Sam Power's paper, which is a joint work with collaborators Jeremias Knoblauch (UCL), Chris Oates (Newcastle), and Zheyang Shen (Newcastle), is titled “Prediction-Centric Uncertainty Quantification via MMD”. It focuses on how to make informed predictions on the basis of statistical models which may be mis-specified. For example, one might choose to model some physical system as evolving according to an ordinary differential equation (ODE), even though the reality might be more accurately described by a noisy, stochastic differential equation (SDE). In these settings, even though the ODE leaves something to be desired in terms of capturing the underlying physics of the model, it might still serve as a useful tool for making useful predictions about the future behaviour of the system.
Read the paper "Prediction-Centric Uncertainty Quantification via MMD" for free online.
Rahil Morjaria and collaborators, including Professor Sidarth Jaggi, authored the paper "Density-Dependent Group Testing". In the information-theoretic field of Group Testing, in which it is attempted to discern a subset of items with a trait identifiable by a binary test (which allows groups of items to be in one test), this paper answers a natural open problem: What if the probability of a positive test outcome is a function of the density of the trait in the test? For example, a single diseased individual in a test with 100 healthy people could be harder to detect than if they were in a test with 10 healthy people, or more formally for a binary testing apparatus modelled by some function $f$ the probability of a positive test in the first case would be $f(1/101)$ while in the other case it would be $f(1/11)$. In this paper, using tools in information theory and analysis, they provide both lower and upper bounds on the number of tests needed to identify the subset of items for a large class of functions.
Song Liu's paper is titled 'High-Dimensional Differential Parameter Inference in Exponential Family using Time Score Matching'. In many applications, it is important to understand the evolution of a time-varying process (e.g., stock price fluctuations). A key challenge is recovering the time-score—the instantaneous change of a distribution—from data and given a dataset evolving over time, how can these instantaneous changes be described? A common approach assumes the process is smooth and estimates a probabilistic model at each time point and then compare them, however, this can be computationally and statistically challenging, especially for complex time-series distributions. In this paper, Dr Liu and collaborators propose a method to estimate changes directly from the dataset without learning the full high-dimensional probabilistic model. This approach enables efficient estimation for high-dimensional distributions.
Read the paper "High-Dimensional Differential Parameter Inference in Exponential Family using Time Score Matching" for free online.