Machine Learning services are being deployed in a large range of applications that make it easy for an adversary, using the algorithm and/or the model, to gain access to sensitive data. This paper investigates fundamental bounds on information leakage. First, we identify and bound the success rate of the worst-case membership inference attack, connecting it to the generalization error of the target model. Second, we study the question of how much sensitive information is stored by the algorithm about the training set and we derive bounds on the mutual information between the sensitive attributes and model parameters. Although our contributions are mostly of theoretical nature, the bounds and involved concepts are of practical relevance. Inspired by our theoretical analysis, we study linear regression and DNN models to illustrate how these bounds can be used to assess the privacy guarantees of ML models.

Hawkes Processes have undergone increasing popularity as default tools for modeling self- and mutually exciting interactions of discrete events in continuous-time event streams. A Maximum Likelihood Estimation (MLE) unconstrained optimization procedure over parametrically assumed forms of the triggering kernels of the corresponding intensity function are a widespread cost-effective modeling strategy, particularly suitable for data with few and/or short sequences. However, the MLE optimization lacks guarantees, except for strong assumptions on the parameters of the triggering kernels, and may lead to instability of the resulting parameters .In the present work, we show how a simple stabilization procedure improves the performance of the MLE optimization without these overly restrictive assumptions.This stabilized version of the MLE is shown to outperform traditional methods over sequences of several different lengths.

This is a course on Machine Learning, Deep Learning and Bayesian Learning (​yes all 3 topics in one place!!!). We start off by analysing data

An information cascade is a circumstance where agents make decisions in a sequential fashion by following other agents. Bikhchandani et al., predict that once a cascade starts it continues, even if it is wrong, until agents receive an external input such as public information. In an information cascade, even if an agent has its own personal choice, it is always overridden by observation of previous agents' actions. This could mean agents end up in a situation where they may act without valuing their own information. As information cascades can have serious social consequences, it is important to have a good understanding of what causes them. We present a detailed Bayesian model of the information gained by agents when observing the choices of other agents and their own private information. Compared to prior work, we remove the high impact of the first observed agent's action by incorporating a prior probability distribution over the information of unobserved agents and investigate an alternative model of choice to that considered in prior work: weighted random choice. Our results show that, in contrast to Bikhchandani's results, cascades will not necessarily occur and adding prior agents' information will delay the effects of cascades.

For Bayesian learning, given likelihood function and Gaussian prior, the elliptical slice sampler, introduced by Murray, Adams and MacKay 2010, provides a tool for the construction of a Markov chain for approximate sampling of the underlying posterior distribution. Besides of its wide applicability and simplicity its main feature is that no tuning is necessary. Under weak regularity assumptions on the posterior density we show that the corresponding Markov chain is geometrically ergodic and therefore yield qualitative convergence guarantees. We illustrate our result for Gaussian posteriors as they appear in Gaussian process regression, as well as in a setting of a multi-modal distribution. Remarkably, our numerical experiments indicate a dimension-independent performance of elliptical slice sampling even in situations where our ergodicity result does not apply.

There is a range of applications where the interest is in understanding interactions between a set of time series, including in neuroscience, genomics, econometrics, climate science, and social media analysis. For example, in neuroscience, one may seek to understand whether activity in one brain region correlates with later activity in another region, or to decipher instantaneous correlations between regions--both notions of functional connectivity. In genomics, there is an analogous study of gene regulatory networks. In econometrics, one may be interested in how various macroeconomic indicators predict one another. We also have unprecedented levels of data on people's actions--whether they be social media posts, purchase histories, or political voting records--and want to understand the dependencies between the actions of these individuals. Modern recording modalities and the ability to store and process large amounts of data have escalated the scale at which we seek to do such analyses. In many cases, one may seek notions of causal interactions amongst the time series, but be limited to drawing inferences from observational data without opportunities for experimentation and without known mechanistic models for the observed phenomena.

This paper produces an efficient Semidefinite Programming (SDP) solution for community detection that incorporates non-graph data, which in this context is known as side information. SDP is an efficient solution for standard community detection on graphs. We formulate a semi-definite relaxation for the maximum likelihood estimation of node labels, subject to observing both graph and non-graph data. This formulation is distinct from the SDP solution of standard community detection, but maintains its desirable properties. We calculate the exact recovery threshold for three types of non-graph information, which in this paper are called side information: partially revealed labels, noisy labels, as well as multiple observations (features) per node with arbitrary but finite cardinality. We find that SDP has the same exact recovery threshold in the presence of side information as maximum likelihood with side information. Thus, the methods developed herein are computationally efficient as well as asymptotically accurate for the solution of community detection in the presence of side information. Simulations show that the asymptotic results of this paper can also shed light on the performance of SDP for graphs of modest size.

We develop simple methods for constructing parameter priors for model choice among Directed Acyclic Graphical (DAG) models. In particular, we introduce several assumptions that permit the construction of parameter priors for a large number of DAG models from a small set of assessments. We then present a method for directly computing the marginal likelihood of every DAG model given a random sample with no missing observations. We apply this methodology to Gaussian DAG models which consist of a recursive set of linear regression models. We show that the only parameter prior for complete Gaussian DAG models that satisfies our assumptions is the normal-Wishart distribution. Our analysis is based on the following new characterization of the Wishart distribution: let $W$ be an $n \times n$, $n \ge 3$, positive-definite symmetric matrix of random variables and $f(W)$ be a pdf of $W$. Then, f$(W)$ is a Wishart distribution if and only if $W_{11} - W_{12} W_{22}^{-1} W'_{12}$ is independent of $\{W_{12},W_{22}\}$ for every block partitioning $W_{11},W_{12}, W'_{12}, W_{22}$ of $W$. Similar characterizations of the normal and normal-Wishart distributions are provided as well.

AMP algorithms have two features that make them particularly attractive. First, they can easily be tailored to take advantage of prior information on the structure of the signal, such as sparsity or other constraints. Second, under suitable assumptions on a design or data matrix, AMP theory provides precise asymptotic guarantees for statistical procedures in the high-dimensional regime where the ratio of the number of observations n to dimensions p converges to a constant (Bayati and Montanari, 2012; Donoho et al., 2013; Sur et al., 2017). More generally, AMP has been also used to obtain lower bounds on the estimation error of first-order methods (Celentano et al., 2020), and in linear regression and low rank matrix estimation, it plays a fundamental role in understanding the performance gap between information-theoretically optimal and computationally feasible estimators (Reeves and Pfister, 2019; Barbier et al., 2019; Lelarge and Miolane, 2019). In these settings, it is conjectured that AMP achieves the optimal asymptotic estimation error among all polynomial-time algorithms (cf.

Energy-based models (EBMs) are a simple yet powerful framework for generative modeling. They are based on a trainable energy function which defines an associated Gibbs measure, and they can be trained and sampled from via well-established statistical tools, such as MCMC. Neural networks may be used as energy function approximators, providing both a rich class of expressive models as well as a flexible device to incorporate data structure. In this work we focus on shallow neural networks. Building from the incipient theory of overparametrized neural networks, we show that models trained in the so-called "active" regime provide a statistical advantage over their associated "lazy" or kernel regime, leading to improved adaptivity to hidden low-dimensional structure in the data distribution, as already observed in supervised learning. Our study covers both maximum likelihood and Stein Discrepancy estimators, and we validate our theoretical results with numerical experiments on synthetic data.