emergence of oriented circuits driven ... - jiglesia.alawa.ch · introduction: synaptogenesis and...
TRANSCRIPT
Emergence of Oriented Circuits driven bySynaptic Pruning associated with
Spike-Timing-Dependent Plasticity (STDP)
Javier Iglesias1,2
under the supervision of
Alessandro E.P. Villa2 and Marco Tomassini1
1 Information Systems Department, University of Lausanne, Switzerland
2 Laboratory of Neurobiophysics, University Joseph-Fourier, France
introduction: synaptogenesis and synaptic pruning 1
NB 0.5 1 2 5 10 adult aged(74-90)years
10
6
8
4
12
2
ne
uro
ns /
mm
x1
03
4
NB 0.5 1 5 10 15 20 40 60 80 100
years
15
10
20
5
syn
ap
se
s /
mm
x1
03
8
modified from Huttenlocher, Synaptic density in human frontal cortex – developmentalchanges and effects of aging, Brain Research, 163:195–205, 1979
introduction: synaptic plasticity 2
-40 400
-60 600 -25 250 -40 400
-50 500-100 100
syn
ap
tic c
ha
ng
esyn
ap
tic c
ha
ng
e
syn
ap
tic c
ha
ng
e
syn
ap
tic c
ha
ng
e
syn
ap
tic c
ha
ng
e
a b
c d e
time [ms]
time [ms] time [ms] time [ms]
time [ms]
modified from Roberts and Bell, Spike timing dependent synaptic plasticityin biological systems, Biol. Cybern., 87:392–403, 2002
introduction: existing work 3
• The memory performance of a network is optimally maximized if,under limited metabolic energy resources restricting their numberand strength, synapses are first overgrown and then pruned.
Chechik et al., Synaptic pruning in development:A computational account, Neural Computation, 10(7):1759–77, 1998
• Neuronal regulation might maintain the memory performance ofnetworks undergoing synaptic degradation.
Horn et al., Memory maintenance via neuronal regulationNeural Computation, 10(1):1–18, 1998
• STDP has been shown to maintain the postsynaptic input field.
Abbott et al., Synaptic plasticity: taming the beastNature Neuroscience, 3:1178–83, 2000
model: network 4
500
-50
x -50
0.0
0.2
0.4
0.6probability
e
500
-50
x
50
0
-50y
0.0
0.2
0.4
0.6probability
+500-50
x
+50
0
-50
y
f
+500-50x
+50
0
-50
y
+50
0
-50+500-50
y
x
+50
0
-50+500-50
y
x
g
0
250
500
0 200 400
cell
coun
t
connection count
0
250
500
0 200 400
cell
coun
t
connection count
a b c d
h
50
0 y
i e i i
600
e e
e i
600
Iglesias et al., Dynamics of Pruning in Simulated Large-ScaleSpiking Neural Networks, Biosystems, 79(1-3):11-20, 2005
introduction: leaky integrate and fire neuromimetic model 5
V(t)S(t)
B(t)
w(t)
~190 excitations
~115 inhibitions
Type I = excitatory 80%Type II = inhibitory 20%
Vrest = -76 [mV]θi = -40 [mV]τmem = 8 [ms]trefract = 1 [ms]λi = 10 [spikes/s]n = 50
Vi(t+1) = Vrest[q]+(1−Si(t))·((Vi(t)−Vrest[q])·kmem[q])+∑
j
wji(t)+Bi(t)
Si(t) = H(Vi(t)− θqi)
wji(t + 1) = Sj(t) ·Aji(t) · P[qj,qi]
Bi(t + 1) = Preject(λqi) · n · P[q1,qi]
model: STDP and pruning 6
Lji(t + 1) = Lji(t) · kact[qj,qi]
+(Si(t) ·Mj(t))−(Sj(t) ·Mi(t))
post
pre
S (t)i
S (t)j time
time
model: STDP and pruning 6
Lji(t + 1) = Lji(t) · kact[qj,qi]
+(Si(t) ·Mj(t))−(Sj(t) ·Mi(t))
post
pre
S (t)i
S (t)j time
timepost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
time
model: STDP and pruning 6
Lji(t + 1) = Lji(t) · kact[qj,qi]
+(Si(t) ·Mj(t))−(Sj(t) ·Mi(t))
post
pre
S (t)i
S (t)j time
timepost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
timepost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
time
L (t)ji
model: STDP and pruning 6
Lji(t + 1) = Lji(t) · kact[qj,qi]
+(Si(t) ·Mj(t))−(Sj(t) ·Mi(t))
post
pre
S (t)i
S (t)j time
timepost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
timepost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
time
L (t)ji
Lji(
t)A
ji(t)
time [s]50 100 1500
4
2
1
0
wji(t + 1) = Sj(t) ·Aji(t) · P[qj,qi]
model: synaptic adaptation examples 7
timeL0
L1
L2
L3
L4
timeL0
L1
L2
L3
L4
timeL0
L1
L2
L3
L4
a b cA4
A3
A2
A1
results: simulation evolution 8
0
20
40
60
80
100
0 200 400 600 800 1,000time [s]
cum
ula
ted
dis
trib
ution [%
]
A (t)=0ji
A (t)=4ji
A (t)=2ji
A (t)=1ji
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
200
100
0
2001000
t=5s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
200
100
0
2001000
t=5s
k_in
k_out
200
100
0
2001000
t=10s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
200
100
0
2001000
t=5s
k_in
k_out
200
100
0
2001000
t=10s
k_in
k_out
200
100
0
2001000
t=20s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
200
100
0
2001000
t=5s
k_in
k_out
200
100
0
2001000
t=10s
k_in
k_out
200
100
0
2001000
t=20s
k_in
k_out
200
100
0
2001000
t=50s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
200
100
0
2001000
t=5s
k_in
k_out
200
100
0
2001000
t=10s
k_in
k_out
200
100
0
2001000
t=20s
k_in
k_out
200
100
0
2001000
t=50s
k_in
k_out
200
100
0
2001000
t=100s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
200
100
0
2001000
t=5s
k_in
k_out
200
100
0
2001000
t=10s
k_in
k_out
200
100
0
2001000
t=20s
k_in
k_out
200
100
0
2001000
t=50s
k_in
k_out
200
100
0
2001000
t=100s
k_in
k_out
200
100
0
2001000
t=200s
k_in
k_out
results: simulation evolution (cont.) 9
200
100
0
2001000
t=0s
k_in
k_out
200
100
0
2001000
t=2s
k_in
k_out
200
100
0
2001000
t=5s
k_in
k_out
200
100
0
2001000
t=10s
k_in
k_out
200
100
0
2001000
t=20s
k_in
k_out
200
100
0
2001000
t=50s
k_in
k_out
200
100
0
2001000
t=100s
k_in
k_out
200
100
0
2001000
t=200s
k_in
k_out
200
100
0
2001000
t=500s
k_in
k_out
results: scaling factor 10
φ∗[exc] = f · φ[exc]
f = 3
√104
(10·N)2
P ∗[exc,exc] =
(1 + f−1
2
)· P[exc,exc]
network sizeis defined as(10×N)2
1
10
100
100010010
active s
ynapses [A
4] (%
)
time [s]
123456789
10
model: simple stimulus 11
t = 1
stimulation
onset
t = 2 t = 3 t = 4 t = 5 t = 6 t=duration
(50, 100
or 200 ms)
[...]
every 2 secondsN vertical bars moving to the rightduring 50, 100, or 200 time steps.
input units ratio: 3, 5, 7, or 10% excitatory units.randomly selected before stimulation or in the beginning
See an animated sequence.
model: graph considerations 12
Strongly Interconnected (SI) unitsat the end of the simulationset of cells (discarding input units)maintaining kout ≥ 3 and kin ≥ 3with strongest activation level (Aji(t) = 4)with units with the same properties.
model: graph considerations 12
Strongly Interconnected (SI) unitsat the end of the simulationset of cells (discarding input units)maintaining kout ≥ 3 and kin ≥ 3with strongest activation level (Aji(t) = 4)with units with the same properties.
Neighbourhoodall excitatory units (including input units)with at least one projection to or from SI-units.
model: graph considerations 12
Strongly Interconnected (SI) unitsat the end of the simulationset of cells (discarding input units)maintaining kout ≥ 3 and kin ≥ 3with strongest activation level (Aji(t) = 4)with units with the same properties.
Neighbourhoodall excitatory units (including input units)with at least one projection to or from SI-units.
index of connected units (ICU)#input units in neighbourhood
#SI−units
results: with and without scaling factor 13
0
5
10
15
10 7 5 3
ICU
ratio of input units [%]
90 x 90
100 x 100
110 x 110
0
5
10
15
10 7 5 3
ratio of input units [%]
a b
ICU
Iglesias et al., Stimulus-Driven Unsupervised Pruningin Large Neural Networks, LNCS (in press, October 2005)
results: emergence of circuits from t=0 to t=50s 14
stimulus amplitude = 0mV stimulus amplitude = 60mV
stimulus protocol = random
stimulus amplitude = 60mV
stimulus protocol = fixed
k_o
ut
k_o
ut
k_in k_in k_in
0 100 200 0 100 200 0 100 200
0
100
200
0
100
200
k_
out
k_o
ut
0
100
200
100
200
0
a
b c d
t =
0 s
eco
nd
st =
50
seco
nds
results: emergence of circuits from t=50 to t=200s 15
stimulus amplitude = 0mV stimulus amplitude = 60mV
stimulus protocol = random
stimulus amplitude = 60mV
stimulus protocol = fixed
k_o
ut
k_o
ut
k_in k_in k_in
0 100 200 0 100 200 0 100 200
0
100
200
0
100
200
k_
out
k_o
ut
0
100
200
100
200
0
b c d
e f g
t =
50 s
econ
ds
t =
20
0 s
eco
nds
results: emergence of circuits from t=200 to t=500s 16
stimulus amplitude = 0mV stimulus amplitude = 60mV
stimulus protocol = random
stimulus amplitude = 60mV
stimulus protocol = fixed
k_o
ut
k_o
ut
k_in k_in k_in
0 100 200 0 100 200 0 100 200
0
100
200
0
100
200
k_
out
k_o
ut
0
100
200
100
200
0
t =
200 s
eco
nd
st =
50
0 s
eco
nds
e f g
h i j
results: emergence of circuits from t=200 to t=500s (cont.) 17
stimulus amplitude = 0mV stimulus amplitude = 60mV
stimulus protocol = random
stimulus amplitude = 60mV
stimulus protocol = fixed
k_o
ut
k_o
ut
k_in k_in k_in
0 100 200 0 100 200 0 100 200
0
0
k_
out
k_o
ut
0
0
t =
200 s
eco
nd
st =
50
0 s
eco
nds
20
40
20
40
20
40
20
40e f g
h i j
results: circuit 18
Iglesias et al., Emergence of Oriented Cell AssembliesAssociated with Spike-Timing-Dependent Plasticity, LNCS 3696:127-132, 2005
layer 1 layer 2 layer 3 layer 4 layer 5
5326
5697
2800
5550
7199
7297
9661
992
4695
1539
1202
8402
1608
8889
931931
81208120
80222267
3027
2007
1152
2666
8467
1151
7928
5724 5724
490 490490
45478863
4761
2279
804
826782678267
1234
1048
1435
8300
7794
1049
2532
1842 4622
24272427
9983
1845
2848
492492
147214721472
7235
time = 500 s
results: circuit 18
Iglesias et al., Emergence of Oriented Cell AssembliesAssociated with Spike-Timing-Dependent Plasticity, LNCS 3696:127-132, 2005
layer 1 layer 2 layer 3 layer 4 layer 5
5326
5697
2800
5550
7199
7297
9661
992
4695
1539
1202
8402
1608
8889
931931
81208120
80222267
3027
2007
1152
2666
8467
1151
7928
5724 5724
490 490490
45478863
4761
2279
804
826782678267
1234
1048
1435
8300
7794
1049
2532
1842 4622
24272427
9983
1845
2848
492492
147214721472
7235
time = 500 s
layer 1 layer 2 layer 3 layer 4 layer 5
5326
5697
2800
5550
7199
7297
9661
992
4695
1539
1202
8402
1608
8889
931931
81208120
80222267
3027
2007
1152
2666
8467
1151
7928
5724 5724
490 490490
45478863
4761
2279
804
826782678267
1234
1048
1435
8300
7794
1049
2532
1842 4622
24272427
9983
2848
1845
492492
147214721472
7235
time = 0 s
model: complex stimulus 19
t = 1
stimulation
onset
t = 2 t = 3 t = 4 t = 5 t = 6 t=duration
(50, 100
or 200 ms)
[...]A
B [...]
every 2 seconds10 groups of 40 units activated in sequence (5% input units)during 200 time steps (AA, BB, AB, BA, AB|BA)
continuous and interrupted pruning
Animated sequences are available for both stimuli A and B.
results: firing rates vs. connection degrees 20
26 units with kin = kout = 0 in absence of stimulationand kin > 0 and kout > 0 with all other stimuli (AA, BB, ...)
0
50
100
150
200
250
0
10
20
30
40
50
0 20 40 60 80 100 120
a b
0 20 40 60 80 100 120
mean firing rate [spikes/s]mean firing rate [spikes/s]
k_
in
k_
ou
t
results: spatio-temporal pattern of activity 21
<5, 9, 4, 2; 13, 53, 109>
10008006005004002000time [seconds]
10008006005004002000time [seconds]
a b
dctime [ms]
0-100 500400100 200 300time [ms]
0-100 500400100 200 3000
40
0
40
continuous pruning pruning stopped at t=500s
#5
#9
#4
#2
#5
#9
#4
#2
for methods, see Villa et al., 1999; Tetko and Villa, 2001
conclusion 22
• network size effect
• emergence of oriented circuitsby synaptic pruning associated with STDP
• circuits are able to producespatiotemporal patterns of activity
• role of firing rate
discussion: firing rates vs. connection degrees (cont.) 23
post
pre
S (t)i
-M (t)i
S (t)j
M (t)j
a bpost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
time
time
time
L (t)ji L (t)ji
discussion: firing rates vs. connection degrees (cont.) 23
post
pre
S (t)i
-M (t)i
S (t)j
M (t)j
a bpost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
time
time
time
L (t)ji L (t)ji
post
pre
S (t)i
-M (t)i
S (t)j
M (t)j
c dpost
pre
S (t)i
-M (t)i
S (t)j
M (t)j
time
time
time
time
L (t)ji L (t)ji