Topology Applied to Machine Learning: From Global to Local

Singular-Value-Decomposition-Based Matrix Surgery

This paper is motivated by the need to stabilise the impact of deep learning (DL) training for medical image analysis on the conditioning of convolution filters in relation to model overfitting and robustness. We present a simple strategy to reduce square matrix condition numbers and investigate its effect on the spatial distributions of point clouds of well- and ill-conditioned matrices. For a square matrix, the SVD surgery strategy works by: (1) computing its singular value decomposition (SVD), (2) changing a few of the smaller singular values relative to the largest one, and (3) reconstructing the matrix by reverse SVD. Applying SVD surgery on CNN convolution filters during training acts as spectral regularisation of the DL model without requiring the learning of extra parameters. The fact that the further away a matrix is from the non-invertible matrices, the higher its condition number is suggests that the spatial distributions of square matrices and those of their inverses are correlated to their condition number distributions. We shall examine this assertion empirically by showing that applying various versions of SVD surgery on point clouds of matrices leads to bringing their persistent diagrams (PDs) closer to the matrices of the point clouds of their inverses.

Topological measures of order for imperfect two-dimensional Bravais lattices

Motivated by patterns with defects in natural and laboratory systems, we develop two quantitative measures of order for imperfect Bravais lattices in the plane. A tool from topological data analysis called persistent homology combined with the sliced Wasserstein distance, a metric on point distributions, are the key components for defining these measures. The measures generalize previous measures of order using persistent homology that were applicable only to imperfect hexagonal lattices in two dimensions. We illustrate the sensitivities of these measures to the degree of perturbation of perfect hexagonal, square, and rhombic Bravais lattices. We also study imperfect hexagonal, square, and rhombic lattices produced by numerical simulations of pattern-forming partial differential equations. These numerical experiments serve to compare the measures of lattice order and reveal differences in the evolution of the patterns in various partial differential equations.

Nutrigonometry I: using right-angle triangles to quantify nutritional trade-offs in performance landscapes

AbstractAnimals regulate their food intake to maximize the expression of fitness traits but are forced to trade off the optimal expression of some fitness traits because of differences in the nutrient requirements of each trait ("nutritional trade-offs"). Nutritional trade-offs have been experimentally uncovered using the geometric framework for nutrition (GF). However, current analytical methods to measure such responses rely on either visual inspection or complex models of vector calculations applied to multidimensional performance landscapes, making these approaches subjective or conceptually difficult, computationally expensive, and, in some cases, inaccurate. Here, we present a simple trigonometric model to measure nutritional trade-offs in multidimensional landscapes (nutrigonometry) that relies on the trigonometric relationships of right-angle triangles and thus is both conceptually and computationally easier to understand and use than previous quantitative approaches. We applied nutrigonometry to a landmark GF data set for comparison of several standard statistical models to assess model performance in finding regions in the performance landscapes. This revealed that polynomial (Bayesian) regressions can be used for precise and accurate predictions of peaks and valleys in performance landscapes, irrespective of the underlying structure of the data (i.e., individual food intakes vs. fixed diet ratios). We then identified the known nutritional trade-off between life span and reproductive rate in terms of both nutrient balance and concentration for validation of the model. This showed that nutrigonometry enables a fast, reliable, and reproducible quantification of nutritional trade-offs in multidimensional performance landscapes, thereby broadening the potential for future developments in comparative research on the evolution of animal nutrition.

Personalized Medicine for the Critically Ill Patient: A Narrative Review

Personalized Medicine (PM) is rapidly advancing in everyday medical practice. Technological advances allow researchers to reach patients more than ever with their discoveries. The critically ill patient is probably the most complex of all, and personalized medicine must make serious efforts to fulfill the desire to “treat the individual, not the disease”. The complexity of critically ill pathologies arises from the severe state these patients and from the deranged pathways of their diseases. PM constitutes the integration of basic research into clinical practice; however, to make this possible complex and voluminous data require processing through even more complex mathematical models. The result of processing biodata is a digitized individual, from which fragments of information can be extracted for specific purposes. With this review, we aim to describe the current state of PM technologies and methods and explore its application in critically ill patients, as well as some of the challenges associated with PM in intensive care from the perspective of economic, approval, and ethical issues. This review can help in understanding the complexity of, P.M.; the complex processes needed for its application in critically ill patients, the benefits that make the effort of implementation worthwhile, and the current challenges of PM.

'Holey' niche! finding holes in niche hypervolumes using persistence homology

Living organisms are limited in the range of values of ecological factors they can explore. This defines where animals exist (or could exist) and forms an ecological fingerprint that explains species' distribution at global scales. Species' ecological fingerprints can be represented as a n-dimensional hypervolume - known as Hutchinson's niche hypervolume. This concept has enabled significant progress in our understanding of species' ecological needs and distributions across environmental gradients. Nevertheless, the properties of Hutchinson's n-dimensional hypervolumes can be challenging to calculate and several methods have been proposed to extract meaningful measurements of hypervolumes' properties. One key property of hypervolumes are holes, which provide important information about the ecological occupancy of species. However, to date, current methods rely on volume estimates and set operations to identify holes in hypervolumes. Yet, this approach can be problematic because in high-dimensions, the volume of region enclosing a hole tends to zero. We propose the use of persistence homology (PH) to identify holes in hypervolumes and in ecological datasets more generally. PH allows for the estimates of topological properties in n-dimensional niche hypervolumes independent of the volume estimates of the hypervolume. We demonstrate the application of PH to canonical datasets and to the identification of holes in the hypervolumes of five vertebrate species with diverse niches, highlighting the potential benefits of this approach to gain further insights into animal ecology. Overall, our approach enables the study of a yet unexplored property of Hutchinson's hypervolumes, and thus, have important implications to our understanding of animal ecology.