BigData@Chalmers logo

Machine Learning Workshop


Thank you all who attended the Machine Learning Workshop on 14 April 2016. Below you will find speaker presentations.

During this one day workshop, researchers in the Gothenburg area (and guests) shared with us how they use machine learning to solve complex research questions in medicine, transport, biology, language technology and urban planning.

Organizer: The BigData@Chalmers Initiative
Contact: hans.salomonsson@chalmers.se

Programme


9.00 Welcome
9.05 Introduction to Machine Learning
Fredrik Johansson, PhD Student, Computer Science and Engineering, Chalmers
9.30 Learning to Interpret Medical Images
Fredrik Kahl, Professor, Signals and Systems, Chalmers
10:15 Coffee & Sandwich
Irene Yu-Hua Gu, Professor, Signals and Systems, Chalmers
11:15 Synthetic information systems for high performance simulations of complex systems
Chris Barrett, Professor and Scientific Director, Virginia Biocomplexity Institute
12:00 Lunch (not provided)
13:00 High dimensional statistics for analysis of biological networks
Rebecka Jörnsten, Associate Professor in Mathematical Statistics, Mathematical Sciences, Chalmers
Shalom Lappin, Professor Computational Linguistics, Gothenburg University
14:30 Coffee & Cake
14:45 Small cars, big data?
Henk Wymeersch, Associate Professor, Signals and Systems, Chalmers
Hans Salomonsson Big Data Application Expert, Chalmers


Abstracts

Introduction to Machine Learning
Fredrik Johansson, PhD Student, Computer Science and Engineering, Chalmers

The ability to learn from and infer properties of our surroundings is one of the great reasons for the success of the human species. Machine learning tries to repeat this feat in artificial systems that automatically improve through experience. Building on, and in many cases leading innovation in optimization, statistics and computer hardware, this area has grown into one of the major driving forces behind both research and industry, and is still rapidly increasing in popularity. Recent progress, partly due to the availability of large datasets of images, text and speech recordings, has shown that machine learning methods are sufficiently mature to apply to a wide variety of tasks, in applications ranging from health care and robotics to playing video games.


Learning to Interpret Medical Images
Fredrik Kahl, Professor, Signals and Systems, Chalmers
In this talk, I will present some ongoing research work at the Computer Vision Group (Chalmers) using machine learning for interpreting medical images. In particular, I will focus on a new framework based on deep convolutional neural networks for localizing and segmenting organs and other anatomical structures in CT images.

Machine Learning methods and their applications to autonomous driving and e-healthcare using video sensor data
Irene Yu-Hua Gu Professor, Signals and Systems, Chalmers
There has been a growing research interest in autonomous driving and healthcare. In this presentation, we focus on techniques based on video/image analysis and machine learning methods using multi-camera (RGB, D, IR) sensors. First, we will show several recent applications from our research on i) Road traffic sign recognition using street view videos for autonomous driving and driving assistance systems; ii) Privacy-preserving fall detection and activity recognition for e-healthcare; iii) Object tracking for road traffic analysis and for medical applications. We will then describe several methodologies behind these applications. These include target object characterization by effective spatio-temporal features, Riemannian manifold-based visual object tracking, as well as multiclass learning and classification. (Presentation slides, Irene Gu)

Synthetic information systems for high performance simulations of complex systems
Chris Barrett, Professor and Scientific Director, Virginia Biocomplexity Institute
We will give an overview of an approach to high performance computing (HPC) simulations of large scale complex socially coupled technological systems using synthetic information systems. Chalmers is part of a new EU Horizon 2020 Center of Excellence in HPC for Global Systems Science and a project on analysis of electric vehicle (EV) adoption in Sweden has been initiated.

Network modeling of TCGA data: integrative and disease comparative approaches
Rebecka Jörnsten, Associate Professor in Mathematical Statistics, Mathematical Sciences, Chalmers
Statistical network modeling techniques have the potential to increase our understanding of cancer genomics data. Here, we analyze multiple TCGA data sets via a generalized sparse inverse covariance model, carefully addressing such challenges as unbalanced sample sizes, local network topology, model selection and robust estimation. The method integrates genetic, epigenetic and transcriptional data from multiple cancers, to define links that are present in multiple cancers, a subset of cancers, or a single cancer. The modeling results are available at cancerlandscapes.org, where the derived networks can be explored as interactive web content and be compared with several pathway and pharmacological databases. Network components are shown to fall in mainly two categories: common to all cancers or unique to one type of cancer. We also discuss how network models can be used to construct diagnostic markers (predictors of survival).

Do We Need Grammar for Natural Language Semantics?
Shalom Lappin, Professor Computational Linguistics, Department of Philosophy, Linguistics, Theory of Science, GU
On the classical view of semantics, formal grammars assign hierarchical syntactic structures to natural language phrases and sentences. While formally elegant this tradition has not yielded robust, wide coverage computational treatments of natural language meaning. The current success of deep neural networks in generating descriptions of graphic images suggests an alternative strategy. We can formulate semantic interpretation as a problem in machine translation from natural language phrases and sentences to multimodal representations consisting of images, video sequences, speech and audio elements, text, and data structures combining some or all of these objects. To the extent that such a language model is successful it will achieve the traditional requirement on a semantic theory that it assign truth or satisfaction conditions to each sentence of a language. (Presentation slides, Shalom Lappin)

Small cars, big data?
Henk Wymeersch, Associate Professor, Signals and Systems, Chalmers
As autonomous cars are entering society, new possibilities emerge by wirelessly connecting these cars with other cars and the traffic infrastructure. Our ongoing research deals with microscopic and macroscopic traffic control, and includes traffic monitoring, infrastructure control, and inter-vehicle coordination. Due to the possibly massive amounts of data exchanged (a car generates about 1 gigabyte every second), storage, learning from data, and communication of information is challenging. In this talk, we will give a broad overview of our current research challenges and tools.


Machine Learning Tools And Libraries
Hans Salomonsson, Big Data Application Expert, Chalmers
Machine learning libraries and numerical libraries, such as BLAS, are fundamental for working effectively with machine learning. But do we really need to program at a low abstraction level? What happens when we go up in abstraction levels? Can we gain more flexibility when designing our models, in less time and still keep almost the raw performance of writing the code in e.g. C? We will cover what might be a good setup in terms of flexibility and time/money/knowledge constraints for your project. Moreover, some hints will be given for those interested in implementing their own high performance (machine learning) library and automatically create wrappers for many programming languages. (Presentation slides, Hans Salomonsson)


Published: Wed 20 Apr 2016. Modified: Thu 21 Apr 2016