The field of machine learning has become strongly compute driven, such that emerging research and applications require larger amounts of specialised hardware or smarter algorithms to advance beyond the state-of-the-art. This thesis develops specialised techniques and algorithms for a subset of computationally difficult machine learning problems. The applications under investigation are quantile approximation in the limited-memory data streaming setting, interpretability of decision tree ensembles, efficient sampling methods in the space of permutations, and the generation of large numbers of pseudorandom permutations. These specific applications are investigated as they represent significant bottlenecks in real-world machine learning pipelines, where improvements to throughput have significant impact on the outcomes of machine learning projects in both industry and research. To address these bottlenecks, we discuss both theoretical improvements, such as improved convergence rates, and hardware/software related improvements, such as optimised algorithm design for high throughput hardware accelerators. Some contributions include: the evaluation of bin-packing methods for efficiently scheduling small batches of dependent computations to GPU hardware execution units, numerically stable reduction operators for higher-order statistical moments, and memory bandwidth optimisation for GPU shuffling. Additionally, we apply theory of the symmetric group of permutations in reproducing kernel Hilbert spaces, resulting in improved analysis of Monte Carlo methods for Shapley value estimation and new, computationally more efficient algorithms based on kernel herding and Bayesian quadrature. We also utilise reproducing kernels over permutations to develop a novel statistical test for the hypothesis that a sample of permutations is drawn from a uniform distribution. The techniques discussed lie at the intersection of machine learning, high-performance computing, and applied mathematics. Much of the above work resulted in open source software used in real applications, including the GPUTreeShap library [38], shuffling primitives for the Thrust parallel computing library [2], extensions to the Shap package [31], and extensions to the XGBoost library [6].