connected populations: oscillations, competition and spatial continuum (field equations) lecture 12...
Post on 22-Dec-2015
220 views
TRANSCRIPT
Connected Populations: oscillations, competition and spatial continuum (field equations)
Lecture 12Course: Neural Networks and Biological Modeling
Wulfram Gerstner, EPFL
Book: W. Gerstner and W. Kistler, Spiking Neuron Models Chapter 9.1
Population of neurons
h(t)
I(t) ?
0t
A(t) ))('),(()( tItIgtA
))()(())(()( dsstIsgthgtA A(t)
A(t) ))(()( tIgtA
potential
A(t) ))(()()( thgtAtAdt
d
t
2 2( ; )( )
t tn t tA t
N t
populationactivity
Blackboard:Pop. activity
Signal transmission in populations of neurons
Connections4000 external4000 within excitatory1000 within inhibitory
Population- 50 000 neurons- 20 percent inhibitory- randomly connected
-low rate-high rate
input
Population- 50 000 neurons- 20 percent inhibitory- randomly connected
Signal transmission in populations of neurons
100 200time [ms]
Neuron # 32374
50
u [mV]
100
0
10
A [Hz]
Neu
ron
#
32340
32440
100 200time [ms]50
-low rate-high rate
input
Theory of transients
low noise
I(t)h(t)
noise-free
(escape noise/fast noise) noise model A
low noise
fast transient
noise model A
I(t)h(t)
(escape noise/fast noise)
high noise
slow transient
))(()( tIgtA
But transient oscillations
))(()( thgtA
High-noise activity equation
noise model A
I(t)h(t)
(escape noise/fast noise)
high noise
slow transient
))(()( thgtA
))(()( thgtA
)()()( tIRththdtd
Population activity
Membrane potential caused by input
blackboard
In the limit of high noise,
Full connectivity
Part I: Single Population - Population Activity, definition - high noise - full connectivity
Fully connected network
N
Jwij
0 f
jtt j f
ijnet wtI )(
All spikes, all neurons
fjtt
Synaptic coupling
fullyconnected N >> 1
)()()( tIRththdtd
)()()( tItItI netext
blackboard
N
Jwij
0
)(tItt extfj
j fiji wtI )(
All spikes, all neurons
)(tI ext dsstAsJtI i )()( 0 fullyconnected
All neurons receive the same total input current (‘mean field’)
)()()( tIRththdtd
Index i disappears
Stationary solution)( 0hg
0h
0A
fullyconnected N >> 1
Homogeneous network, stationary,All neurons are identical,Single neuron rate = population rate
0 0( )g h A
( ) ( ) ( ) ( )ext netwddt h t h t R I t R I t
0( ) ( ) ( ) ( ) ( )extddt h t h t R I t R J ds s A t s
0( ) 0 ( )ddt h t h t h
Exercises 1, now
Next lecture 10:15
Exercise 1: homogeneous stationary solution
)( 0hg
0h
0A
fullyconnected N >> 1
Homogeneous networkAll neurons are identical,Single neuron rate = population rate
A
Connected Populations:Part I: Single Population - Population Activity, definition - high noise - full connectivity - stationary state - mean rate in stationary state
0 0( )A g hWhat is this function g?
Wulfram Gerstner, EPFL
Stationary State/Asynchronous State
1 s
fully connected coupling J/N tt ˆ ttu ˆ| 0h
input potential
1
0
ˆ ˆ|Is P t s t ds
)( 0hg
)( 0hg
extIAqJRRIh 00000
frequency (single neuron)
qdss
blackboard
Homogeneous networkAll neurons are identical,Single neuron rate = population rate
A(t)= A0= const
Single neuron
High-noise activity equation
noise model A
I(t)h(t)
(escape noise/fast noise)
high noise
slow transient
))(()( thgtA
))(()( thgtA
)()()( tIRththdtd
Population activity
Membrane potential caused by input
Note: total membrane potential
)ˆ()()( ttthtu
Effect of last spike
0 0( )A g hWhat is this function g?
Spike Response Model
iuij
fjtt
Spike reception: EPSP
fjtt
Spike reception: EPSP
itt ˆ
itt ˆ
Spike emission: AP
fjtt itt ˆ tui
j fijw
tui Firing: tti ˆ
Spike emission
Last spike of i All spikes, all neurons
hi(t)
)()()( tIRthth iiidtd
Fully connected network
N
Jwij
0
Spike emission: AP
fjtt ( )netw
ijj f
I t wAll spikes, all neurons
fjtt
itt ˆ
Synaptic coupling
fullyconnected N >> 1
)(ˆ thtt ii tui
)()()( tIRththdtd
( ) ( ) ( )ext netwI t I t I t
0( ) ( ) ( )netwI t J s A t s ds
Stationary State/Asynchronous State
1 s
fully connected coupling J/N
tt ˆ ttu ˆ| 0h
input potential
1
0
ˆ ˆ|Is P t s t ds
)( 0hg
)( 0hg
0h
extIAqJRRIh 00000
0A
A(t)=const
extRIhqRJ
A 000
01
frequency (single neuron)
typical mean field (Curie Weiss)
qdss
Homogeneous networkAll neurons are identical,Single neuron rate = population rate
High-noise activity equation
noise model A
I(t)h(t)
(escape noise/fast noise)
high noise
slow transient
))(()( thgtA
))(()( thgtA
)()()( tIRththdtd
Population activity
Membrane potential caused by input
)()()( tItItI netext
)()()( 0 tAqJtItI ext
))(()()( 0 thgqJtItI ext
))(()()()( thgtIRthth extdtd
1 population = 1 differential equation
Connected Populations:Part I: Single Population - Population Activity, definition - high noise - full connectivity - stationary state - mean rate in stationary state
Part II: Multiple Population - Spatial coupling - Background: cortical populations - Continuum model - Stationary state
Wulfram Gerstner EPFL
Microscopic vs. Macroscopic Coupling
I(t)
)(tAn
Cortical columns:Orientation tuning (and coarse coding)
Detour: Receptive fields (see also lecture 4)
visual cortex
electrode
Detour: Receptive field development
visual cortex
Detour: Receptive field development
Receptive fields: Retina, LGN
-+
-
Detour: Receptive field development
Receptive fields: Retina, LGN
Receptive fields: visual cortex V1
Orientation selective
Detour: Receptive field development
Receptive fields: visual cortex V1
Orientation selective
Detour: orientation selective receptive fields
Receptive fields: visual cortex V1
Orientation selective
2
0
rate
Stimulus orientation
Detour: Receptive fields
visual cortex
Neighboring cells in visual cortexHave similar preferred orientation:
cortical orientation map
Detour: orientation selective receptive fields
Receptive fields: visual cortex V1
Orientation selective
2
0
rate
Stimulus orientation
Cell 1
Detour: orientation selective receptive fields
0
rate
Cell 1 Cell 5
2
Oriented stimulus
Course coding
Many cells respond to a single stimulus with different rate
Continuous Networks
Continuous Networks
)(tAi
Several populations
i k
Continuum
Blackboard
Part I: Single Population - Population Activity, definition - high noise - full connectivity - stationary state - mean rate in stationary state Part II: Multiple Population - Background: cortical populations - spatial coupling - continuum model - stationary state - application: head direction cells
Field equations and continuum Models for coupled populations
)(),( AtA
Continuum: stationary profile
Head direction cells (and line attractor)
Neurophysiology of the Rat Hippocampus
rat brain
CA1
CA3
DG
pyramidal cells
soma
axon
dendrites
synapses
electrodePlace fields
Hippocampal Place Cells
Main property: encoding the animal’s location
place field
Head-direction Cells
Main property: encoding the animal ’s heading
Preferred firing direction
r (i
i
Head-direction Cells
Main property: encoding the animal ’s allocentric heading
Preferred firing direction
r (i
i
0
90
180
270300
Basic phenomenology
0
A(theta)
I: Bump formation
strong lateral connectivity
Possible interpretation of head direction cells: always some cells active indicate current orientation
Field Equations:Wilson and Cowan, 1972
Basic phenomenology
x0
A(x)
II. Edge enhancement
Weaker lateral connectivity
I(x)
Possible interpretation of visual cortex cells: contrast enhancement in - orientation - location
Field Equations:Wilson and Cowan, 1972
)(),( AtA
Continuum: stationary profile
Comparison simulation of neurons and macroscopic field equation
Spiridon&Gerstner
See: Chapter 9, book:Spiking Neuron Models,W. Gerstner and W. Kistler, 2002
Exercises 2,3,4 from 11:15-12:45
Exercise 1: homogeneous stationary solution
Exercise 2: bump solution
Exercise 3: bump formation
)( 0hg
0h
0A
Blob forms where the input is
x
0A
x
0A
Part I: Single Population - Population Activity, definition - high noise - full connectivity - stationary state - mean rate in stationary state Part II: Multiple Population - Background: cortical populations - spatial coupling - continuum model - stationary statePart III: Beware of Pitfalls - oscillations - low noise
Field equations and continuum Models for coupled populations
Book: Spiking Neuron Models.W. Gerstner and W. Kistler
Chapter 8
Oscillations
Stability of Asynchronous State
Search for bifurcation points
linearize
tdtAttPtA I
t
ˆ)ˆ(ˆ|)(
)()( 0 tAAtA )()( 0 thhth dsstAsJth )()()(
h: input potential
A(t)
ttieAtA 1)(
0
fully connected coupling J/N
Stability of Asynchronous State A(t)
delayperiod
)()( sess0 for
stable0
03
02
noise
T
)(s
s
)(s
(Gerstner 2000)
High-noise activity equation
noise model A
I(t)h(t)
(escape noise/fast noise)
high noise
slow transient
))(()( thgtA
))(()( thgtA
)()()( tIRththdtd
Population activity
Membrane potential caused by input
))(()()()( thgtIRthth extdtd
Attention: - valid for high noise only, else transients might be wrong - valid for high noise only, else spontaneous oscillations may arise
Random connectivity
Random Connectivity/Asynchronous State
1 s
randomly connected
1
0
,ˆ|ˆ
dststPI ),( 0 hg
),( 0 hg
0h
0A
A(t)=const
frequency (single neuron)
improved mean field
C inputs per neuron
0
0
CJ
ijw
extIAqJI 0000
002 A
C
J
mean drive
variance
(Amit&Brunel 1997, Brunel 2000)
Analogous for column of 1 exc. + 1 inhib. Pop.
The end
Liquid state/Information buffering
Ultra-short-term information buffering (‘echo’)
I(t))()( f
jj f
j ttatout
How much information in out(t) about signal
2)()()( tItoutaE k
)( tI ?
(Maass et al., 2002, H.Jäger 2004)
640 excitatory n.160 inhibitory n.20ms time constant50 connections/n.
population activity
0)( AtA
bias <I>
N parameters
)()( tItout
ms40
Ultra-short-term information buffering (‘echo’)
I(t)
)()( fj
j f
ttatout
dsstAsa )()(
(Mayor&Gerstner)
)(tA
bias <I>
mean-field readoutmicroscopic readout