Nine MSc theses to be presented on Thursday May 30, 2024

Thursday May 30 is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Nine MSc theses will be presented.

You will find information about how to follow along under each presentation. There will be presentations in three different rooms: E:2116, E:2405 (Glasburen) and E:4130 (Lucas). See room for each presentation. A preliminary schedule follows.

Please note that there will also be thesis presentations on Friday May 31, schedule at: https://cs.lth.se/kalendarium/?evenemang=cs-msc-thesis-presentation-day-may-31-2024

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.

09:00-09:45 in E:4130 (Lucas) and via Zoom N.B. No more opponents for this presentation

Presenters: Théodore Zitouni, Axel Beke
Title: Fine-tuning Phi Models for Informed Decision Support in Supply Chain Optimisation
Examiner: Elin Anna Topp
Supervisors: Pierre Nugues (LTH), Bahram Zarrin (Microsoft)

In the rapidly evolving domain of supply chain management, optimising operations for efficiency and reliability is essential. Traditional methods have struggled to keep pace with the complexity of modern logistics networks and the vast amount of data they generate. However, recent advancements in artificial intelligence offer a new avenue for innovation in this field. This research focuses on the application of Phi models, a series of small language models by Microsoft, to address two tasks in supply chain optimisation: code generation for what-if analysis and the job shop scheduling problem. We show that fine-tuning the Phi-2 model enhances its performance on these tasks, demonstrating a marked improvement in code generation capabilities with a BLEU score of 98.6%. Though it encounters challenges with the job shop scheduling problem due to its inherent complexity, it still displays a novel understanding of the problem and how to solve it, attaining a BLEU score of 21.7% using a whitespace tokenizer.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240530_09BekeZitouni.pdf

Link to Zoom presentation: https://lu-se.zoom.us/j/69950449311?pwd=RTRMcmNxYklGL1d0eWtuWFZSWVpoQT09

09:15-10:00 in E:2405 (Glasburen) N.B. No more opponents for this presentation

Presenters: Oskar Hallberg, Oscar Peyron
Title: Exploring the Potential of Generative AI for Corporate Documentation Management
Examiner: Alma Orucevic-Alagic
Supervisor: Elizabeth Bjarnason (LTH)

Efficient and accurate documentation is an important element for maintaining effective workflows in corporations. This thesis investigates the integration of generative AI into documentation processes of a multinational corporation to address inefficiencies in manual documentation management. Using Design Science research complemented by Cross-Industry Standard Process for Data Mining, we developed and evaluated an AI-driven solution tailored to automate and enhance the creation and of documentation. Through interviews and thematic analysis, we identified key challenges such as scalability, time consumption, inaccuracies, and resistance to technology adoption. Our solution employs fine-tuning, Retrieval-Augmented Generation, and prompt engineering to generate accurate and contextually relevant documents. The solution demonstrated improvements in documentation efficiency and quality while reducing manual errors. However, integration challenges and the need for continuous model training were noted. The findings suggest that while AI can improve documentation processes, ongoing adjustments and adaptations are essential for maintaining alignment with corporate standards and practices.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240530_09PeyronHallberg.pdf

10:15-11:00 in E:2116 N.B. No more opponents for this presentation

Presenters: Madeleine Berild, Jakob Sinclair
Title: Optimizing Memory Usage of Tensors in Neural Networks
Examiner: Flavius Gruian
Supervisors: Jonas Skeppstedt (LTH), Kristofer Jonsson (Arm)

Tensors, the data objects in the neural network graph, can easily be allocated by obtaining their own space in GPU memory. However, this approach is highly inefficient, as many tensors are only required at specific points during execution and are not needed thereafter. A more effective strategy involves reusing memory that is no longer active. The challenge lies in determining when a tensor can be reused, particularly as branches in the graph are executed in parallel. In this thesis, different algorithms and strategies are explored and tested to reduce the memory consumption of parallel neural networks. The result is a new algorithm called live-tensor analysis that determines which tensors depend on each other. Together with best-fit allocation and various search algorithms, we can find an effective allocation. The implementation can greatly reduce memory usage depending on the neural network structure while maintaining the correct output.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240530_10BerildSinclair.pdf

11:15-12:00 in E:2116

Presenters: Richard Lundberg, Marcus Rettig
Title: Modelling Profiling Data in a Graph Database for Performance Analysis
Examiner: Per Andersson
Supervisors: Jonas Skeppstedt (LTH), Simon Priisalu (Neo4j), Jaroslaw Palka (Neo4j)

Benchmarking is an important part of the development process for any mission-critical application. We propose a framework for identifying bottlenecks and regressions by modeling profiling data as call-stack trees in a graph database. We demonstrate the usefulness of the framework for cross-profile analysis such as time series analysis and aggregation-based methods. We conclude that there is much potential in this approach and our thesis can be used as a decision basis for organizations wanting to implement a similar framework. Using a graph database to model profiling data has many advantages and is suitable for the tree-like structure of the data. It makes the data more accessible and facilitates flexible querying in which the user can ask questions about the data and perform non-trivial aggregation. The main disadvantage is the complexity involved in importing large quantities of data.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240530_11LundbergRettig.pdf

13:15-14:00 in E:2116 N.B. No more opponents for this presentation

Presenters: Maria Svensson, Patrik Gyllvin
Title: Segmenting a power consumption profile for approximating battery behavior
Examiner: Flavius Gruian
Supervisors: Jonas Skeppstedt (LTH), Björn Rosqvist (Qoitech AB), Andreas Olausson (Qoitech AB)

This thesis aims to analyze how a power consumption profile for a battery powered device can be segmented into a number of different sized steps while keeping the most important characteristics for preserving the behavior of the battery under load of said profile. It attempts to investigate what those important characteristics are and how they can be preserved using proposed algorithms. Peaks in the profile in particular were deemed of significant meaning because of how batteries react to high peak loads. Our algorithms aim to recreate peaks as close as possible given the constraints of the segmentation, and our analysis is focused on the effect these have.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240530_13SvenssonGyllvin.pdf

14:15-15:00 in E:2405 (Glasburen)

Presenters: Lisa Bybro, Joel Engström
Title: Applying Hirschberg’s Algorithm on Matrices
Examiner: Görel Hedin
Supervisor: Jonas Skeppstedt (LTH)

There are many different ways of optimising video and other forms of image sequences in use today. This thesis attempts to apply an existing algorithm, called Hirschberg's algorithm, to this very same problem. The algorithm has previously mainly been used within DNA sequence comparison, but is here used to compare two consecutive images and extract an efficient representation of the changes between them (delta). Applying an algorithm meant for one-dimensional data on two-dimensional data can be done in many different ways, some of which are explored here. While our findings show that Hirschberg's algorithm does lead to some minor improvements on the delta data size, it also creates other problems including time complexity and inability to apply other optimisations. We test the algorithm on image sets containing small differences, and compare the results to other, more naive algorithms.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240530_14BybroEngstrom.pdf

14:15-15:00 in E:4130 (Lucas)

Presenters: Hashim Siddig Hashim Ismail, Jialong Li
Title: Language Guided Object Picking for Robots
Examiner: Michael Doggett
Supervisor: Volker Krueger (LTH)

The project aims to establish a deep-learning-based end-to-end pipeline from RGB-D and natural language instructions input to feasible grasping, and to apply and validate this pipeline on a real robot arm, identifying its feasibility and limitations. The pipeline specifically employs the Grounding Dino model for open-set object detection, the Segment Anything model for zero-shot object segmentation, and Contact-GraspNet for generating feasible grasps on unknown objects based on 2.5D point cloud inputs. Through testing, it was found that the pipeline performs well in most scenarios, with limitations primarily attributed to the inconsistencies of the visionlanguage model and the inherent limitations of 2.5D point cloud inputs in representing global scene information.

Link to popular science summary: To be added

15:15-16:00 in E:4130 (Lucas)

Presenters: Daniel Kärde, Jonathan Permfors
Title: Accelerating NeRF-based Rendering Using Bounding Volume Hierarchies
Examiner: Michael Doggett
Supervisor: Rikard Olajos (LTH)

Originally introduced in 2020, Neural Radiance Fields (NeRFs) have become a great area of interest within the research community. NeRFs combines neural networks with volumetric rendering in order to synthesise new images of 3D scenes by training on a sparse set of 2D images.. The neural network describes the scene as a continuous field and as such must be sampled from in order to render the final image. The computational cost of inference on these samples are very high, which incentivizes reducing the number of samples to a minimum while maintaining image quality. We present a method of accelerating the rendering of NeRFs through the usage of Bounding Volume Hierarchies that are constructed from point clouds exported from a pre-trained model. From the point clouds we utilise K-Means to group the points into clusters, from which we instantiate our bottom-level bounding boxes. We experiment with varying point cloud sizes and cluster counts, along with different tree construction methods. By modifying a CUDA-based ray tracer we are able to efficiently compute intersection points which we use to place samples where the density exceeds some threshold value. The reduction in samples results in a significant decrease in rendering times while maintaining image quality. We limit our work to that of bounded scenes and note that there are many interesting areas of future research.

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/240530_15PermforsKärde.pdf

16:15-17:00 in E:2116 N.B. Change of time and also no more opponents for this presentation

Presenter: Christian Benson
Title: Optimizing MLPs for NeRF rendering
Examiner: Flavius Gruian
Supervisor: Michael Doggett (LTH)

NeRF rendering is a technique that renders a scene by utilizing neural networks (MLPs). When using this, the inference time of the network is of great interest as it dictates how fast the rendering process will be. In this thesis we look closer at what techniques can be used in order to shorten the inference time as much as possible. Two different type of networks, both built in CUDA, are compared. In order to judge what makes these networks different the inference time is measured, as well as tensor core utilization, cache utilization and startup time of kernels on the GPU. It is found that a fully fused approach is significantly faster as well as scales better when rendering larger images. This is mainly credited to how this network utilizes the cache memory, gaining a significantly higher cache hit rate compared to the CUTLASS network it is compared to.

Link to popular science summary: To be added