CS MSc Presentation Day August 30!
Plats: E-huset E:4130 LUCAS & E:2116
Four MSc theses to be presented on August 30, 2018
August 30 is the day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Four MSc theses will be presented.
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 [dot] lastname [at] cs [dot] lth [dot] 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 (E:4130 LUCAS)
Presenter: Alexander Olsson
Title: High-performance signal processing for digital AESA-radar
Examiner: Flavius Gruian
Supervisor: Jörn Janneck
The demands on signal processing in digital AESA-systems is rising rapidly. It's not only tough demands on performance, but also on how to improve the engineering efficiency. There is thus a need for a high-performance signal processing architecture with a stable and engineering efficient API. This project aims to explore the possibilities of utilizing the emerging computational trend of complete frameworks to meet the signal processing demands of the future. Several frameworks belonging to a decades old computing paradigm called stream processing will be examined. Flink and RaftLib are the main candidates being studied. The results show that both frameworks parallelize tasks well and contribute towards engineering efficiency, Flink looking the most promising. RaftLib performing parallel matrix multiplication operations nearly as fast as a custom written thread-pool style benchmark. Flink is faster still, beating even the benchmark in execution time.
11:15-12:00 (E:4130 LUCAS)
Presenter: Frida Hammarberg
Title: Extracting named locations from text using machine learning and natural language processing
Examiner: Christoph Reichenbach
Supervisors: Pierre Nugues (LTH), John Ardelius (Hedvig)
This master thesis aims to investigate different machine learning approaches in order to find named locations in text. Finding named entities, for instance locations, in a text is a problem with increasing popularity. This thesis has been conducted in collaboration with Hedvig, which is an insurance company. For Hedvig, it is important to automatically extract information from text such as where a damage or loss has occurred. Logistic regression, feedforward neural networks and briefly also bidirectional recurrent neural network with long-short-term memory were studied. The best performing model achieves an F1 score of 74.75%, a competitive result in comparison with the state-of-the-art for Swedish that has an F1 score of 77.64% for finding locations in text. The promising result shows that it is possible to build a system that detects locations in text for Swedish text within the insurance domain.
Presenters: Therese Magnusson & Alexander Olsson
Title: Implementing and Evaluating a Breadth-First Search in Cypher
Examiner: Flavius Gruian
Supervisors: Krzysztof Kuchcinski (LTH), Tobias Lindaaker (Neo4j Sweden AB)
This thesis covers the implementation and evaluation of a Breadth-First Search operator for the Neo4j graph database and the Cypher query language. The evaluation compared the pre-existing Depth-First Search operator with both a non-optimized and two different optimized Breadth-First Searches. The focus of this evaluation was the runtime and memory usage for the different operands and to find out for which kind of graph and query that each operand is the better choice. The evaluation was performed through different benchmarks and self-constructed test cases. We concluded that using Breadth-First Search on graph databases is beneficial for smaller datasets where we got improvements up to $90%$ faster and that the Breadth-First Search uses more memory than Depth-First Search. The optimizations gave huge improvements, however they are harder to fit into real world usage. The Breadth-First Search operator should in theory be used on smaller datasets with smaller queries with more restrictions on the relationships. The better the machine the better results. Breadth-First Search should not be applied to the whole path, instead each relationship shall be evaluated case by case.
16:15-17:00 (E:4130 LUCAS) N.B. Change of time and place!
Presenters: Axel Bojrup & Tony Ngo
Title: Machine learning assisted scene detection
Examiner: Jacek Malec
Supervisors: Pierre Nugues (LTH), Sebastian Raase (Sony Mobile Communications AB)
In this thesis, we are exploring the possibility of using machine learning to do scene recognition on a mobile device. The output scenes are then to be mapped to a camera parameter setting, hence the scenes have been based on the Android Camera2 API. A dataset which corresponds to those scenes was created. This was done by investigating different convolutional neural network architectures. We present evaluations on the architectures and show how they perform in terms of speed, accuracy and number of parameters. We demonstrate that it is feasible to use convolutional neural networks to do scene recognition on mobile devices. With a convolutional neural network architecture called MobilenetV1-1, an F1-score of 86.9% was achieved. An ensemble method was also performed, where the collective result achieved a higher F1-score than what any of the networks did separately. Runtime results for inference of various architectures on Sony Xperia Z3 are also presented, ranging from 41 ms to 2750 ms.