brain anatomy and artificial intelligence -...
TRANSCRIPT
Brain anatomy and artificial intelligence
L. Andrew CowardAustralian National University, Canberra, ACT 0200, Australia
The Fourth Conference onArtificial General Intelligence August 2011
Functional Architecture
Control ArithmeticLogic
RegistersProcessor
Physical Architecture
Architectures and Information Processes
Memory
CallProcessing
Diagnostics
Maintenance
Billing
Medulla
Thalamus
Hypothalamus
Basalganglia
Cerebellum
PrecentralGyrus
CentralSulcus
Post CentralGyrus
MammillaryBody
LateralVentricle
CorpusCallosum
Fornix
Pons
Tectum
Tegmentum
Cortex
SpinalCord
Major Brain Structures
Behaviourselection
ResourceManagement
Rewardmanagement
Behavioursequencemanagement
Information flowmanagement
Behaviour Implementation
condition definitionand detection
behaviourselection
Behaviour typeprobabilitymanagement
Amygdala
Dorsal BasalGanglia
Hippocampus
Ventral BasalGanglia
Cerebellum
Brain stem;Spinal cord
Cortex
Thalamus
Brain Architecture and Information Processes
400 µm column All pyramidal neurons in all layersof a column have similar receptive fields
Receptive Fields of Cortical Columns
68B
8A
9
9/46d
9/46v10
46
47/1245A 45B
44
3,1,2 57
17
1819
4
43
40
22
3837
2120
52 42 41
39
24
33
25
29
34
27
38 20
36
2835 37
86 4
32
11
9
10
2331
7
19
1918
1817
5
26 30
12
Remembering past event
Elaborating
Conceiving
Imagining future event
Activity of Cortical Areas During Mental Imaging
apicaldendrite
basal dendrite
axon
synapses from differentcortical pyramidals withvarious weights
inhibitory synapsesfrom local interneuron
A
C
B
a1
a2
a3
a4
soma
Inputs
Outputs
Conditions that arecombinations of inputs frompreceding array
Conditions that arecombinations of conditionsdetected by previous level
Conditions that arecombinations of conditionsdetected by previous level
IV
II/III
V/VI
Cortical layers Type of condition detected in layer
Receptive Fields Must Change as Little and as Rarely as Possible
Cortex
Perirhinal andparahippocampalcortices
Hippocampus
Entorhinal cortex
CA1 CA3 DG
Hippocampus gets Information on Internal Column Activityfrom All Over the Cortex
pyramidalneuron
granule cell
inhibitiveinterneuron
mossycell
dentate gyrus
CA1
CA3
Inputs from entorhinal cortex
Outputs to cortex
Hippocampal Competition Determines Which Cortical Columnswill Change Receptive Fields
BIRD
inactivecolumn
column detecting itsreceptive field
column expanding andtherefore detecting itsreceptive field
familiarbirds
unfamiliarbird
Receptive Field Detection in One Area in Response to Different Objectsof Same Category
columns oftenactive whenbirds areperceived
BIRD
Some Columns are Often (but Not Always) Active When a Bird is Seen
≈ groups ofobjects
≈ visualobjects
≈ group ofgroups of
objects
Columns active when perceiving a novel event
inactivecolumn
column detecting itsreceptive field
column expanding andtherefore detecting itsreceptive field
≈ groups ofobjects
≈ visualobjects
≈ group ofgroups of
objects
Columns active when hearing words that relate to novel past event
≈ groups ofobjects
≈ visualobjects
≈ group ofgroups of
objects
Capability to indirectly activate columns on the basis ofpast simultaneous receptive field expansion
~ groups ofobjects
~ visualobjects
~ group ofgroups of
objects
= group ofgroups duringoriginalexperience
Column population indirectly activated on the basis ofpast simultaneous receptive field expansion
Behaviourselection
ResourceManagement
Rewardmanagement
Behavioursequencemanagement
Information flowmanagement
Behaviour Implementation
condition definitionand detection
behaviourselection
Behaviour typeprobabilitymanagement
Amygdala
Dorsal BasalGanglia
Hippocampus
Ventral BasalGanglia
Cerebellum
Brain stem;Spinal cord
Cortex
Thalamus
SNc
STN
GPe
Thalamus
Directpathway
Indirectpathway
D2D1Striatum
Cortex
GPi/SNr
Modulationpathway
Excitatory(glutamatergic)
Inhibitory(GABAergic)
Modulatory(dopaminergic)
DorsalBasalganglia
Behaviour Selection, Including At Least and Only One Behaviour
Shell}}
}}
Core
Centralstriatum
Dorsolateralstriatum
}}
}}
orbital andmedialprefrontal
dorsolateralprefrontal premotor motor
VTA
SNc
Cortex
Shell
Core
Centralstriatum
Dorsolateralstriatum
Midbrain dopamine neurons
Striatum
Striatum
Implementation of Strategic, Tactical and Detailed Reward Behaviours
Behaviourselection
ResourceManagement
Rewardmanagement
Behavioursequencemanagement
Information flowmanagement
Behaviour Implementation
condition definitionand detection
behaviourselection
Behaviour typeprobabilitymanagement
Amygdala
Dorsal BasalGanglia
Hippocampus
Ventral BasalGanglia
Cerebellum
Brain stem;Spinal cord
Cortex
Thalamus
VA AN
VL
VPLP
P
LD
MD
Thalamicreticularnucleus
front
right
Massaintermedia
supplementarymotor cortex
motor cortex;supplementarymotor cortex
hippocampus
somatosensorycortex
visual and otherassociation areas
prefrontalcortex
entorhinalcortex
medialprefrontalcortex
touch, propiocepticsensory information
LGN MGN
primaryauditorycortex
primaryvisualcortex
visualinformation
auditoryinformation
VI
V
IV
thalamicreticularnucleus
thalamicnucleus
corticallayers
TRN inhibitoryinterneuron
thalamocorticalprojectionneuron
corticalpyramidalneuron
basalganglia
inhibitoryconnection
excitatoryconnection
inhibitoryconnection
Implementation of Release Behaviours by Imposing Frequency Modulation
Behaviourselection
ResourceManagement
Rewardmanagement
Behavioursequencemanagement
Information flowmanagement
Behaviour Implementation
condition definitionand detection
behaviourselection
Behaviour typeprobabilitymanagement
Amygdala
Dorsal BasalGanglia
Hippocampus
Ventral BasalGanglia
Cerebellum
Brain stem;Spinal cord
Cortex
Thalamus
Parallel Cerebellar Path for ImplementingFrequently Used Sequences of Actions
To design an artificial general intelligence system
Focus on the information processesthat are requiredIf a system needs to
Perform many different behavioursDetect many different conditionsDefine most of the conditions heuristicallyLimit the resources required
Then the information processes willneed to be
Condition definition and detectionCondition resource managementInformation flow managementBehaviour selectionReward behaviour selectionBehaviour priority managementBehaviour sequence management
Behaviourselection
Rewardmanagement
Informationflowmanagement
behaviourselection
Behavioursequencemanagement
BehaviourImplementation
ResourceManagement
conditiondefinition
anddetection
Behaviourtypeprobabilitymanagement
a. Boundary of objectin visual field
b. Some of boundaryelement receptivefields detected in V1
c. Retinal area forwhich one boundaryelement receptivefield detectionrecommendsmodulation of inputsfrom area
d. Total recommendationstrengths for modulation muchstronger for area within objectboundary
20105 15 25 30 35
msecinput actionpotential spikes
threshold
postsynapticpotential
total
time window ofone third of themodulation cycle
integration windowclose to modulationmaximum
integration windowclose to modulation
minimum
source neurons target neuron
I
II
III
dentategyrus
entorhinalcortex
II
III
V/VICA2
CA1pyramidalneurongranulecellinhibitiveinterneuronmossycell
supra-mammillaryarea
mammillarybodies
anteriorthalamus
amygdala
subicular complex
CA3
cortex
Connectivity
parallel fibreoutputs fromgranule cells
Purkinje celldendritic treeperpendicular toparallel fibres
Purkinje cellbodies
granulecells
Purkinje cell outputtargets cerebellumnuclei
mossy fibreinput targetsgranule cells
climbing fibre inputtargets small group ofPurkinje cells (~10)
InferiorOlive
PontineNucleus
24
33
25
29
34
27
38 20
36
2835 37
86 4
32
11
9
10
2331
7
19
1918
1817
5
26 30
12
68B
8A
9
9/46d
9/46v10
46
47/1245A 45B
44
3,1,2 57
17
1819
4
43
3940
22
3837
2120
52 42 41
Areas concerned with working memory
Areas with receptive fields corresponding withgroups of cortical columns often active atsimilar times in the past
Areas with receptive fields corresponding withgroups of cortical columns that expanded theirreceptive fields at similar times in the past
left cerebralhemisphere
basal nucleicerebellum lateral ventricles some of basal
gangliathalamus
striatum
claustrum
amygdalahippocampus
hypothalamus
i ii iii
ivvvi
fornix
septal nucleinucleusaccumbens
substantianigra
caudatenucleus putamen
the striking commonalities in medial left prefrontal and parietal activity during theelaboration of (a) past and (b) futureevents (relative to the control tasks)
From Addis, D.A., Wong, A.T. and Schacter, D.L. (2007). Remembering the past andimagining the future: Common and distinct neural substrates during event construction andelaboration. Neuropsychologia 45, 1363-1377.