Two Computer Science MSc theses to be presented on 14 June

Friday, 14 June there will be two master thesis presentations in Computer Science at Lund University, Faculty of Engineering.

The presentations will take place in E:4130 (Lucas) and E:2116. See information for each presentation.

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname.lastname@cs.lth.se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions are found on this page.

10:15-11:00 in E:4130 (Lucas)

Presenters: Adi Creson, Mamdollah Amini
Title: Cross-Domain Generalizability in Image Feature Extraction
Examiner: Volker Krüeger
Supervisors: Alexander Dürr (LTH), Simon Kristofferson Lind (LTH)

This thesis examines the effectiveness of using ResNet-50, a pre-trained deep convolutional neural network, as a feature extractor in the reinforcement learning environment of the Atari game Breakout. The study evaluates the performance of features extracted from the last block of different stages of ResNet-50 in training a reinforcement learning agent and compares these across the stages. Through a multi-phase experimental setup, the research explores ResNet-50's ability to adapt to domains outside its original training, without fine-tuning the model. The findings reveal that all stages of ResNet-50 underperformed, particularly in comparison to a established benchmark. Notably, the last stage, stage 4, showed some potential for learning despite overall poor performance. The results suggest that ResNet-50 as a feature extractor has limited success in Breakout and depends heavily on careful integration and design of the reinforcement learning pipeline. This study contributes to the ongoing discussion about the practicality of leveraging large pre-trained models in new domains, underscoring both the challenges and opportunities of repurposing these models for diverse applications.

Link to popular science summary: To be uploaded

13:15-14:00 in E:2116

Presenter: Alfred Clemedtson
Title: Finding candidate pairs for heterogeneous link prediction
Examiner: Maj Stenmark
Supervisors: Jonas Skeppstedt (LTH), Jacob Sznajdman (Neo4j)

Knowledge graphs can be viewed as an extension of traditional graphs, where edges are replaced with triplets on the form (node-relation-node). In this thesis, we look into link prediction, the task of predicting new triplets based on patterns in the graph. We explore interesting models used, such as random walk based methods, trainable embeddings and graph neural networks. Scoring every possible triplet of a graph is unfeasible for large graphs and we recognize the need for efficient candidate-pair sampling methods to be able to apply models that score triplets indivudually. We present a novel, scalable candidate sampling method, WAPPR, an extension and approximation of the Personalized PageRank algorithm to support multi-relational graphs. We benchmark WAPPR as a candidate sampling method, on available, standard datasets but find that it fails to beat approximate nearest neighbor search on a trained node embeddings in terms of recall in most cases.

Link to popular science summary: To be uploaded