Wenzao Ursuline College of Languages
Kaohsiung, Taiwan
On the Neurocognitive Basis of Language
Sydney Lamb
[email protected]
2010 November 12
Why is it important to consider the brain?
“I gather…that the status of linguistic theories continues
to be a difficult problem. … I would wish, cautiously, to
make the suggestion, that perhaps a further touchstone
may be added: to what esxtent does the throry tie in with
other, non-linguistic information, for example, the
anatomical aspects of language? In the end such bridges
link a theory to the broader body of scientific knowledge.”
Norman Geschwind
“The development of the brain and the evolution of language”
Georgetown Round Table on Languages and Linguistics, 1964
Topics
•
•
•
•
•
A little neuroanatomy
Functional webs
Nodes and links: Cortical columns
Basic operations in the cortex
More operations: Learning
Topics
•
•
•
•
•
•
A little neuroanatomy
Functional webs
Nodes and links: Cortical columns
Basic operations in the cortex
Syntax
More operations: Learning
The brain
•
•
•
•
•
Medulla oblongata – Myelencephalon
Pons and Cerebellum – Metencephalon
Midbrain – Mesencephalon
Thalamus and hypothalamus – Diencephalon
Cerebral hemispheres – Telencephalon
– Cerebral cortex
– Basal ganglia
– Basal forebrain nuclei
– Amygdaloid nucleus
Two hemispheres
Left
Interhemispheric fissure
(a.k.a. longitudinal
fissure)
Right
Corpus Callosum Connects Hemispheres
Corpus
Callosum
Major Left Hemisphere landmarks
Central Sulcus
Sylvian fissure
Major landmarks and the four lobes
Central Sulcus
Frontal
Lobe
Sylvian fissure
Parietal
Lobe
Temporal
Lobe
Occipital
Lobe
Some brain facts – now well established
• Locations of various kinds of “information”
– Visual, auditory, tactile, motor, …
• The brain is a network
– Composed, ultimately, of neurons
• Neurons are interconnected
– Axons (with branches)
– Dendrites (with branches)
• Activity travels along neural pathways
– Cortical neurons are clustered in columns
• Columns come in different sizes
– The smallest: minicolumn – 70-110 neurons
• Each minicolumn acts as a unit
– When it becomes active all its neurons are active
Deductions from known facts
• Everything represented in the brain has the form of a
network
– (the “human information system”)
• Therefore a person’s linguistic and conceptual system
is a network
– (part of the information system)
• Every lexical entry and every concept is a sub-network
– Term: functional web (Pulvermüller 2002)
Primary Areas
Central Sulcus
Primary Somatosensory Area
Primary
Motor Area
Primary Auditory
Area
Sylvian fissure
Primary
Visual Area
Divisions of Primary Motor and Somatic Areas
Leg
Primary
Motor Area
Primary Somatosensory Area
Trunk
Arm
Hand
Fingers
Mouth
Primary Auditory
Area
Primary
Visual Area
Higher level motor areas
Actions perFormed by leg
Actions
performed
by hand
Leg
Primary Somatosensory Area
Trunk
Arm
Hand
Fingers
Actions
performed
by mouth
Mouth
Primary Auditory
Area
Primary
Visual Area
Hierarchy in cortical development
Topics
•
•
•
•
•
•
A little neuroanatomy
Functional webs
Nodes and links: Cortical columns
Basic operations in the cortex
Syntax
More operations: Learning
Hypothesis I: Functional Webs
•
•
A word is represented as a functional web
Spread over a wide area of cortex
– Meaning includes perceptual information
– As well as specifically conceptual information
• For nominal concepts, mainly in
– Angular gyrus
– (?) For some, middle temporal gyrus
– (?) For some, supramarginal gyrus
Example: The concept DOG
• We know what a dog looks like
– Visual information, in occipital lobe
• We know what its bark sounds like
– Auditory information, in temporal lobe
• We know what its fur feels like
– Somatosensory information, in parietal lobe
• All of the above..
– constitute perceptual information
– are subwebs with many nodes each
– have to be interconnected into a larger web
– along with further web structure for conceptual information
Building a model of
a functional web:
first steps
Each node in this diagram
represents the cardinal node*
of a subweb of properties
T
For example
C
M
V
Let’s
zoom in
on this
one
*to be defined in a moment!
Zooming in on the “V” Node..
Cardinal
V-node
A network of
visual features
Etc. etc.
(many layers)
Add phonological recognition
For example, FORK
C
T
M
P
V
These are all
cardinal nodes –
each is supported
by a subweb
Labels for Properties:
C – Conceptual
M – Motor
P – Phonological image
T – Tactile
V – Visual
The phonological image
of the spoken form [fork]
(in Wernicke’s area)
Add node in primary auditory area
For example, FORK
C
T
M
P
PA
V
Labels for Properties:
C – Conceptual
M – Motor
P – Phonological image
PA – Primary Auditory
T – Tactile
V – Visual
Primary Auditory: the cortical structures in the primary
auditory cortex that are activated when the ears receive
the vibrations of the spoken form [fork]
Add node for phonological production
For example, FORK
C
T
M
P
PP
PA
V
Labels for Properties:
C – Conceptual
M – Motor
P – Phonological image
PA – Primary Auditory
PP – Phonological Production
T – Tactile
V – Visual
Part of the functional web for DOG
(showing cardinal nodes only)
Each
node
shown
here is
the
cardinal
node of
a
subweb
T
M
PP
C
P
PA
V
For example,
the cardinal
node of the
visual subweb
An activated functional web
(with two subwebs partly shown)
T
C
PP
PR
PA
M
C – Cardinal concept node
M – Memories
PA – Primary auditory
PP – Phonological production
PR – Phonological recognition
T – Tactile
V – Visual
V
Visual features
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Ignition of a functional web from visual input
T
C
PR
Art
PA
M
V
Speaking as a response to ignition of a web
T
C
PR
Art
PA
M
V
Speaking as a response to ignition of a web
T
C
PR
Art
PA
M
V
Speaking as a response to ignition of a web
T
C
PR
Art
PA
M
From here (via subcortical
structures) to the muscles that
control the organs of articulation
V
An MEG study from Max Planck Institute
Topics
•
•
•
•
•
A little neuroanatomy
Functional webs
Nodes and links: Cortical columns
Basic operations in the cortex
More operations: Learning
Hypothesis 2: Nodes as Cortical Columns
• Nodes are implemented as cortical columns
• The interconnections are represented by inter-columnar neural
connections and synapses
– Axonal fibers – neural output
– Dendritic fibers – neural input
The node as a cortical column
• The properties of the cortical column are
approximately those described by Vernon
Mountcastle
– Mountcastle, Perceptual Neuroscience, 1998
• Additional properties of columns and functional
webs can be derived from Mountcastle’s
treatment together with neurolinguistic findings
Quote from Mountcastle
“[T]he effective unit of operation…is not
the single neuron and its axon, but
bundles or groups of cells and their
axons with similar functional properties
and anatomical connections.”
Vernon Mountcastle,
Perceptual Neuroscience
(1998), p. 192
Three views of the gray matter
Different stains
show different
features
Layers of the Cortex
From top to
bottom,
about 3 mm
The Cerebral Cortex
 Grey matter
• Columns of neurons
White matter
• Inter-column connections
Microelectrode penetrations in the
paw area of a cat’s cortex
Columns for orientation of lines (visual cortex)
Microelectrode
penetrations
K. Obermayer & G.G. Blasdell, 1993
The (Mini)Column
• Width is about (or just larger than) the
diameter of a single pyramidal cell
– About 30–50 m in diameter
• Extends thru the six cortical layers
– Three to six mm in length
– The entire thickness of the cortex
is accounted for by the columns
• Roughly cylindrical in shape
• If expanded by a factor of 100, the
dimensions would correspond to a
tube with diameter of 1/8 inch and
length of one foot
Simplified model of minicolumn I:
Activation of neurons in a column
Other
cortical
locations
Cell Types
II
III
Pyramidal
Spiny
Stellate
Thalamus
IV
Inhibitory
Connections to
neighboring
columns not
shown
V
VI
Subcortical
locations
Cortical column structure
• Minicolumn 30-50 microns diameter
• Recurrent axon collaterals of pyramidal
neurons activate other neurons in same
column
• Inhibitory neurons can inhibit neurons of
neighboring columns
– Function: contrast
• Excitatory connections can activate
neighboring columns
– In this case we get a bundle of
contiguous columns acting as a unit
Cortical minicolumns: Quantities
•
•
•
•
•
•
Diameter of minicolumn: 30 microns
Neurons per minicolumn: 70-110 (avg. 75-80)
Minicolumns/mm2 of cortical surface: 1460
Minicolumns/cm2 of cortical surface: 146,000
Neurons under 1 sq mm of cortical surface: 110,000
Approximate number of minicolumns in Wernicke’s area:
2,920,000 (at 20 sq cm for Wernicke’s area)
Adapted from Mountcastle 1998: 96
Large-scale cortical anatomy
• The cortex in each hemisphere
– Appears to be a three-dimensional structure
– But it is actually very thin and very broad
• The grooves – sulci – are there because the
cortex is “crumpled” so it will fit inside the skull
Topologically, the cortex of each hemisphere
(not including white matter) is..
• Like a thick napkin, with
– Area of about 1300 square centimeters
• 200 sq. in.
• 2600 sq cm for whole cortex
– Thickness varying from 3 to 5 mm
– Subdivided into six layers
• Just looks 3-dimensional because it is
“crumpled” so that it will fit inside the skull
Topological essence of cortical structure
(known facts from neuroanatomy)
• The thickness of the cortex is entirely
accounted for by the columns
• Hence, the cortex is an array of nodes
– A two-dimensional structure of
interconnected nodes (columns)
• Third dimension for
– Internal structure of the nodes (columns)
– Cortico-cortical connections (white matter)
Nodal interconnections
(known facts from neuroanatomy)
• Nodes (columns) are connected to
– Nearby nodes
– Distant nodes
• Connections to nearby nodes are either excitatory or inhibitory
– Via horizontal axons (through gray matter)
• Connections to distant nodes are excitatory only
– Via long (myelinated) axons of pyramidal neurons
Local and distal connections
excitatory
inhibitory
Simplified model of minicolumn I:
Activation of neurons in a column
Other
cortical
locations
Cell Types
II
III
Pyramidal
Spiny
Stellate
Thalamus
IV
Inhibitory
Connections to
neighboring
columns not
shown
V
VI
Subcortical
locations
Simplified model of minicolumn II:
Inhibition of competitors
Other
cortical
locations
Cell Types
II
III
Pyramidal
Spiny
Stellate
Thalamus
IV
Inhibitory
V
VI
Cells in
neighboring
columns
Local and distal connections
excitatory
inhibitory
Findings relating to columns
(Mountcastle, Perceptual Neuroscience, 1998)
• The column is the fundamental module of perceptual systems
– probably also of motor systems
• This columnar structure is found in all mammals that have
been investigated
• The theory is confirmed by detailed studies of visual, auditory,
and somatosensory perception in living cat and monkey brains
Functional webs and subwebs
• A functional web for a word consists of multiple subwebs
• Every such subweb
– has a specific function
– occupies an area that fits the portion of cortex in
which it located
• For example,
– Phonological recognition in Wernicke’s area
– Visual subweb in occipital and lower temporal lobe
– Tactile subweb in parietal lobe
Hypothesis 3: Nodal Specificity in functional webs
• Every node in a functional web has a specific function
• Each node of a subweb also has a specific function
within that of the subweb
Support for Nodal Specificity:
the paw area of a cat’s cortex
Column (node) represents
specific location on paw
Support for Nodal Specificity:
Columns for orientation of lines (visual cortex)
Microelectrode
penetrations
K. Obermayer & G.G. Blasdell, 1993
Hypothesis 3a: Adjacency
• Nodes of related function are in adjacent locations
– More closely related function, more closely adjacent
• Examples:
– Adjacent locations on cat’s paw represented by
adjacent cortical locations
– Similar line orientations represented by adjacent
cortical locations
Support for Nodal adjacency:
the paw area of a cat’s cortex
Adjacent column in cortex
for adjacent location on paw
Hypothesis 4: Extrapolation to Humans
• Hypothesis: The findings about cortical structure and
function from experiments on cats, monkeys, and rats can
be extrapolated to human cortical structure and function
• In fact, this hypothesis is simply assumed to be valid by
neuroscientists
• Why? We know from neuroanatomy that, locally,
– Cortical structure is relatively uniform across mammals
– Cortical function is relatively uniform across mammals
Hypothesis 4a: Linguistic and conceptual structure
• The extrapolation can be extended to linguistic and
conceptual structures and functions
• Why?
– Local uniformity of cortical structure and function across
all human cortical areas except for primary areas
• Primary visual and primary auditory are known to
have specialized structures, across mammals
• Higher level areas are – locally – highly uniform
Conceptual systems and perceptual systems
• Likewise, conceptual systems in humans evidently use
the same structures as perceptual systems
• Therefore it is not too great a stretch to suppose that
experimental findings on the structure of perceptual
systems in monkeys can be applied to an understanding
of the structure of conceptual systems of human beings
• In particular to the structures of conceptual categories
Extrapolation to Language?
• Our knowledge of cortical columns comes mostly from studies
of perception in cats, monkeys, and rats
• Such studies haven’t been done for language
– Cats and monkeys don’t have language
– That kind of neurosurgical experiment isn’t done on human
beings
• Are they relevant to language anyway?
– Relevant if language uses similar cortical structures
– Relevant if linguistic functions are like perceptual functions
Objection
• Cats and monkeys don’t have language
• Therefore language must have unique properties of its
structural representation in the cortex
• Answer: Yes, language is different, but
– The differences are a consequence not of different
(local) structure but differences of connectivity
– The network does not have different kinds of
structure for different kinds of information
• Rather, different connectivities
Summary of the argument
• Cortical structure and function, locally, are essentially
the same in humans as in cats and monkeys
• Moreover, in humans,
– The regions that support language have the same
structure locally as other cortical regions
Support for the connectionist claim
•
•
•
Lines and nodes (i.e., columns) are approximately the same
all over
Uniformity of cortical structure
– Same kinds of columnar structure
– Same kinds of neurons
– Same kinds of connections
Conclusion: Different areas have different functions because
of what they are connected to
Uniformity of cortical function
• Claims:
– Locally, all cortical processing is the same
– The apparent differences of function are consequences
of differences in larger-scale connectivity
• Conclusion (if the claim is supported):
– Understanding language, even at higher levels, is
basically a perceptual process
Hypothesis 5: Hierarchy in functional webs
• A functional web is hierarchically organized
– Bottom levels in primary areas
– Lower levels closer to primary areas
– Higher (more abstract) levels in
• Associative areas – e.g., angular gyrus
• Executive areas – prefrontal
• These higher areas are much larger in humans than
in other mammals
• Corollary: Each subweb is likewise hierarchically organized
Properties of Hierarchy
• Each level has fewer nodes than lower
levels, more than higher levels
– Compare the organization of
management of a corporation
• Top level has just one node
– Compare the “CEO”
Hypothesis 6: Cardinal nodes
• Every functional web has a cardinal node
– At the top of the entire functional web
– Unique to that concept
– For example, C/cat/ at “top” of the web for CAT
• Corollary:
– Each subweb likewise has a cardinal node
• At the top level of the subweb
• Unique to that subweb
• For example, V/cat/
– At the top of the visual subweb
Cardinal nodes of a functional web
Some of the cortical structure relating to dog
Each
node
shown
here is
the
cardinal
node of
a
subweb
Cardinal
node of the
whole web
T
M
PP
C
P
PA
V
Cardinal
node of
the visual
subweb
Support for the cardinal node hypothesis
It follows from the hypotheses of nodal specificity and
hierarchy
– A hierarchy must have a highest level
– The node at this level must have a specific function
2. It is needed to account for the arbitrariness of the
linguistics sign
3. It is automatically recruited in learning anyway,
according to the Hebbian learning hypothesis
1.
Cardinal nodes and the linguistic sign
•
•
Connection of conceptual to phonological representation
Consider two possibilities
1. A cardinal node for the concept connected to a
cardinal node for the phonological image
2. No cardinal nodes: multiple connections between
concept representation and phonological image
• supported by Pulvermüller (2002)
Implications of possibility 2
•
•
•
•
No cardinal nodes: multiple connections between
concept representation and phonological image
I.e., different parts of meaning connected to
different parts of phonological image
Consider fork
– Maybe /f-/ connects to the shape?
– Maybe /-or-/ connects to the feeling of holding
a fork in the hand?
– Maybe /-k/ connects to the knowledge that
fork is related to knife?
Conclusion: Possibility 2 must be rejected
Topics
•
•
•
•
•
A little neuroanatomy
Functional webs
Nodes and links: Cortical columns
Basic operations in the cortex
More operations: Learning
Cortical columns do not store symbols
• They only
– Receive activation
– Maintain activation
– Inhibit competitors
– Transmit activation
• Important consequence:
– We have linguistic information represented
in the cortex without the use of symbols
– It’s all in the connectivity
• Challenge:
– How?
Columnar Functions:
Integration and Broadcasting
• Integration: A column is activated if it receives enough
activation from
– Other columns
– Thalamus
• Can be activated to varying degrees
• Can keep activation alive for a period of time
• Broadcasting: An activated column transmits activation to
other columns
– Exitatory
– Inhibitory
• Learning : adjustment of connection strengths and thresholds
Integration and Broadcasting
 Broadcasting
• To multiple
locations
• In parallel
 Integration
Integration and Broadcasting
Broadcasting
Integration
Now I’ll tell my friends!
Wow, I got activated!
Processing in the cortex
•
•
•
•
Parallel (distributed) and serial
Hierarchical
Bidirectional
Variable
– Varying strengths of connections
– Varying degrees of activation
– Variation over time
• Adaptability
• Learning
• Plasticity
Uniformity of structure and function
• Locally,
– All cognitive and perceptual information, of any kind,
is represented as nodes and their interconnections
– All cognitive processing, of any kind, consists of
broadcasting and integration
Complexity from simplicity
• Complexity: what the brain can do
• Simplicity: every node is a simple processor
– Integration
– Broadcasting
– Changes in connection strengths and thresholds
• Problem: how can such simplicity produce such complexity?
• Answer:
– Huge quantity of nodes and connections
– Parallel distributed processing
– Hierarchical organization
Topics
•
•
•
•
•
A little neuroanatomy
Functional webs
Nodes and links: Cortical columns
Basic operations in the cortex
More operations: Learning
Additional operations: Learning
• Links get stronger when they are successfully used
(Hebbian learning)
– Learning consists of strengthening them
– Hebb 1948
• Threshold adjustment
– When a node is recruited its threshold increases
– Otherwise, nodes would be too easily satisfied
Requirements that must be assumed
(implied by the Hebbian learning principle)
• Links get stronger when they are successfully used (Hebbian learning)
– Learning consists of strengthening them
• Prerequisites:
– Initially, connection strengths are very weak
• Term: Latent Links
– They must be accompanied by nodes
• Term: Latent Nodes
– Latent nodes and latent connections must be available for learning
anything learnable
• The Abundance Hypothesis
– Abundant latent links
– Abundant latent nodes
Support for the abundance hypothesis
• Abundance is a property of biological systems generally
– Cf.: Acorns falling from an oak tree
– Cf.: A sea tortoise lays thousands of eggs
• Only a few will produce viable offspring
– Cf. Edelman: “silent synapses”
• The great preponderance of cortical synapses are
“silent” (i.e., latent)
– Electrical activity sent from a cell body to its axon travels
to thousands of axon branches, even though only one or
a few of them may lead to downstream activation
Locations of available latent connections
• Local
– Surrounding area
– Horizontal connections (not white matter)
• Intermediate
– Short-distance fibers in white matter
– For example from one gyrus to neighboring gyrus
• Long-distance
– Long-distance fiber bundles
– At ends, considerable branching
Learning – The Basic Process
Latent
nodes
Latent
links
Dedicated
nodes and
links
Learning – The Basic Process
Latent
nodes
Let these
links get
activated
Learning – The Basic Process
Latent
nodes
Then these
nodes will get
activated
Learning – The Basic Process
That will
activate
these links
Learning – The Basic Process
This node
gets enough
activation to
satisfy its
threshold
Learning – The Basic Process
This node is
therefore
recruited
B
A
These links
now get
strengthened
and the node’s
threshold gets
raised
Learning – The Basic Process
This node is
now dedicated
to function AB
AB
B
A
Learning
Next time it
gets activated
it will send
activation on
these links to
next level
AB
B
A
Learning: Deductions from
the basic process
• Learning is generally bottom-up.
• The knowledge structure as learned by the cognitive
network is hierarchical — has multiple layers
• Hierarchy and proximity:
– Logically adjacent levels in a hierarchy can be
expected to be locally adjacent
• Excitatory connections are predominantly from one
layer of a hierarchy to the next
• Higher levels will tend to have larger numbers of
nodes than lower levels
Learning in cortical networks:
A Darwinian process
• It works by trial-and-error
– Thousands of possibilities available
• The abundance hypothesis
– Strengthen those few that succeed
• “Neural Darwinism” (Edelman)
• The abundance hypothesis
– Needed to allow flexibility of learning
– Abundant latent nodes
• Must be present throughout cortex
– Abundant latent connections of a node
• Every node must have abundant latent links
Learning – Enhanced understanding
• This “basic process” is not the full story
• The nodes of this depiction:
– Are they minicolumns, maxicolumns, or what?
– Nodes of the model may be represented by
• Minicolumns
or
• Contiguous bundles of minicolumns
– Of different sizes
» “maxicolumns”, “hypercolumns”
Findings of Mountcastle:
Columns of different sizes for
categories and subcategories
• Minicolumn
– The smallest unit
– 70-110 neurons
• Functional column
– Variable size – depends on experience
– Intermediate between minicolumn and
maxicolumn
• Maxicolumn (a.k.a. column)
– 100 to a few hundred minicolumns
• Hypercolumn
– Several contiguous maxicolumns
Hypercolums: Modules of maxicolumns
A visual area
in temporal
lobe of a
macaque
monkey
Perceptual subcategories and
columnar subdivisions of larger columns
• Nodal specificity applies for maxicolumns as well
as for minicolumns
• The adjacency hypothesis likewise applies to
larger categories and columns
– Adjacency applies for adjacent maxicolumns
• Subcategories of a category have similar function
– Therefore their cardinal nodes should be in
adjacent locations
Functional columns
• The minicolumns within a maxicolumn respond to a
common set of features
• Functional columns are intermediate in size between
minicolumns and maxicolumns
• Different functional columns within a maxicolumn are
distinct because of non-shared additional features
– Shared within the functional column
– Not shared with the rest of the maxicolumn
Mountcastle: “The neurons of a [maxi]column have
certain sets of static and dynamic properties in
common, upon which others that may differ are
superimposed.”
Similarly..
• Neurons of a hypercolumn may have similar response features,
upon which others that differ may be superimposed
• Result is maxicolumns in the hypercolumn sharing certain basic
features while differing with respect to others
• Such maxicolumns may be further subdivided into functional
columns on the basis of additional features
• That is, columnar structure directly maps categories and
subcategories
(!)
Hypercolumns: Modules of maxicolumns
A visual area
in the
temporal lobe
of a macaque
monkey
Category
(hypercolumn)
Subcategory
(can be further
subdivided)
Learning in a system with
columns of different sizes
• At early learning stage, maybe a whole hypercolumn gets
recruited
• Later, subdivided into maxicolumns for further distinctions
• Still later, functional columns as subcolumns within
maxicolumns
• New term: Supercolumn – a group of minicolumns of whatever
size, hypercolumn, maxicolumn, functional column
• Links between supercolumns will thus consist of multiple fibers
Revisit the diagram: Each node of the diagram represents a
group of minicolumns – a supercolumn
Latent
supercolumns
Bundles
of latent
links
Dedicated
supercolumns
and links
Learning – The Basic Process
Let these
links get
activated
Learning – The Basic Process:
Refined view
Then these
supercolumns
get activated
Learning – The Basic Process:
Refined view
That will
activate
these links
Learning – Refined view
This
supercolumn
gets enough
activation to
satisfy its
threshold
Learning – Refined view
This supercolumn is
recruited for
function AB
AB
B
A
Learning:
Refined view
Next time it
gets activated
it will send
activation on
these links to
next level
AB
B
A
Learning
Refined view
Can get
subdivided for
finer
distinctions
AB
B
A
A further enhancement
• Minicolumns within a supercolumn have
mutual horizontal excitatory connections
• Therefore, some minicolumns can get
activated from their neighbors even if they
don’t receive activation from outside
Learning:
Refined view
AB
Hypercolumn
composed of 3
maxicolumns
Can get
subdivided for
finer
distinctions
B
A
Learning: refined view
If, later, C is
activated along with
A and B, then
maxicolumn ABC is
recruited for ABC
ABC
AB
B
A
C
Learning:
And the
connection from
C to ABC is
strengthened –it
is no longer
latent
refined view
ABC
AB
B
A
C
Topics
•
•
•
•
•
A little neuroanatomy
Functional webs
Nodes and links: Cortical columns
Basic operations in the cortex
More operations: Learning
Thank you for your attentIon!
References
Lamb, Sydney, 1999. Pathways of the Brain: The
Neurocognitive Basis of Language. John Benjamins.
Mountcastle, Vernon, 1998. Perceptual Neuroscience:
The Cerebral Cortex. Harvard University Press.
Pulvermüller, Friedemann, 2002. The Neuroscience of
Language. Cambridge University Press
Internet Sources
www.rice.edu/langbrain
www.owlnet.rice.edu/ling411/ClassNotes
For further information..
www.rice.edu/langbrain
[email protected]
The two big problems of neurosyntax
How does the brain handle..
Sequencing – ordering of words in a sentence
1.
And ordering of phonemes in a word
–
2.
Categories
–
Noun, Verb, Preposition, etc.
• Subtypes of nouns, verbs, etc.
–
–
–
–
What categories are actually used in syntax?
How are syntactic categories defined?
How represented in the brain?
How does a child build up knowledge of such categories
based on just his/her ordinary language experience?
First step: accounting for sequence
• Important not just for language
– Dancing
– Eating a meal
– Events of the day, of the year, etc.
– Etc., etc.
• In language, not just syntax (lexotactics)
– Ordering of morphemes in a word
• Morphotactics
– Order of phonological elements in syllables
• Phonotactics
Neurological Structures for Sequence
• How is sequencing implemented in neural
structure?
• For an answer, consider the structure of the
cortical column
Lasting activation in minicolumn
Cell Types
Recurrent axon
branches keep
activation alive in
the column –
Until is is turned
off by inhibitory
cell
Pyramidal
Spiny
Stellate
Inhibitory
Connections to
neighboring
columns not
shown
Subcortical
locations
The ‘Wait’ Element
1
2
www.ruf.rice.edu/~lngbrain/neel
Lasting activation in minicolumn
Cell Types
Recurrent axon
branches keep
activation alive in
the column –
Until is is turned
off by inhibitory
cell
Pyramidal
Spiny
Stellate
Inhibitory
Connections to
neighboring
columns not
shown
Subcortical
locations
Simple notation for lasting activation
Thick border for a node
that stays active for a
relatively long time
Thin border for a node
that stays active for a
relatively short time
N.B.: Nodes are implemented as cortical columns
Recognizing items in sequence
This link
stays
active
This node
recognizes the
sequence ab
c
a
b
Node c is satisfied by activation from both a and b
If satisfied it sends activation to output connections
Node a keeps itself active for a while
Suppose that node b is activated after node a
Then c will recognize the sequence ab
Example: eat apple
(structure for recognition)
(Just labels,
not part of the
structure)
eat apple
eat
apple
Example: eat apple, eat banana
(structure for recognition)
eat apple
eat
apple
eat banana
banana
Producing items in sequence
Wait element
ab
a
b
First a, then b
How does the delay element work?
• Remember: each node is implemented as a cortical column
– Within the column are 75-110 neurons
• Enough for considerable internal structure
• When node ab receives activation, it
– Sends activation on down to node a
– And to the delay element, which
• Waits for activation from clock timer or feedback
– Will come in on line labeled ‘f’ in diagram
• Upon receiving this signal, sends activation on to node b
Producing items in sequence
Delay element
ab
a
f
b
Carries feedback
or clock signal
Producing items in sequence
May be within one
cortical column
ab
a
f
b
Producing items in sequence
a different means
a
b
f
This would apply for items ‘a’ and ‘b’ in
sequence where there is no ‘ab’ to be
recognized as a unit.
Example: Adjectives of size precede
adjectives of color, which precede
adjectives of material in the English noun
phrase, as in big brown wooden box
Two different network notations
Narrow notation
•
•
•
•
Nodes represent cortical columns
Links represent neural fibers
Uni-directional
Close to neurological structure
eat apple
eat
apple
Abstract notation
•
•
•
•
Nodes show type of relationship (OR,
AND)
Easier for representing linguistic
relationships
Bidirectional
Not as close to neurological structure
eat apple
eat
apple
Two different network notations
Narrow notation
b
a
ab
Upward
b
f
a
ab
Abstract notation

Bidirectional
a
b
Downward
b
Constructions have meanings and functions
• They are also signs
Meaning/Function
Form/Expression
The sign
relationship:
a (neural)
connection
The difference is that for a construction the
expression is variable rather than fixed
The transitive verb phrase construction
Semantic
function
Syntactic
function
CLAUSE
DO-TO-SMTHG
Transitive verb phrase
Variable
expression
Vt
NP
Linked constructions
The clause
construction
CL
DO-TO-SMTHG
NP
Transitive verb phrase
Vt
NP
Add a few more connections
ACTOR-DO
CL
DO-TO-SMTHG
Transitive verb phrase
Vt
NP
Add other types of predicate
THING-DESCR
CL
BE-SMTHG
DO-TO-SMTHG
Vi
(A rough first
approximation)
Vt
be
Adj
NP
Loc
The other big problem for syntax
• Categories
• Problems of categories are considered in a separate
presentation
Descargar

Document