiment based on the MNIST data set.

V arious problems in Engineering and Statistics require the computation of the likelihood ratio function of two probability densities. In classical approaches the two densities are assumed known or to belong to some known parametric family. In a data-driven version we replace this requirement with the availability of data sampled from the densities of interest. For most well known problems in Detection and Hypothesis testing we develop solutions by providing neural network based estimates of the likelihood ratio or its transformations. This task necessitates the definition of proper optimizations which can be used for the training of the network. The main purpose of this work is to offer a simple and unified methodology for defining such optimization problems with guarantees that the solution is indeed the desired function. Our results are extended to cover estimates for likelihood ratios of conditional densities and estimates for statistics encountered in local approaches. HE likelihood ratio of two probability densities is a function that appears in a variety of problems in Engineering and Statistics. Characteristic examples [1], [2] constitute Hypothesis testing, Signal detection, Sequential hypothesis testing, Sequential detection of changes, etc. Many of these problems also use the likelihood ratio under a transformed form with the most frequent example being the log-likelihood ratio. In all these problems the main assumption is that the corresponding probability densities are available under some functional form. What we aim in this work is to replace this requirement with the availability of data sampled from each of the densities of interest. As we mentioned, the computation of the likelihood ratio function relies on the knowledge of the probability densities which, for the majority of applications, is an unrealistic assumption. One can instead propose parametric families of densities and, with the help of available data, estimate the parameters and form the likelihood ratio function. However, with the advent of Data Science and Deep Learning there is a phenomenal increase in need for processing data coming from images, videos etc. For most of these cases it is very difficult to propose any meaningful parametric family of densities that could reliably describe their statistical behavior. Therefore, these techniques tend to be unsuitable for most of these datasets. If parametric families cannot be employed one can always resort to nonparametric density estimation [3] and then form the likelihood ratio. These approaches are purely data-driven but require two different approximations, namely one for each density.

We propose a probabilistic framework for modelling and exploring the latent structure of relational data. Given feature information for the nodes in a network, the scalable deep generative relational model (SDREM) builds a deep network architecture that can approximate potential nonlinear mappings between nodes' feature information and the nodes' latent representations. Our contribution is two-fold: (1) We incorporate high-order neighbourhood structure information to generate the latent representations at each node, which vary smoothly over the network. (2) Due to the Dirichlet random variable structure of the latent representations, we introduce a novel data augmentation trick which permits efficient Gibbs sampling. The SDREM can be used for large sparse networks as its computational cost scales with the number of positive links. We demonstrate its competitive performance through improved link prediction performance on a range of real-world datasets.

Modern machine learning models are often so complex that they achieve vanishing classification error on the training set. Max-margin linear classifiers are among the simplest classification methods that have zero training error (with linearly separable data). Despite this simplicity, their high-dimensional behavior is not yet completely understood. We assume to be given i.i.d. data $(y_i,{\boldsymbol x}_i)$, $i\le n$ with ${\boldsymbol x}_i\sim {\sf N}({\boldsymbol 0},{\boldsymbol \Sigma})$ a $p$-dimensional Gaussian feature vector, and $y_i \in\{+1,-1\}$ a label whose distribution depends on a linear combination of the covariates $\langle {\boldsymbol \theta}_*,{\boldsymbol x}_i\rangle$. We consider the proportional asymptotics $n,p\to\infty$ with $p/n\to \psi$, and derive exact expressions for the limiting prediction error. Our asymptotic results match simulations already when $n,p$ are of the order of a few hundreds. We explore several choices for the the pair $({\boldsymbol \theta}_*,{\boldsymbol \Sigma})$, and show that the resulting generalization curve (test error error as a function of the overparametrization ratio $\psi=p/n$) is qualitatively different, depending on this choice. In particular we consider a specific structure of $({\boldsymbol \theta}_*,{\boldsymbol \Sigma})$ that captures the behavior of nonlinear random feature models or, equivalently, two-layers neural networks with random first layer weights. In this case, we observe that the test error is monotone decreasing in the number of parameters. This finding agrees with the recently developed `double descent' phenomenology for overparametrized models.

In variational inference, the complicated target is approximated by a closest member relative to the Kullback-Leibler (KL) divergence in a pre-specified family of tractable densities. In many large-scale machine learning applications including clustering problems [11, 32], image classification [25, 27] and topic models [21, 7], variational inference can be orders of magnitude faster than the traditional sampling based approaches such as Markov Chain Monte Carlo (MCMC). In particular, by turning the integration, or sampling, problem into an optimization problem, variational inference can take advantage of modern optimization tools such as stochastic optimization techniques [20, 17] and distributed optimization architecture [1, 8] for further improving its efficiency. Among various approximating schemes, mean-field approximation is the most common type of variational inference that is conceptually simple, implementation-wise easy and particularly suitable for problems involving large numbers of latent variables. The word "mean-field" is originated from the mean-field theory in physics where despite complex interactions among many particles in a many (infinite) body system, all interactions to any one particle can be approximated by a single averaged effect from a "mean-field". In variational inference, by restricting the approximating family of the mean-field to be all density functions that are fully factorized over (blocks of) unknown variables, the associated optimization problem of finding a closest weih2@illinois.edu

Machine learning techniques have been successfully applied to super-resolution tasks on natural images where visually pleasing results are sufficient. However in many scientific domains this is not adequate and estimations of errors and uncertainties are crucial. To address this issue we propose a Bayesian framework that decomposes uncertainties into epistemic and aleatoric uncertainties. We test the validity of our approach by super-resolving images of the Sun's magnetic field and by generating maps measuring the range of possible high resolution explanations compatible with a given low resolution magnetogram.

How secure automatic speaker verification (ASV) technology is? More concretely, given a specific target speaker, how likely is it to find another person who gets falsely accepted as that target? This question may be addressed empirically by studying naturally confusable pairs of speakers within a large enough corpus. To this end, one might expect to find at least some speaker pairs that are indistinguishable from each other in terms of ASV. To a certain extent, such aim is mirrored in the standardized ASV evaluation benchmarks. However, the number of speakers in such evaluation benchmarks represents only a small fraction of all possible human voices, making it challenging to extrapolate performance beyond a given corpus. Furthermore, the impostors used in performance evaluation are usually selected randomly. A potentially more meaningful definition of an impostor - at least in the context of security-driven ASV applications - would be closest (most confusable) other speaker to a given target. We put forward a novel performance assessment framework to address both the inadequacy of the random-impostor evaluation model and the size limitation of evaluation corpora by addressing ASV security against closest impostors on arbitrarily large datasets. The framework allows one to make a prediction of the safety of given ASV technology, in its current state, for arbitrarily large speaker database size consisting of virtual (sampled) speakers. As a proof-of-concept, we analyze the performance of two state-of-the-art ASV systems, based on i-vector and x-vector speaker embeddings (as implemented in the popular Kaldi toolkit), on the recent VoxCeleb 1 & 2 corpora. We found that neither the i-vector or x-vector system is immune to increased false alarm rate at increased impostor database size.

Machine learning algorithms are extensively used to make increasingly more consequential decisions, so that achieving optimal predictive performance can no longer be the only focus. This paper explores intersectional fairness, that is fairness when intersections of multiple sensitive attributes -- such as race, age, nationality, etc. -- are considered. Previous research has mainly been focusing on fairness with respect to a single sensitive attribute, with intersectional fairness being comparatively less studied despite its critical importance for modern machine learning applications. We introduce intersectional fairness metrics by extending prior work, and provide different methodologies to audit discrimination in a given dataset or model outputs. Secondly, we develop novel post-processing techniques to mitigate any detected bias in a classification model. Our proposed methodology does not rely on any assumptions regarding the underlying model and aims at guaranteeing fairness while preserving good predictive performance. Finally, we give guidance on a practical implementation, showing how the proposed methods perform on a real-world dataset.

Monte Carlo methods are a class of techniques for randomly sampling a probability distribution. There are many problem domains where describing or estimating the probability distribution is relatively straightforward, but calculating a desired quantity is intractable. This may be due to many reasons, such as the stochastic nature of the domain or an exponential number of random variables. Instead, a desired quantity can be approximated by using random sampling, referred to as Monte Carlo methods. These methods were initially used around the time that the first computers were created and remain pervasive through all fields of science and engineering, including artificial intelligence and machine learning.