Skip to main content

Program

SweDS 19's conference program includes invited presenters and keynote speakers.

Program

Venue

SweDS19 is held in room Q1, Malvinas väg 4, Q-huset, floor 3, KTH Campus.

Map to location

15 October

08:30 - 09:00 Registration

09:00 - 09:05 Welcome and SweDS19 opening
Information (pdf 5.0 MB)

09:05 - 10:00 Keynote: Professor Virginia Dignum (Umeå University)

10:00 - 10:40 Posters and coffee (see poster information further down)

10:40 - 11:50 Presentation session A

  1. Ruben Buendia, Halsey Lea, Bo Zhang, Faisal Khan and Glynn Dennis "Replacing Site-Based Exercise Tests with Wearable Activity Monitors to Reduce Patient Burden and Time to Signal Detection"
  2. Elena Malakhatka "Big data ecosystem for smart and sustainable residential buildings: System boundaries and reference architecture"
  3. György Kovács, Foteini Liwicki and Marcus Liwicki "Towards automatic support of mental well-being assessment with audio-visual surveying"

11:50 - 13:30 Lunch at campus

13:30 - 14:00 Invited presentation: Dr. Meg Schedel (Stony Brook University, NY)

14:00 - 14:30 Invited presentation: Dr. Anders Arpteg (Peltarion, Stockholm)

14:30 - 15:00 Invited presentation: Professor Michael Höhle (Stockholm University)

15:00 - 15:30 Fika and posters (see poster information further down)

15:30 - 16:40 Presentation session B

  1. Henrik Imberg, Marina Axelson-Fisk and Johan Jonasson "Optimal sampling in unbiased active learning"
  2. Lorenzo Andraghetti, Panteleimon Myriokefalitakis, Pier Luigi Dovesi, Belen Luque, Matteo Poggi, Alessandro Pieropan and Stefano Mattoccia "Enhancing self-supervised monocular depth estimation with traditional visual odometry"
  3. Pier Luigi Dovesi, Matteo Poggi, Lorenzo Andraghetti, Miquel Marti, Hedvig Kjellström, Alessandro Pieropan and Stefano Mattoccia "Real-Time Semantic Stereo Matching"

16:40 - 16:45 Day 1, wrap up

20:00 - 21:00 Free concert at KMH ("Music/Data -> Data/Music")

16 October

09:00 - 09:30 Invited presentation: Professor Sven Ahlbäck (KMH, DoReMIR)

09:30 - 10:00 Invited presentation: Dr. Josephine Sullivan (KTH)

10:00 - 10:40 Posters and coffee (see poster information further down)

10:40 - 11:50 Presentation session C

  1. Edvin Listo Zec and Olof Mogren "Grammatical gender of Swedish nouns are predictable with recurrent neural networks"
  2. Thomas Hörberg "The processing of grammatical functions in Swedish is expectation-based"
  3. Thomas Hörberg and Jonas Olofsson "The semantic organization of the English odor vocabulary"

11:50 - Lunch at campus

Keynote speaker

Virginia Dignum

Professor Virginia Dignum

Department of Computing Science, Umeå University

Virginia Dignum's personal webpage

Title: Responsible Artificial Intelligence

Description: As Artificial Intelligence (AI) systems are increasingly making decisions that directly affect users and society, many questions raise across social, economic, political, technological, legal, ethical and philosophical issues. Can machines make moral decisions? How should moral, societal and legal values be part of the design process? In this talk, I look at ways to ensure that behaviour by artificial systems is aligned with human values and ethical principles. Given that ethics are dependent on the socio-cultural context and are often only implicit in deliberation processes, methodologies are needed to elicit the values held by designers and stakeholders, and to make these explicit leading to better understanding and trust on artificial autonomous systems. We will in particular focus on the ART principles for AI: Accountability, Responsibility, Transparency.

Invited presentations

Dr. Anders Arpteg

Anders Arpteg

Head of Research, peltarion.com
Title: The unreasonable effectiveness of using deep learning for NLP

Description: A significant step change in AI happened back in 2012 for computer vision with deep learning models such as AlexNet. Last year, in 2018, another significant step change happened in AI for NLP. New deep learning models such as Google BERT and OpenAI’s GPT2 demonstrated the unreasonable effectiveness in being able to have pre-trained generic models that could be used not only for a specific task, but that could be fine-tuned to break state-of-the-art for many NLP tasks using the same pre-trained model. In this talk, I will give a quick introduction to what has happened in this field, and also talk specifically about a project we have had at Peltarion that makes use of the latest AI techniques for NLP and how surprisingly well these techniques work.

Bio: Anders Arpteg has been working with artificial intelligence for 20 years in both academia and industry, with a PhD in AI from Linköping University, Sweden. Previously headed a research group at Spotify making use of big data and machine learning techniques to optimize user experience. Now working with the latest AI techniques at Peltarion as Principal Data Scientist, where we have the ambitious goal of making deep learning and the latest AI techniques available for all, not just the large technology giants. Also founder of Agent Central AB, AI adviser for the Swedish government, organizer of Machine Learning Stockholm meetup group, and member of several advisory boards.

Professor Sven Ahlbäck

Sven Ahlbäck

Kungliga Musikhögskolan, CEO på Doremir Music Research AB

Title: What is the ground truth in music modeling? A musician’s approach to computer modelling in the crossroads between music practice and computer engineering. Do we need music theory when we have machine learning? What challenges does music practice present when applying computer modelling?

Description: Thirty years ago, fiddle player Sven Ahlbäck’s doctoral project in Musicology took a new turn, when he met prof. Johan Sundberg at KTH, who advised Sven to collaborate with someone with computer engineering skills to be able to model perception and cognition of music structure. Sven’s project was originally about tonality, i.e. the perception of hierarchical pitch structure in music, specifically in the obscure style of Swedish Folk Music, but in order to get there he needed the basic structure of the music, including accent and beat structure, phrase segments etc. He met a master student in computer science, also named Sven, who became interested in the subject and together they started working on a project of automatically determining segment structure in music, estimating some 3-4 months for the entire project. However, not until 15 years later Sven defended his thesis in musicology and have had to learn a bit of coding by then, to build the model that was an important foundation of his thesis, basically a gestalt psychological and cognitive psychology approach to music. Later on, Sven started the company Doremir together with his original collaborator Sven Emtell, since then developing e.g. the award-winning software ScoreCloud for music notation, building on his original research in music cognition. In this talk, Sven Ahlbäck will discuss the need for technological and scientific approaches, including computer modelling, in relation to music study and practice, as well as principal and practical pitfalls and challenges that may appear in the application of technological approaches to art practice – and the other way around.

Professor Michael Höhle

Department of Mathematics, Stockholm University

Michael Höhle

Title: Finding the needle in the epidemiological haystack - monitoring time series of counts”

Description: Digitalization in healthcare has lead to an increased amount of data being collected by public health authorities. A common aim of this data collection is to obtain information on particular health events of interest in order to monitor the occurrence of these events over time. Possible applications are the monitoring of adverse drug reactions, detection of disease outbreaks and the quality assurance of hospital procedures. Because several drugs, diseases or quality indicators are monitored simultaneously for different entities such as regions and hospitals the data can, from a statistical perspective, be represented as multivariate time series of counts.

The statistical challenge is to design algorithms, which are able to detect aberrations in these time series. We illustrate the problem by work from the German monitoring system for infectious diseases, which routinely monitors more than 100,000 count time series arising from different combinations of pathogen, age, sex and region. The example is contrasted with German quality assurance of hospitals, where approximately 300 quality indicators are tracked for each hospital. In both cases we discuss the challenges of using algorithmic decisions in the workflow of public health institutions, e.g., the role uncertainty, validation of the decisions and user acceptance.

Bio: Michael Höhle works part-time as a professor in mathematical statistics at Stockholm University. His research interests cover biostatistical methods in general and statistical methods for infectious disease epidemiology in particular. He is the initiator of the R package ‘surveillance’ available from the Comprehensive R Archive Network (CRAN) and has a general interest in getting statistical methodology implemented in practice – a topic which he blogs about in ‘Theory meets Practice’ (http://staff.math.su.se/hoehle/blog/).

Dr. Josephine Sullivan

Associate Professor, Division of robotics, perception and learning, KTH

Title: Multi-modal Processing: Vision and language

Josephine Sullivan

Description: In the last 5 years the intersection of vision and language processing has become a fertile area of research producing some impressive results: The automatic description of images with non-trivial text captions (extending beyond the description of objects present) and answering non-trivial text based questions about visual images. The advancements made have embraced one of the main successes of deep learning - its ability to learn powerful feature representations. These learnt feature representations have enabled the translation between the modalities of text and vision. In this talk I will give an overview of theses developments which make use of attention, highlight some of the perils of being too excited by the results and review our current research where we have shown the potential benefits of learning feature representations for images and text jointly to allow for concept grounding. 

Dr. Margaret Schedel

Margaret Schedel

Associate Professor, Stony Brook University, NY, USA

Title: From Logo to Logopenic: The Sound of Data

Description: Data sonification exists along a continuum of sound for the sake of data, to sound for the sake of music. In each case practitioners blend aesthetic and scientific data to create sonic material. In this talk I detail how a scientific project, sonifying nano materials, resulted in a scientific project in use at Brookhaven National Laboratories, a VR experience for a science museum, and a new piece of acoustic music premiering at this conference. I explain the algorithm behind this particular sonification, and the process of turning 10,000 audio samples into sound design and how I then turned those samples into a new music composition for string trio, horn and percussion.

Bio: With an interdisciplinary career blending classical training in cello and composition, sound/audio data research, and innovative computational arts education, Margaret Anne Schedel transcends the boundaries of disparate fields to produce integrated work at the nexus of computation and the arts. Her research focuses on gesture in music, the sustainability of technology in art, and sonification of data; she co-authored a paper published in Frontiers of Neuroscience on using familiar music to sonify the gaits of people with Parkinson's Disease. She serves as a regional editor for Organised Sound and is an editor for the open access journal Cogent Arts and Humanities. Schedel currently serves as the co-director of computer music and leads the Making Sense of Data Workgroup at the Institute of Advanced Computational Science at Stony Brook University. She also teaches composition for new media at the Peabody Institute of the Johns Hopkins University. 

Poster sessions

  1. Sara Abbaspour, David Krýsl, Fredrik Ohlsson, Jan Wipenmyr and Kristina Malmgren "Classification of tonic–clonic seizures using wearable accelerometer sensors"
  2. Sophie Grape, Erik Branger, Zsolt Elter, Li Caldeira Balkeståhl and Vaibhav Mishra "Machine learning in nuclear safeguards"
  3. Mobyen Uddin Ahmed, Shaibal Barua, Shahina Begum, Ekrem Güclü, Manasi Jayapal and Sharmin Sultana Sheuly "Intelligent Data Analytics for Maintenance in Industry 4.0"
  4. Susanna Pozzoli, Amira Soliman, Leila Bahri, Rui Mamede Branca and Sarunas Girdzijauskas "Domain Expertise–Agnostic Feature Selection for the Analysis of Breast Cancer Data”
  5. Rickard Brännvall and Tor-Björn Minde, “National Space Data Lab”

  6. Carl Dath, Carl Nyströmer, Jakob Vyth & Joel Ekelöf, "Investigating Hit Song Prediction"

Belongs to: School of Electrical Engineering and Computer Science
Last changed: Oct 14, 2019