This thesis is concerned with two main areas: approximate inference in discrete graphical models, and random embeddings for dimensionality reduction and approximate inference in kernel methods. Approximate inference is a fundamental problem in machine learning and statistics, with strong connections to other domains such as theoretical computer science. At the same time, there has often been a gap between the success of many algorithms in this area in practice, and what can be explained by theory; thus, an important research effort is to bridge this gap. Random embeddings for dimensionality reduction and approximate inference have led to great improvements in scalability of a wide variety of methods in machine learning. In recent years, there has been much work on how the stochasticity introduced by these approaches can be better controlled, and what further computational improvements can be made.

In the first part of this thesis, we study approximate inference algorithms for discrete graphical models. Firstly, we consider linear programming methods for approximate MAP inference, and develop our understanding of conditions for exactness of these approximations. Such guarantees of exactness are typically based on either structural restrictions on the underlying graph corresponding to the model (such as low treewidth), or restrictions on the types of potential functions that may be present in the model (such as log-supermodularity). We contribute two new classes of exactness guarantees: the first of these takes the form of particular hybrid restrictions on a combination of graph structure and potential types, whilst the second is given by excluding particular substructures from the underlying graph, via graph minor theory. We also study a particular family of transformation methods of graphical models, uprooting and rerooting, and their effect on approximate MAP and marginal inference methods. We prove new theoretical results on the behaviour of particular approximate inference methods under these transformations, in particular showing that the triplet relaxation of the marginal polytope is unique in being universally rooted. We also introduce a heuristic which quickly picks a rerooting, and demonstrate benefits empirically on models over several graph topologies.

In the second part of this thesis, we study Monte Carlo methods for both linear dimensionality reduction and approximate inference in kernel machines. We prove the statistical benefit of coupling Monte Carlo samples to be almost-surely orthogonal in a variety of contexts, and study fast approximate methods of inducing this coupling. A surprising result is that these approximate methods can simultaneously offer improved statistical benefits, time complexity, and space complexity over i.i.d. Monte Carlo samples. We evaluate our methods on a variety of datasets, directly studying their effects on approximate kernel evaluation, as well as on downstream tasks such as Gaussian process regression.