Presentation of Master's theses in Mathematical statistics

Date: Wednesday 9 September 2026
Time: 09:00 - 11:10
Location: Mittag-Leffler room, Department of Mathematics, floor 3, house 1, Albano

On Wednesday, September 9, three master's theses in mathematical statistics will be presented

09:00-9:40 Isaac Zúñiga
Title: Scrutinizing the Training Dynamics of Neural Network Representations: An Approach from Graph-based Representation Dissimilarity
Supervisor: Chun-Biu Li

Abstract:
Deep neural networks achieve remarkable predictive performance, yet understanding
how their internal representations evolve throughout training remains a fundamental
challenge. This thesis introduces the Shape-Aware Graph Distance (SAGD), a graphbased
framework for comparing neural network representations through their intrinsic
geometric structure. Representations are modeled as weighted graphs, whose pairwise
Commute Time Distances are summarized by empirical cumulative distribution functions
and compared using the 1-Wasserstein distance. This construction provides a global
measure of representational dissimilarity that is naturally invariant to permutations of
node labels and captures structural relationships beyond standard Euclidean comparisons.
The theoretical framework combines spectral graph theory, random walks, and 1-
Wasserstein distance to establish the mathematical foundations of SAGD and to analyze
its principal properties. Robustness is evaluated through controlled experiments involving
synthetic Gaussian data, high-dimensional Gaussian mixtures, and Watts–Strogatz smallworld
networks, demonstrating that the proposed methodology is stable under moderate
variations in graph construction and stochastic perturbations.
Finally, SAGD is applied to study the evolution of internal representations in a ResNet-
18 model trained on the CIFAR-10 dataset. By combining SAGD with the graph-based
visualization method; Shape-Aware Stochastic Neighbor Embedding (SASNE), the analysis
reveals distinct geometric trajectories across network layers and training epochs. The
results show that representations undergo rapid structural reorganization during the early
stages of optimization, followed by gradual refinement, and continue to evolve even after
predictive performance has largely converged. Overall, this work presents a graph-theoretic
perspective on representation learning and provides a mathematically grounded framework
for analyzing the geometry and dynamics of neural network representations.

9:45-10:25 Yahani Wathsala
Title: Simulation based analysis of infectious disease transmission trees
Supervisor: Martina Favero

Abstract:
In this thesis, we investigate how different stochastic transmission
models and superspreader behavior are reflected in transmission trees.
As our models of interest we use four continuous-time Markov chain epidemics
in a closed population, a homogeneous SIR model, an SEIR model
with a latent period, an SIR model with Gamma-distributed individual
infectivity to capture superspreading, and an SIR model with households
and global mixing. Drawing inspiration from tree-based epidemic analysis,
we simulate each model using the Gillespie algorithm and record
who infected whom, constructing a transmission tree for every outbreak.
In our simulations, the superspreading model produces highly overdispersed
offspring distributions, large Gini coefficients and strong ”top x%
cause y%” effects, with many infections generated by a small number of
high degree nodes. The household model yields comparatively shallow
trees, similar in depth to the superspreading model despite arising from a
very different mechanism. We also examine backward, ancestral offspring
statistics and show that ancestors tend to have more offspring than a
typical infectious individual, consistent with size biased sampling. As a
further assessment of the models, we compare tree-based statistics with
final epidemic sizes. Based on these results, we argue that forward and
backward transmission summaries can help to identify the superspreading
and mixing structure and to guide the selection of simple stochastic
models for real transmission-tree data.

10:30-11:10 Li Wu
Title: A Comparative Study of Unsupervised Anomaly Detection Methods: Isolation Forest and LOF
Supervisor: Chun-Biu Li

Abstract:
This study compares two unsupervised anomaly detection methods: Isolation Forest
and Local Outlier Factor (LOF). Experiments are conducted on two datasets with
varying sizes, dimensionalities, and anomaly ratios. Performance is evaluated using
Receiver Operating Characteristic (ROC) analysis and Precision-Recall (PR) analysis,
together with an analysis of hyperparameter sensitivity and computational efficiency.
The results indicate that Isolation Forest provides a strong balance between accuracy
and efficiency, making it a robust general-purpose method. In contrast, LOF performs
well only when anomalies are locally separable and the data are low-dimensional,
while being sensitive to parameter settings.

Last updated: 2026-09-07

Source: Department of Mathematics