Transcript Kein Folientitel - Institut für Grundlagen der Informationsverarbeitung

Neural Networks B 2009
1
Neural Networks B
http://www.igi.tugraz.at/lehre/NNB/SS09/
Lecture 1
Wolfgang Maass
http://www.igi.tugraz.at/maass/
Institut für Grundlagen der Informationsverarbeitung
Technische Universität Graz
Institute for Theoretical Computer Science
http://www.igi.tugraz.at/maass/
Neural Networks B 2009
Which scientific disciplines are involved ?
•
•
•
•
•
Computational Neuroscience
Cognitive Neuroscience
Molecular Biology
Neuroinformatics
Neuromorphic Engineering
2
Neural Networks B 2009
Some links
on http://www.igi.tugraz.at/lehre/NNB/SS09/
•
Tutorials and other teaching material
www.sinauer.com/cogneuro
http://www.science.smith.edu/departments/NeuroSci/courses/bio330/pedsites.ht
ml
http://psych.hanover.edu/krantz/neurotut.html
http://en.wikipedia.org/wiki/Computational_neuroscience
http://www.scholarpedia.org/article/Encyclopedia_of_computational_neuroscienc
e
http://www.scholarpedia.org/article/Encyclopedia_of_cognitive_neuroscience
http://en.wikipedia.org/wiki/Cognitive_neuroscience
http://bluebrain.epfl.ch/
Conferences, Researchers, Resources
http://home.earthlink.net/~perlewitz/
Technological Applications and Related Research Projects
http://facets.kip.uni-heidelberg.de/
http://www.seco-project.eu/
http://reservoir-computing.org/
http://www.eventmakeronline.com/dso/View/index.asp?MeetingID=561
http://www.outlookseries.com/news/Science/3589.htm
http://siliconretina.ini.uzh.ch/wiki/index.php
http://www.smart-systems.at/rd/rd_smart_sensors_de.html
http://www.ini.uzh.ch/
3
Neural Networks B 2009
Difference of course content in comparison with
last years
• Introduction to the PCSIM simulator of biological networks of
neurons (with Python-interface)
• Recent research results on models for cortical micorcircuits
• Inclusion of results, models, and problems of cognitive
neuroscience (memory, top-level-control)
• Discussion of work in related EU-research projects (in which
students could become involved)
• Discussion of results and open problems in neuromorphic
engineering
4
Neural Networks B 2009
Why is it so difficult to understand how the brain works ?
One reason: Different spatial scales are relevant, and in general
mechanisms on different scales interact for each information
processing task
(whereas for a digital computer an algorithm designer does not have to look to
levels below that of a logic gate)
•
•
•
•
•
system level (the whole brain)
brain areas
microcircuits
neurons and synapses
molecular level (including gene regulation)
5
Another reason: Time (on several time scales) plays an
Neural Networks B 2009
essential role for information processing in the brain, in quite
different ways than in computers
Various time scales are relevant, and different processes are not only
superimposed on different spatial scales, but also on different time scales :
•
Weeks/months (replacement of all active neuron components)
•
A day (consolidation of changes of synaptic weights resulting from learning)
•
Minutes (initiation of synaptic weight changes through training)
•
Seconds (behavioural time scale, delay of „rewards“, fMRI)
•
150 – 500 ms (time for completing a fast computation in the brain, also time-scale of
spatio-temporal patterns that encode memory items)
•
100s of ms (short term dynamics of neurons and synapses)
•
10s of ms (integration time for a neuron, learning window for spike-timing-dependent
plasticity: STDP)
•
1- a few ms (spike, relevance of spike-order for STDP)
Analogously as for space, it is also not clear which time scale is „the right one“
for analyzing information processing algorithms of the brain
6
Neural Networks B 2009
A third reason: It is not clear to what extent there exists a
division of labor for information processing tasks of the brain
Older models for information processing in the brain had assumed that the brain
• first builds a model of the external world (e.g. of a visual scence) in the
sensory cortex
• then draws conclusions from that in the association cortices („inference“)
• then initiates motor outputs on the basis of these conclusions in the motor cortex
7
Neural Networks B 2009
More recent experimental data suggest a
quite different perspective
•
The brain does not care to build a model of the external world (and
probably could not even do that for a visual scene)
•
Instead it aims at satisfying certain goals, and actively searches the
sensory inputs for hints how these goals could be achieved in the
current environment
•
Sensory processing and motor processing cannot be separated. Rather
behaviours are encoded as whole entities by the brain (integrating
sensory and motor components).
•
Cortical areas and circuits do not have genetically assigned processing
tasks, rather their computational function emerges on the basis of
genetically encoded learning algorithms and the statistics of their
environment.
8
Neural Networks B 2009
Yet another reason for the difficulty in
understanding how the brain works
• „Innenschau“
• All kinds of heuristic models that we have from
everyday life, old ideas, naive psychology, etc
9
Neural Networks B 2009
Cortical microcircuits of neurons as are the primary
information processing devices in the brain
10
Neural Networks B 2009
The output of neurons consists of „spikes“ in continuous time
(without a „clock“)
Each spike of a presynaptic neuron
causes a temporary change of the
postsynaptic membrane potential
(EPSPs and IPSPs)
The EPSPs and IPSPs are added
linearly, and the neurons emits a
spike whenever the postsynaptic
membrane reaches the
„firing threshold“
11
Neural Networks B 2009
A protocol of a neural computation („spike raster“)
Shown here are the times when spikes are emitted by neurons in
the primary visual cortex of a cat (when it was shown the letters A, D)
12
Neural Networks B 2009
Both the timing of spikes and the „firing rates“ of neurons
change from trial to trial
Shown are spike rasters for 5 trials with each of the two stimuli A and D.
13
Neural Networks B 2009
Program of this course for the next few weeks
• Visualization of a model for a small network of spiking
neurons
• Mathematical models for spiking neurons
• Why are neurons and synapses so difficult to model ?
(a brief look at their molecular components)
• Methods for studying how the brain works
14