arXiv subjects

Paul Matthews

Publications and source records attributed to Paul Matthews.

6 recordsLinked to original sources

arXiv2021

Neurological Consequences of COVID-19 Infection

COVID-19 infections have well described systemic manifestations, especially respiratory problems. There are currently no specific treatments or vaccines against the current strain. With higher case numbers, a range of neurological symptoms are becoming apparent. The mechanisms responsible for these are not well defined, other than those related to hypoxia and microthrombi. We speculate that sustained systemic immune activation seen with SARS-CoV-2 may also cause secondary autoimmune activation in the CNS. Patients with chronic neurological diseases may be at higher risk because of chronic secondary respiratory disease and potentially poor nutritional status. Here, we review the impact of COVID-19 on people with chronic neurological diseases and potential mechanisms. We believe special attention to protecting people with neurodegenerative disease is warranted. We are concerned about a possible delayed pandemic in the form of an increased burden of neurodegenerative disease after acceleration of pathology by systemic COVID-19 infections.

q-bio.NC↗

arXiv2020

Tensor Dropout for Robust Learning

CNNs achieve remarkable performance by leveraging deep, over-parametrized architectures, trained on large datasets. However, they have limited generalization ability to data outside the training domain, and a lack of robustness to noise and adversarial attacks. By building better inductive biases, we can improve robustness and also obtain smaller networks that are more memory and computationally efficient. While standard CNNs use matrix computations, we study tensor layers that involve higher-order computations and provide better inductive bias. Specifically, we impose low-rank tensor structures on the weights of tensor regression layers to obtain compact networks, and propose tensor dropout, a randomization in the tensor rank for robustness. We show that our approach outperforms other methods for large-scale image classification on ImageNet and CIFAR-100. We establish a new state-of-the-art accuracy for phenotypic trait prediction on the largest dataset of brain MRI, the UK Biobank brain MRI dataset, where multi-linear structure is paramount. In all cases, we demonstrate superior performance and significantly improved robustness, both to noisy inputs and to adversarial attacks. We rigorously validate the theoretical validity of our approach by establishing the link between our randomized decomposition and non-linear dropout.

cs.LG↗

arXiv2016

Structural and configurational properties of nanoconfined monolayer ice from first principles

Understanding the structural tendencies of nanoconfined water is of great interest for nanoscience and biology, where nano/micro-sized objects may be separated by very few layers of water. Here we investigate the properties of ice confined to a quasi-2D monolayer by a featureless, chemically neutral potential, in order to characterize its intrinsic behaviour. We use density-functional theory simulations with a non-local van der Waals density functional. An ab initio random structure search reveals all the energetically competitive monolayer configurations to belong to only two of the previously-identified families, characterized by a square or honeycomb hydrogen-bonding network, respectively. We discuss the modified ice rules needed for each network, and propose a simple point dipole 2D lattice model that successfully explains the energetics of the square configurations. All identified stable phases for both networks are found to be non-polar (but with a topologically non-trivial texture for the square) and, hence, non-ferroelectric, in contrast to previous predictions from a five-site empirical force-field model. Our results are in good agreement with very recently reported experimental observations.

physics.chem-ph↗

arXiv2014

Orientation-dependent pinning and homoclinic snaking on a planar lattice

We study homoclinic snaking of one-dimensional, localised states on two-dimensional, bistable lattices via the method of exponential asymptotics. Within a narrow region of parameter space, fronts connecting the two stable states are pinned to the underlying lattice. Localised solutions are formed by matching two such stationary fronts back-to-back; depending on the orientation relative to the lattice, the solution branch may `snake' back and forth within the pinning region via successive saddle-node bifurcations. Standard continuum approximations in the weakly nonlinear limit (equivalently, the limit of small mesh size) do not exhibit this behaviour, due to the resultant leading-order reaction-diffusion equation lacking a periodic spatial structure. By including exponentially small effects hidden beyond all algebraic orders in the asymptotic expansion, we find that exponentially small but exponentially growing terms are switched on via error function smoothing near Stokes lines. Eliminating these otherwise unbounded beyond-all-orders terms selects the origin (modulo the mesh size) of the front, and matching two fronts together yields a set of equations describing the snaking bifurcation diagram. This is possible only within an exponentially small region of parameter space---the pinning region. Moreover, by considering fronts orientated at an arbitrary angle $ψ$ to the $x$-axis, we show that the width of the pinning region is non-zero only if $\tanψ$ is rational or infinite. This is the first time a formula explicitly relating the orientation of a front to the width of its pinning region has been derived. The asymptotic results are compared with numerical calculations, with good agreement.

math.AP↗

arXiv2013

Dissipation in a topological Josephson junction

Topological features of low dimensional superconductors have caused a lot of excitement recently because of their broad range of applications in quantum information and their potential to reveal novel phases of quantum matter. A potential problem for practical applications is the presence of phase-slips that break phase coherence. Dissipation in non-topological superconductors suppresses phase-slips and can restore long-range order. Here we investigate the role of dissipation in a topological Josephson junction. We show that the combined effects of topology and dissipation keeps phase and anti-phase slips strongly correlated so that the device is superconducting even under conditions where a non-topological device would be resistive. The resistive transition occurs at a critical value of the dissipation which is four times smaller than that expected for a conventional Josephson junction. We propose that this difference could be employed as a robust experimental signature of topological superconductivity.

cond-mat.supr-con↗

arXiv2010

Do absolutely irreducible group actions have odd dimensional fixed point spaces?

In his volume [5] on "Symmetry Breaking for Compact Lie Groups" Mike Field quotes a private communication by Jorge Ize claiming that any bifurcation problem with absolutely irreducible group action would lead to bifurcation of steady states. The proof should come from the fact that any absolutely irreducible representation possesses an odd dimensional fixed point space. In this paper we show that there are many examples of groups which have absolutely irreducible representations but no odd dimensional fixed point space. This observation may be relevant also for some degree theoretic considerations concerning equivariant bifurcation. Moreover we show that our examples give rise to some interesting Hamiltonian dynamics and we show that despite some complications we can go a long way towards doing explicit computations and providing complete proofs. For some of the invariant theory needed we will depend on some computer aided computations. The work presented here greatly benefited from the computer algebra program GAP [6], which is an indispensable aid for doing the required group theory computations.

math.DS↗