spaun: a biologically realistic large-scale functional ...compneuro.uwaterloo.ca › files ›...
TRANSCRIPT
Spaun: A biologically realistic large-scale functional brain modelChris Eliasmith (Centre for Theoretical Neuroscience, University of Waterloo)
Trevor Bekolay, Xuan Choo, Terrence C. Stewart, Travis DeWolf, Yichuan Tang, Daniel Rasmussen, Jan Gosmann
The Semantic Pointer Architecture UnifiedNetwork (Spaun) is a network of 2.5 millioninterconnected artificial spiking neurons.
Spaun is composed of groups of neurons that perform functions nec-essary to complete cognitive tasks. It flexibly coordinates thosegroups depending on the cognitive task being performed.
Working memory
Visualinput
Informationencoding
Transformcalculation
Rewardevaluation
Informationdecoding
Motorprocessing
Motoroutput
Action selection
These functionally related groups of neurons are mapped onto brainareas consistent with our current understanding of functional neu-roanatomy. Spaun can be manipulated in order to test hypothesesin neuroscience.
PPCM1 SMA
PM
DLPFC
VLPFC
OFC
AIT
Thalamus
ITV4V2
V1
Str (D2)
GPe
STN Str (D1)
GPi/SNr
SNc/VTA
vStr
TheoryCNS
Systems
Maps
Networks
Neurons
Synapses
Molecules
1m
10cm
1cm
1mm
100μm
1μm
10nm
⇐ Nengo modelling software1
⇐ Semantic pointer architecture2
List=Pos1~Item1�Pos2~Item2 . . .
⇐ Neural Engineering Frameworkωij =αjdiej x̂=∑
ai(x)di
⇐ Single-cell modellingdVdt =− 1
RC (V(t)−J(t)R)⇐ Post-synaptic current curves
h(t)= e−t/τ if t> 0
Spaun performs tasks like humans
PFC
DLFPC1
PFC
SMA
stimulus
IT
Str
GPi
DLPFC2
DLFPC1
DLPFC2
arm
Time (s)
*
**
*
*
* * * *
**
*
*
* *+
+ +
+
+ +
+
A 7 [ 1 ] [ 1 1 ] [ 1 1 1 ] [ 4 ] [ 4 4 ] [ 4 4 4 ] [ 5 ] [ 5 5 ] ?
Chance Humans SpaunImage recognition 10% 98% 94%
Fluid reasoning 13% 89% 88%
Number of positions counted
Tim
e (s
)
1 2 3 4 50.5
0.75
1
1.25
1.5
1.75
2
2.25
2.5
Counting
1 2 3 4 5 6 70
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Item position
Accu
racy
Question answering
What position is X?What is in position Y?
Spaun makes mistakes like humans
0 1 5 8 7 3
stimulus
IT
Str
GPi
PFC
DLPFC
DLPFC
SMA
Arm
Time (s)0 1 2 3 4 5 6
A 3 [ 0 1 5 8 7 3 ] ?
1 2 3 4 5 6 70
0.1
0.5
0.6
0.7
0.8
0.9
1
4 items5 items6 items7 items
1 2 3 4 5 6 70
0.1
0.5
0.6
0.7
0.8
0.9
1
4 items5 items6 items7 items
Item position
Acc
ura
cy
List memory (Human data)
Acc
ura
cy
List memory (Spaun)
Item position
Spaun matches experimental data
0
2
4
6
8
Spik
es p
er s
econ
d
Delay Approach Reward Return
Experimental vStrSimulated vStr
5
15
25
35
Tria
l num
ber
0.0 0.1 0.2 0.3 0.4 0.5
Time (s)0
255075
100125150
Neu
ron
num
ber
Data Spaun
Data Spaun
Spaun can perform 8 cognitive tasks using only visual informationand without external intervention.
Copy drawing“Copy this 2.”
A 0 [ ] ? ⇒
Image recognition“Write a 2.”
A 1 [ 2 ] ? ⇒
Gambling“Try slot machine 0, 1 or 2.”
A 2 ? ⇒ ⇒ 0A 2 ? ⇒ ⇒ 1A 2 ? ⇒ ⇒ 1
List memory“Write the list 0 1 5 8 7 3.”
A 3 [ 0 1 5 8 7 3 ] ?⇒
Counting“Starting from 3, count 5.”
A 4 [ 3 ] [ 5 ] ? ⇒
Question answering“Write the 2nd number in 0 1 5 8 7 3.”
A 5 [ 0 1 5 8 7 3 ] [ P ] [ 2 ] ?⇒
Rapid variable creation“Complete the pattern:
0014→14, 0094→94, 0074→?”
A 6 [ 0 0 1 4 ] [ 1 4 ][ 0 0 9 4 ] [ 9 4 ][ 0 0 7 4 ] ?⇒
Fluid reasoning“Fill in the last cell.”
A 7 [ 1 ] [ 1 1 ] [ 1 1 1 ][ 4 ] [ 4 4 ] [ 4 4 4 ][ 5 ] [ 5 5 ] ?⇒
Recent developments
General instruction processingConductance
neuron model3
1.0
0.5
0.0
0.5
1.0
Input B (50 standard LIF)
0 1 2 3 4 5 6
1.0
0.5
0.0
0.5
1.0
0 1 2 3 4 5 6
Input B (50 compartmental)
Time (s) Time (s)
0 10 20 30 40 50 60 70 800
20
40
60
80
100
Percentage of Na+ Channels Blocked (%)
Dig
it R
ecog
nitio
n A
ccur
acy
(%
)
A large-scale model of the functioning brain. Science, 338:1202–1205.Chris Eliasmith, Terrence C. Stewart, Xuan Choo, Trevor Bekolay, Travis DeWolf,Yichuan Tang, and Daniel Rasmussen (2012).
1) Trevor Bekolay et al. (2013). Nengo: a Python tool for building large-scalefunctional brain models. Frontiers in Neuroinformatics 7.Available at http://github.com/nengo/nengo
2) Chris Eliasmith (2013). How to build a brain: A neural architecture for biologicalcognition. Oxford University Press.
3) Armin Bahl, Martin B. Stemmler, Andreas V.M. and Arnd Roth (2012). Automatedoptimization of a reduced layer 5 pyramidal cell model based on experimental data.Journal of Neuroscience Methods, 210:22–34.
View this poster online at http://compneuro.uwaterloo.ca/files/spaun-2015.pdf