The accurate predictions and principled uncertainty measures provided by GP regression incur $O(n^3)$ cost which is prohibitive for modern-day large-scale applications. This has motivated extensive work on computationally efficient approximations. We introduce a new perspective by exploring robustness properties and limiting behaviour of GP nearest-neighbour (GPnn) prediction. We demonstrate through theory and simulation that as the data-size $n$ increases, accuracy of estimated parameters and GP model assumptions become increasingly irrelevant to GPnn predictive accuracy. Consequently, it is sufficient to spend small amounts of work on parameter estimation in order to achieve high MSE accuracy, even in the presence of gross misspecification. In contrast, as $n \rightarrow \infty$, uncertainty calibration and NLL are shown to remain sensitive to just one parameter, the additive noise-variance; but we show that this source of inaccuracy can be corrected for, thereby achieving both well-calibrated uncertainty measures and accurate predictions at remarkably low computational cost. We exhibit a very simple GPnn regression algorithm with stand-out performance compared to other state-of-the-art GP approximations as measured on large UCI datasets. It operates at a small fraction of those other methods' training costs, for example on a basic laptop taking about 30 seconds to train on a dataset of size $n = 1.6 \times 10^6$.

With Moore's law progressively running out of steam, heterogeneous computing architectures have been powering the top supercomputers in the world for many years and are now finding broader adoption across the industry. The trend towards sustainable computing also requires domain-specific heterogeneous hardware architectures, which promise further gains in energy efficiency. At the same time, today's high performance computing applications have evolved from monolithic simulations in a single domain to multidisciplinary complex workflows. In this paper, we explore how these trends affect system design decisions and what this means for future computing system architectures.

Much of machine learning relies on the use of large amounts of data to train models to make predictions. When this data comes from multiple sources, for example when evaluation of data against a machine learning model is offered as a service, there can be privacy issues and legal concerns over the sharing of data. Fully homomorphic encryption (FHE) allows data to be computed on whilst encrypted, which can provide a solution to the problem of data privacy. However, FHE is both slow and restrictive, so existing algorithms must be manipulated to make them work efficiently under the FHE paradigm. Some commonly used machine learning algorithms, such as Gaussian process regression, are poorly suited to FHE and cannot be manipulated to work both efficiently and accurately. In this paper, we show that a modular approach, which applies FHE to only the sensitive steps of a workflow that need protection, allows one party to make predictions on their data using a Gaussian process regression model built from another party's data, without either party gaining access to the other's data, in a way which is both accurate and efficient. This construction is, to our knowledge, the first example of an effectively encrypted Gaussian process.

Particle Filters (PFs) are Sequential Monte Carlo methods which are widely used to solve filtering problems of dynamic models under Non-Linear Non-Gaussian noise. Modern PF applications have demanding accuracy and run-time constraints that can be addressed through parallel computing. However, an efficient parallelization of PFs can only be achieved by effectively parallelizing the bottleneck: resampling and its constituent redistribution step. A pre-existing implementation of redistribute on Shared Memory Architectures (SMAs) achieves O(NT log2N) time complexity over T parallel cores. This redistribute implementation is, however, highly computationally intensive and cannot be effectively parallelized due to the inherently limited number of cores of SMAs. In this paper, we propose a novel parallel redistribute on OpenMP 4.5 which takes O(N/T + log2N) steps and fully exploits the computational power of SMAs. The proposed approach is up to six times faster than the O(N/T log2N) one and its implementation on GPU provides a further three-time speed-up vs its equivalent on a 32-core CPU. We also show on an exemplary PF that our redistribution is no longer the bottleneck.

Research on the microbiome is an emerging and crucial science that finds many applications in healthcare, food safety, precision agriculture and environmental studies. Huge amounts of DNA from microbial communities are being sequenced and analyzed by scientists interested in extracting meaningful biological information from this big data. Analyzing massive microbiome sequencing datasets, which embed the functions and interactions of thousands of different bacterial, fungal and viral species, is a significant computational challenge. Artificial intelligence has the potential for building predictive models that can provide insights for specific cutting edge applications such as guiding diagnostics and developing personalised treatments, as well as maintaining soil health and fertility. Current machine learning workflows that predict traits of host organisms from their commensal microbiome do not take into account the whole genetic material constituting the microbiome, instead basing the analysis on specific marker genes. In this paper, to the best of our knowledge, we introduce the first machine learning workflow that efficiently performs host phenotype prediction from whole shotgun metagenomes by computing similaritypreserving compact representations of the genetic material. Our workflow enables prediction tasks, such as classification and regression, from Terabytes of raw sequencing data that do not necessitate any pre-prossessing through expensive bioinformatics pipelines. We compare the performance in terms of time, accuracy and uncertainty of predictions for four different classifiers. More precisely, we demonstrate that our ML workflow can efficiently classify real data with high accuracy, using examples from dog and human metagenomic studies, representing a step forward towards real time diagnostics and a potential for cloud applications.

Chemical space is so large that brute force searches for new interesting molecules are infeasible. High-throughput virtual screening via computer cluster simulations can speed up the discovery process by collecting very large amounts of data in parallel, e.g., up to hundreds or thousands of parallel measurements. Bayesian optimization (BO) can produce additional acceleration by sequentially identifying the most useful simulations or experiments to be performed next. However, current BO methods cannot scale to the large numbers of parallel measurements and the massive libraries of molecules currently used in high-throughput screening. Here, we propose a scalable solution based on a parallel and distributed implementation of Thompson sampling (PDTS). We show that, in small scale problems, PDTS performs similarly as parallel expected improvement (EI), a batch version of the most widely used BO heuristic. Additionally, in settings where parallel EI does not scale, PDTS outperforms other scalable baselines such as a greedy search, $\epsilon$-greedy approaches and a random search method. These results show that PDTS is a successful solution for large-scale parallel BO.