53-XX DIFFERENTIAL GEOMETRY (For differential topology, see 57Rxx. For foundational questions of differentiable manifolds, see 58Axx)
- 53-00 General reference works (handbooks, dictionaries, bibliographies, etc.)
- 53-01 Instructional exposition (textbooks, tutorial papers, etc.)
- 53-02 Research exposition (monographs, survey articles)
- 53-03 Historical (must also be assigned at least one classification number from Section 01)
-
53-04 Explicit machine computation and programs (not the theory of computation or programming)
(1)
- 53-06 Proceedings, conferences, collections, etc.
-
53Axx Classical differential geometry
(1)
- 53Bxx Local differential geometry
- 53Cxx Global differential geometry [See also 51H25, 58-XX; for related bundle theory, see 55Rxx, 57Rxx]
- 53Dxx Symplectic geometry, contact geometry [See also 37Jxx, 70Gxx, 70Hxx]
- 53Zxx Applications to physics
Refine
Year of publication
Document Type
- ZIB-Report (6)
- Master's Thesis (3)
- Doctoral Thesis (2)
- Software (1)
Language
- English (12)
Is part of the Bibliography
- no (12)
Keywords
- Ancient sundials (1)
- Bi-invariant statistics (1)
- Brain connectomes (1)
- Bézier splines (1)
- Computational Differential Geometry (1)
- Deep Learning (1)
- Geometric Statistics (1)
- Hierarchical (1)
- Image Analysis (1)
- Information Geometry (1)
Merging several fine-tuned neural networks into a single model is attractive whenever retraining is expensive, the original training data is unavailable, or a collection of specialised checkpoints already exists. The approaches proposed for this task are diverse. Among them, Fisher merging draws on the geometry of model space by weighting each parameter with the diagonal Fisher information of the respective checkpoint — yet it evaluates this geometry only at the checkpoints themselves. This thesis formulates merging as the computation of a Fréchet mean on the neuromanifold and approximates it with the time-discrete Riemannian calculus of Heeren, Rumpf, Wardetzky and Wirth, together with the polygonal-spider construction of Heeren, Zhang, Rumpf and Smith, which is transferred here from shape space to the space of neural network parameters. The quadratic Fisher form is chosen as the discretisation energy; it agrees with the Kullback–Leibler divergence to second order and is shown to be admissible up to a fixed positive constant. The resulting discrete Euler–Lagrange equations decouple coordinate-wise into tridiagonal systems, which are solved by the Thomas algorithm; the metric between the checkpoints is obtained via a softmax-kernel interpolation. The Fisher merge arises exactly as the single-segment case, so that comparing the two methods isolates precisely the contribution of path resolution. The construction is evaluated on a Mandelbrot escape-time regression task, on the medical image classification benchmark DermaMNIST, and on the GLUE tasks RTE, MRPC and SST-2 with BERT-base. When the source models are trained under controlled conditions, the multi-segment construction significantly improves on the Fisher merge: by 25% in test MSE on the toy task, and with large effect sizes in AUC and accuracy on DermaMNIST. At foundation-model scale, the ranking is task-dependent. The analysis traces this back to the interpolated metric becoming only weakly position-dependent further from the anchor points, whereas the Fisher merge is the exact minimiser at the checkpoints, where the Fisher information is estimated directly.
Data analysis has become fundamental to our society and comes in multiple facets and approaches. Nevertheless, in research and applications, the focus was primarily on data from Euclidean vector spaces. Consequently, the majority of methods that are applied today are not suited for more general data types. Driven by needs from fields like image processing, (medical) shape analysis, and network analysis, more and more attention has recently been given to data from non-Euclidean spaces---particularly (curved) manifolds. It has led to the field of geometric data analysis whose methods explicitly take the structure (for example, the topology and geometry) of the underlying space into account.
This thesis contributes to the methodology of geometric data analysis by generalizing several fundamental notions from multivariate statistics to manifolds. We thereby focus on two different viewpoints.
First, we use Riemannian structures to derive a novel regression scheme for general manifolds that relies on splines of generalized Bézier curves. It can accurately model non-geodesic relationships, for example, time-dependent trends with saturation effects or cyclic trends. Since Bézier curves can be evaluated with the constructive de Casteljau algorithm, working with data from manifolds of high dimensions (for example, a hundred thousand or more) is feasible. Relying on the regression, we further develop
a hierarchical statistical model for an adequate analysis of longitudinal data in manifolds, and a method to control for confounding variables.
We secondly focus on data that is not only manifold- but even Lie group-valued, which is frequently the case in applications. We can only achieve this by endowing the group with an affine connection structure that is generally not Riemannian. Utilizing it, we derive generalizations of several well-known dissimilarity measures between data distributions that can be used for various tasks, including hypothesis testing. Invariance under data translations is proven, and a connection to continuous distributions is given for one measure.
A further central contribution of this thesis is that it shows use cases for all notions in real-world applications, particularly in problems from shape analysis in medical imaging and archaeology. We can replicate or further quantify several known findings for shape changes of the femur and the right hippocampus under osteoarthritis and Alzheimer's, respectively. Furthermore, in an archaeological application, we obtain new insights into the construction principles of ancient sundials. Last but not least, we use the geometric structure underlying human brain connectomes to predict cognitive scores. Utilizing a sample selection procedure, we obtain state-of-the-art results.
Estimating 3D Shape of the Head Skeleton of Basking Sharks Using Annotated Landmarks on a 2D Image
(2022)
Basking sharks are thought to be one of the most efficient filter-feeding fish in terms of the throughput of water filtered through their gills. Details about the underlying morphology of their branchial region have not been studied due to various challenges in acquiring real-world data. The present thesis aims to facilitate this, by developing a mathematical shape model which constructs the 3D structure of the head skeleton of a basking shark using annotated landmarks on a single 2D image. This is an ill-posed problem as estimating the depth of a 3D object from a single 2D view is, in general, not possible. To reduce this ambiguity, we create a set of pre-defined training shapes in 3D from CT scans of basking sharks. First, the damaged structures of the sharks in the scans are corrected via solving a set of optimization problems, before using them as accurate 3D representations of the object. Then, two approaches are employed for the 2D-to-3D shape fitting problem–an Active Shape Model approach and a Kendall’s Shape Space approach. The former represents a shape as a point on a high-dimensional Euclidean space, whereas the latter represents a shape as an equivalence class of points in this Euclidean space. Kendall’s shape space approach is a novel technique that has not yet been applied in this context, and a comprehensive comparison of the two approaches suggests this approach to be superior for the problem at hand. This can be credited to an improved interpolation of the training shapes.
The Sasaki metric is the canonical metric on the tangent bundle TM of a Riemannian manifold M. It is highly useful for data analysis in TM (e.g., when one is interested in the statistics of a set of geodesics in M). To this end, computing the Riemannian logarithm is often necessary, and an iterative algorithm was proposed by Muralidharan and Fletcher. In this note, we derive approximation formulas of the energy gradients in their algorithm that we use with success.
In many applications, geodesic hierarchical models are adequate for the study of temporal observations. We employ such a model derived for manifold-valued data to Kendall's shape space. In particular, instead of the Sasaki metric, we adapt a functional-based metric, which increases the computational efficiency and does not require the implementation of the curvature tensor. We propose the corresponding variational time discretization of geodesics and employ the approach for longitudinal analysis of 2D rat skulls shapes as well as 3D shapes derived from an imaging study on osteoarthritis. Particularly, we perform hypothesis test and estimate the mean trends.
In many applications, geodesic hierarchical models are adequate for the study of temporal observations. We employ such a model derived for manifold-valued data to Kendall's shape space. In particular, instead of the Sasaki metric, we adapt a functional-based metric, which increases the computational efficiency and does not require the implementation of the curvature tensor. We propose the corresponding variational time discretization of geodesics and apply the approach for the estimation of group trends and statistical testing of 3D shapes derived from an open access longitudinal imaging study on osteoarthritis.
For Kendall’s shape space we determine analytically Jacobi fields and parallel transport, and compute geodesic regression. Using the derived expressions, we can fully leverage the geometry via Riemannian optimization and reduce the computational expense by several orders of magnitude. The methodology is demonstrated by performing a longitudinal statistical analysis of epidemiological shape data.
As application example we have chosen 3D shapes of knee bones, reconstructed from image data of the Osteoarthritis Initiative. Comparing subject groups with incident and developing osteoarthritis versus normal controls, we find clear differences in the temporal development of femur shapes. This paves the way for early prediction of incident knee osteoarthritis, using geometry data only.
We introduce techniques for the processing of motion and animations of non-rigid shapes. The idea is to regard animations of deformable objects as curves in shape space. Then, we use the geometric structure on shape space to transfer concepts from curve processing in R^n to the processing of motion of non-rigid shapes. Following this principle, we introduce a discrete geometric flow for curves in shape space. The flow iteratively replaces every shape with a weighted average shape of a local neighborhood and thereby globally decreases an energy whose minimizers are discrete geodesics in shape space. Based on the flow, we devise a novel smoothing filter for motions and animations of deformable shapes. By shortening the length in shape space of an animation, it systematically regularizes the deformations between consecutive frames of the animation. The scheme can be used for smoothing and noise removal, e.g., for reducing jittering artifacts in motion capture data. We introduce a reduced-order method for the computation of the flow. In addition to being efficient for the smoothing of curves, it is a novel scheme for computing geodesics in shape space. We use the scheme to construct non-linear “Bézier curves” by executing de Casteljau’s algorithm in shape space.
