lunduniversity.lu.se

Digit@LTH

Faculty of Engineering, LTH

Denna sida på svenska This page in English

Digit@LTH: Events

CS Extra MSc Thesis Presentation Day June 12!

Föredrag

From: 2020-06-12 10:15 to: 15:00
Place: Online Zoom presentations
Contact: birger [dot] swahn [at] cs [dot] lth [dot] se
Save event to your calendar


Four MSc theses to be presented on Friday June 12, 2020

Friday June 12 is an extra day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Four MSc theses will be presented. Other presentations will take place on June 11 and June 15.

The presentations will take place online via Zoom, see separate link for every presentation. A preliminary schedule follows.

Note to potential opponents: Register as an opponent to the presentation of your choice by following the instructions on Doodle at https://doodle.com/poll/cpfr4bh5npvfstra. 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.

 

Schedule


10:15-11.00

Presenter: Johan Pettersson
Title: Real-time rendering and dynamics of sparse voxel octree clouds
Examiner: Michael Doggett
Supervisors: Pierre Moreau (LTH)

State-of-the-art real-time cloud rendering in games has transitioned from 2D skyboxes to ray marched 3D volumes more closely approximating the volumetric nature of reality. However, most solutions repeat small volumes of details to save memory and show only a few hardcoded cloud layers, which cannot capture the breadth and depth of complexities in the real atmosphere. This thesis develops and presents an atmosphere renderer capable of displaying views from any altitude within or outside an Earth-scale atmosphere, seamlessly transitioning between them, with adaptive level of detail changing the allocation of memory and computational resources depending on the current camera location. By separating cloud creation from cloud rendering, this method in principle enables rendering of cloud data from any source. Ultimately the presented implementation renders on average 30 frames per second with adequate quality settings on a high end GPU; however, there may be further optimisation potential.

Link to presentation: https://lu-se.zoom.us/j/67278255908

Code for opponents: 23 MD

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/200612/10Pettersson.pdf


13:00-14.00

Presenter: Lucas Molsby
Title: Exploring methods for hate speech detection in text
Examiner: Jacek Malec
Supervisors: Pierre Nugues (LTH), Petter (Säkerhetspolisen)

Detection of hate speech can be used for many applications. Most commonly it is used for creating a safe and just setting for online communication but it can also be an asset when working with prevention of violent extremism. In this paper we train classifiers to detect hate in the Gab Hate Corpus, a corpus collected from the social media platform Gab. Our results show that, by fine-tuning pre-trained models and excluding a selection of data, we can outperform the current state of the art on this task and achieve a 0.15 points greater macro F1 score.

Link to presentation: https://lu-se.zoom.us/j/63528382897

Code for opponents: 12 JM


13:15-14.00

Presenters: Josefine Myllenberg, Jens Johansson
Title: Optimizing Machine Learning Inference for MCU:s
Examiner: Jörn Janneck
Supervisors: Flavius Gruian (LTH), Johan Björnstedt (Acconeer AB)

Deep neural networks come with high demand for storage and computational resources, which makes it difficult to deploy deep convolutional neural networks on limited resource devices. This thesis investigates different approaches of how to reduce the size of a network in order to run it on a limited resource devices, while keeping the accuracy close to the original network. The network is converted to C code using X-CUBE-AI and Keras2C in order to run on an MCU. Different pruning techniques such as channel based pruning and magnitude based pruning are then applied in order to reduce the network size and inference time. Results show that the execution time can be reduced by up to x8 and memory usage by up to x4.5. Quantization is also applied, which results in a reduction in execution time by x2.5 and memory usage by x4.

Link to presentation: https://lu-se.zoom.us/j/4787325598

Code for opponents: 28 JJ


14:00-15.00

Presenters: Daniel Regefalk, Alexander Goobar
Title: Classification of Short Text Messages using Machine Learning
Examiner: Jacek Malec
Supervisor: Pierre Nugues (LTH), Jianhua Cao (Sinch), Michael Truong (Sinch)

In this Master's thesis, carried out at Sinch Sweden AB, we evaluate several machine learning models and their performance at classifying short text messages, up to 160 characters in length. Sinch provides services for companies to send business-to-consumer SMS messages, and wants to be able to identify unwanted messages with prohibited content. Both traditional machine learning algorithms and more recent deep learning models were evaluated for this task. To generalize the findings, two public datasets from other domains were also used for the evaluation. Our results show that recent deep learning models based on transformers, such as BERT, perform the best on the public datasets. However, some traditional algorithms, such as random forest and support vector machine, perform similarly to these models on the Sinch data. The machine learning models outperform Sinch’s current solution for message blocking, which is based on regular expressions.

Link to presentation: https://lu-se.zoom.us/j/67572330180

Code for opponents: 10 JM

Link to popular science summary: https://fileadmin.cs.lth.se/cs/Education/Examensarbete/Popsci/200612/14RegefalkGoobar.pdf