You are here

ML Seminar Spring 2013

Machine Learning Seminar and Project

(= Machine Learning A, 320542 and 320544)

Jacobs University Bremen, Spring 2013, Herbert Jaeger

First session: Monday Feb 4, 15:45, West Hall 8 (we will then shape the schedule as we like it)

Topic:  Mechanisms of "Attention" in Recurrent Neural Networks

The seminar will explore recently discovered mechanisms for controlling and focussing the information processing in recurrent neural networks (RNNs). Intuitively, think of these mechanisms as "attention control", "focussing one's thoughts", "actively suppressing irrelevant information", "concentration, please"... and more in this spirit. Join the first session if this sounds riddlesome and interesting enough to tickle your curiosity...

The seminar will have two lines of activity, done in parallel. The first line is ordinary seminar work, that is, we will regularly meet for educational sessions where

  • either I simply teach you the essentials of RNNs and some mathy topics that we'll be needing (mostly linear algebra and dynamical systems, as opposed to the probability and statistics load that you sustained in the Fall lecture),
  • or we study original papers from the literature (cognitive neuroscience, AI, robotics, machine learning) that are relevant for us; then all participants will be required to study the paper in advance, with a set of questions given by me to be addressed before the actual meeting.

The second line is hands-on experiments with RNNs, using Matlab. After a warmup phase where you will get up to speed with handling RNNs in general, a reasonably challenging semester project will be formulated (whether each participants gets an individual project, or whether we'll have teams, or whether we'll in fact all team up together - to be decided).

Grading

Two separate grades will be given for the theory/reading part (course Nr 320442) and the implementation project (320444). Each is a co-requisite for the other.  The theory/reading grade will be determined on the basis of seminar presentations and active participation in discussions. The implementation project grade will depend on a written report and a demonstration of the implementation. Grading details will be discussed early in the semester and then documented on this page.

Prerequisites

This is a highly interdisciplinary seminar, and I invite students from other majors than Smart Systems or Cosyp to consider participation. Relevant backgrounds are

  • Machine learning
  • Signal processing and control
  • Mathematical modeling in general, dynamical systems theory in particular
  • Cognitive science
  • Robotics
  • Neuroscience

If you have a solid footing in at least one of these fields please drop by for a personal conversation to find out whether entering this seminar would make sense.

If there are two or one registrations only, the course will not take place. If there are 3 or 4, it is my descretion to run the course or not. With 5 or more registrations the course will surely happen.

Contact

Please feel kindly invited to contact me (Herbert Jaeger) if you are interested.