Seminar: Machine Learning and Artificial Neural Networks in Biomedical Applications

Lehrstuhl Technische Informatik
Dozent Prof. Dr. Bogdan
Betreuer Dr. Walter
Walter
Dr. Sp├╝ler
Peter
Dobler
Sippel
Dr. K├╝bler
Hettich
Vorbesprechung         21.10.2013, 10.00 (s.t.), B226, Sand 13
Umfang 2 SWS / 4 LP
Eintrag im LSF Machine Learning and Artificial Neural Networks in Biomedical Applications

Beschreibung


The Seminar "Machine Learning and Artificial Neural Networks in Biomedical Applications" covers current topics of signal processing on neural signals (e.g. fMRI, EEG or MEG) for their use in biomedical applications (e.g. neuroprosthetics or brain-computer interfaces, BCIs) and related topics; as well as methods and algorithms applied in those fields.

 

Im Seminar "Maschinelles Lernen und K├╝nstliche Neuronale Netze in der biomedizinischen Anwendung" werden aktuelle Themen aus der Signalverarbeitung im Bereich der Verarbeitung von Nervensignalen (z.B.: Neuroprothetik oder Brain-Computer-Interfaces), medizinischer Signalen (z.B.: fMRT oder MEG) oder verwandten Bereichen sowie in diesen Bereichen verwendeten Algorithmen der Signalverarbeitung bearbeitet.

 

Themen:

  • Dynamic Causal Modeling Analysis of Simultaneous EEG and fMRI Data during Reward Anticipation: Dynamic causal modeling (DCM) is a method for the interpretation of functional neuroimaging data. The goal is to infer the causal architecture of coupled or distributed dynamic systems formulated as stochastic or ordinary differential equations. In a Bayesian model comparison procedure of such hypothesized models, the most likely explanation of how the recorded data were generated by the model is sought after.
  • A close-loop human simulator for investigating the role of feedback control in brain-machine interfaces: Decoding algorithms for brain-machine interfaces are typically evaluated "offline", using neural activity previously gathered from a human/animal an the decoded movement is then compared to the real movement that was executed while the neural data was recorded. However, using this offline evaluation method neglects the role feedback has in real brain-machine interfaces (BMIs). And thereby certain properties of a decoding algorithm can not be evaluated, i.e., how easy is it for the user to adapt to the BMIs. To better understand the role of feedback, a close-loop human simulator has been proposed that simulates neural activity based on human interaction and thereby allows to evaluate decoding algorithm under closed-loop conditions and how the feedback influences BMI performance.
    For this seminar, a basic introduction into the topic of brain-machine interfaces should be given and the paper regarding the closed-loop human simulator should be presented.
  • Problem solving, working memory, and motor correlates of association and commissural fiber bundles in normal aging (NeuroImage, Volume 44, Issue 3; Natalie M. Zahra, Torsten Rohlfing, Adolf Pfefferbaum, Edith V. Sullivan): The search for brain mechanisms underlying complex cognitive, motor, and other behavioral functioning has shifted from single structures or loci to systems and circuits. Indeed, numerous functional MR imaging studies suggest that multiple brain regions are invoked to execute even the simplest sychomotor tasks. Until recently, the study of white matter bundles was restricted to postmortem analysis or slice-by-slice evaluation of structural MRI data. By contrast, magnetic resonance diffusion tensor imaging has enabled noninvasive in vivo assessment of the integrity of brain connectivity by detecting the microenvironment of white matter. In the presented paper they used quantitative fiber tracking in 12 young and 12 elderly healthy men an women and tested functional correlates with concurrent assessment of a wide range of neuropsychological abilities. Principal component analysis of cognitive and motor tests on which the elderly achieved significantly lower scores than the young group was used for data reduction and yielded three factors: Problem Solving, Working Memory, and Motor.
  • Kriging: One main branch of Machine Learning is regression, i.e. the prediction of function values given a few samples. The CS curriculum includes regression methods like Linear Regression, Neural Networks and Support Vector Machines. Kriging is another method, that is not as well-known by CS students. It was originally developed for geology, because it compensates clusters of measurements which often arise with field measurement. The prediction is a linear combination of all known sample values, where the weights depend on the location of the known samples and the location that is to be predicted. Variants of Kriging are also known as Best Linear Unbiased Prediction or Gaussian Process.
  • Multimodal Saliency-based Attention modeling: (Schauerte, B., & Kuhn, B. (2011). Multimodal saliency-based attention for object-based scene analysis. Intelligent Robots and ...) A computer model for prediction of human attention by spectral saliency and surprise, based on visual and auditory input is introduced. It reaches state-of-the-art performance in gaze prediction. The method is based on a quanternion PCA in order to build a saliency map: a map of bottom-up attractiveness of certain objects in the visual field.
  • Decision making in the ICU with RBF-based fuzzy logic: Fuzzy Logic can be applied in automated medical decision making in the intensive care unit (ICU) by providing a human-readable and validatable set of rules. To join these advantages with the adaption potential of neural networks, RBF-based networks may be converted to a set of such rules. For this seminar, the fundamental concepts of fuzzy logic and the general idea of incooperating RBF-networks to allow for dynamic adaption in the context of medical application should be presented.

Bemerkung

We'd welcome if you pre-register for this course. Please send a mail with your name, student id ("Matrikelnummer"), branch of study (CS / bioinformatics / ...) and how far you've progressed in your studies to dobler@informatik.uni-tuebingen.de.

The topics will be assigned at the kickoff-meeting.

Voranmeldung zum Seminar ist erw├╝nscht: E-Mail an dobler@informatik.uni-tuebingen.de mit Namen, Matrikelnummer, Studiengang und Semesterzahl.

Die Vergabe der Themen erfolgt auf der Vorbesprechung.

Lernziele    

  • Vermittlung von Faktenwissen
  • wissenschaftliche Literatur zu bestimmten Themen suchen
  • sich in die Terminologie eines Themengebiets einlesen
  • einen gut strukturierten und informativen Vortrag erarbeiten
  • diesen Vortrag sicher und f├╝r die Zuh├Ârer interessant vortragen
  • einen eng begrenzten Zeitrahmen f├╝r einen Vortrag einhalten
  • konstruktive Kritik zu einem Vortragsstil geben und selbst bekommen
  • eine wissenschaftliche Ausarbeitung schreiben
  • wichtig f├╝r das Seminar: Den Inhalt verstehen und an die Zuh├Ârer weitergeben

Der Vortrag sollte im Idealfall genau 20 Minuten dauern, nicht l├Ąnger. An diese Redezeit schlie├čt sich eine kurze Diskussion ├╝ber sowohl den Inhalt als auch den Vortragsstil an, soda├č insgesamt 30 Minuten pro Beitrag zur Verf├╝gung stehen. Damit ein Vortrag f├╝r die ├╝brigen Zuh├Ârer informativ und interessant ist, mu├č er einer gewissen Form gen├╝gen. Jeder Teilnehmer sollte deshalb fr├╝hzeitig mit seinem Betreuer in Kontakt treten.

Es wird regelm├Ą├čige Treffen mit dem Betreuer geben, um die aktuelle T├Ątigkeit zu besprechen.

Milestones

  • Besprechung der Gliederung mit Betreuer: ca. Anfang - Mitte Dezember
  • Fertige Ausarbeitung: Ende Januar
  • Pr├Ąsentation: tba, Ende Vorlesungszeit
  • Probevortrag: ca. eine Woche vor dem Seminartermin.

 

Zus├Ątzliche Literatur zu Eurem Thema sollte eigenst├Ąndig recherchiert und aufbereitet werden, Einstiegspunkte werden von den Betreuern vorgeschlagen.

Die Ausarbeitung sollte 15-20 Seiten umfassen, in Latex erstellt werden, und im Format einer wissenschaftlichen Ver├Âffentlichung vorgelegt werden (Beispiel), bei der s├Ąmtliche Quellen referenziert werden. Weitere Hinweise zur Ausarbeitung finden sich hier.