Transcript Cognitive Neuroscience

Computational Intelligence
Cognitive
Neuroscience
Based on a course taught by
Prof. Randall O'Reilly
University of Colorado,
Prof. Włodzisław Duch
Uniwersytet Mikołaja Kopernika
and http://wikipedia.org/
http://grey.colorado.edu/CompCogNeuro/index.php/CECN_CU_Boulder_OReilly
http://grey.colorado.edu/CompCogNeuro/index.php/Main_Page
Janusz A. Starzyk
EE141
1
The Brain ...

The most interesting and the most complex
object in the known universe

How can we understand the workings of
the brain?

On what level should we attack this
question? An external description won’t
help much.

How can we understand the workings of a TV or computer?

Experiments won’t suffice, we must have a diagram and an
understanding of operational principles.

To make certain that we understand how it works, we must make a
model.
2
EE141
How do we know anything?
An important question: how do we know things?
Example: super diet based on dr. K, Chinese medicine
and other miracle methods. How do we know that
they work? How do we know that they are for real?
Gall noticed that the skull shape decides about ones
abilities. Thousands of cases confirmed his observations.
Craniometry: measuring the bones of the skull
to determine intelligence.
Do I know or I only believe I know?
Not being certain allows to learn, certainty makes
learning difficult. If we know how easy it is to err
we could avoid a scientific fallacy.
3
EE141
How to understand the brain?
To understand: reduce to simpler mechanisms?
Which mechanisms? Analogies with computers? RAM, CPU? Logic?
Those are poor analogies.
Psychology: first you must describe behavior, it looks for explanations
most often on a descriptive level, but how to understand them?
Physical reductionism: mechanisms of the brain.
Reconstructionism: using mechanisms to reconstruct the brain’s functions
We can answer many questions only from an ecological and evolutionary
perspective: why is the world the way it is? Because that’s how it made
itself ... Why does the cortex have a laminar and columnar structure?
To create: what must we know in order to create an artificial brain?
EE141
4
From molecules through neural networks
10-10 m, molecular level: ion channels, synapses, properties of cell
membranes, biophysics, neurochemistry, psychopharmacology;
10-6 m, single neurons: neurochemistry, biophysics, LTP,
neurophysiology, neuron models, specific activity detectors,
emerging.
10-4 m, small networks: synchronization of neuron activity, recurrence,
neurodynamics, multistable systems, pattern generators, memory,
chaotic behaviors, neural encoding; neurophysiology ...
10-3 m, functional neural groups: cortical columns (104-105), group
synchronization, population encoding, microcircuits, Local Field
Potentials, large-scale neurodynamics, sequential memory,
neuroanatomy and neurophysiology.
5
EE141
… to behavior
10-2 m, mesoscope networks: sensory-motor maps, self-organization,
field theory, associative memory, theory of continuous areas, EEG,
MEG, PET/fMRI imaging methods ...
10-1 m, transcortical fields, functional brain areas: simplified cortical
models, subcortical structures, sensory-motor functions, functional
integration, higher psychic functions, working memory,
consciousness; (neuro)psychology, computer psychiatry ...
Cognitive effects
Principles of
interactions
Neurobiological
mechanisms
6
EE141
Levels of description
Summary (Churchland, Sejnowski 1988)
7
EE141
EE141
How does it all work?
8
Systemic level
9
EE141
… to the mind
Now a miracle happens ...
1 m, CNS, the whole brain and organism:
An interior world arises, intentional behaviors, goal-oriented
actions, thought, language, everything that behavioral psychology
examines.
Approximations of neural models:
Finite State Machine, rules of behavior, models based on the
knowledge of cognitive mechanisms in artificial intelligence.
What happened to the psyche, the internal perspective?
Lost in translation: networks => finished machines => behavior
10
EE141
A neurocognitive approach
Computational cognitive neuroscience: detailed models of cognitive
functions and neurons.
Neurocognitive computing: simplified models of higher cognitive
functions, thinking, problem solving, attention, language, cognitive and
behavioral controls.
Lots of speculation, but qualitative models explaining the results of
psychophysical experiments as well as the causes of mental illnesses are
developing quickly.
Even simple brain-like information processing yields results similar to the
real ones! Forewarning against excessive optimism based on behavioral
models.
11
EE141
Model of transformation
Agent Architecture
Reason
Short-term Memory
Perceive
Act
RETRIEVAL
LEARNING
Long-term Memory
INPUT
OUTPUT
Task
Environment
Simulation or
Real-World System
EE141
From
Randolph M. Jones, P : www.soartech.com
12
Model of self-organization
Topographical representations in numerous areas of the brain:
sensory impulses, in the motor cortex and cerebellum, multimodal maps
of orientation inferior colliculus, visual system maps and maps of the
auditory cortex.
o
Model (Kohonen 1981):
competition between groups of
neurons and local cooperation.
x=data
o=weights of
neurons
x
o
o
o
o x
o
o
x
o
xo
N-dimensional
input space
o
o
o
Neurons react to signals
adjusting their parameters so
that similar impulses awaken
neighboring neurons.
Weights locate
points in N-D
neural network
w 2-D
13
EE141
Dynamic model
Strong feedback, neurodynamics.
Hopfield model: associative memory, learning based on Hebb’s law,
synchronized dynamics, two-state neurons.
Vector of input potentials V(0)=Vini , i.e.
input = output.
Dynamics (iterations) 
Hopfield’s network reaches stationary
states,
or the answers of the network (vectors of
elemental activation) to the posed question
Vini (autoassociation).
If the connections are symmetrical then
such a network trends to a stationary state

(local attractor).
Vi  t  1  sgn I i  t  1  sgn 
t = discrete time.


EE141


j WijV14j   j 

Biophysical model – spiking neurons
Synapses
Soma
I syn (t )
Spike
EPSP, IPSP
Rsyn
Csyn
Spike
Cm
Rm
“Spiking Neuron Models”,
W. Gerstner and W. Kistler
Cambridge University Press, 2002
http://icwww.epfl.ch/~gerstner//SPNM/SPNM.html
EE141
15
Molecular foundations
Action potentials are the result of currents
which flow through ionic channels in the
cell membrane
Hodgkin and Huxley measured these
currents and described their dynamics
through differential equations.
-70mV
Na+
Action
potential
K+
Ca2+
Ions/protein
EE141
16
Hodgkin-Huxley model
100
inside
I
K
mV
C
gK
gNa gl
Na
outside
0
Ion channels
sodium
I Na
potassium
Ion pump
leakage
IK
stimulus
I leak
du
C
 g Na m3h(u  ENa )  g K n 4 (u  EK )  gl (u  El )  I (t )
dt
dh
dn
dm hnmhn0m
()u )
0((u
0u)

dt
dt
dt
hn((muu()u) )
EE141
The likelihood the channel is open is described
by extra variables m, n, and h.
17
Impulse response model
Activation
j
 t  ti^ 
i
Stimulus: EPSP
 t  t
f
j
Activation: AP
 t  t
^
i
ui

Stimulus: EPSP
 t  t

ui t    t  t  
ui t    
EE141
 w  t  t 
ij
j
Firing:

All impulses and neurons
Previous impulse i
^
i
f
j

f
j
linear
f
t t
^
i
threshold
18
Integration and activation model
Activation
j
i
ui
I


reset
Stimulus : EPSP
d
  ui  ui  RI (t )
dt
ui t     Fire+reset
EE141
linear
 t  t jf 
threshold
19
Psychological Phenomena
Visual perception: viewing natural imagery
we must understand ways of encoding
obiects and scenes.
Spatial awareness: considering the interaction
between streams of visual information will let
us simulate concentration
Memory: modeling hippocampal structures allows us to understand
various aspects of episodic memory, and learning mechanisms show how
semantic memory arises.
Working memory: explaining the capacity to simultaneously hold in the
mind several numbers while performing calculations requires specific
mechanisms in the neural model.
20
EE141
Psychological Phenomena
Reading words: the network will learn to read and pronounce words and
then to generalize its knowledge to the pronunciation of new words as
well as to recreate certain forms of dyslexia.
Semantic representations: analyzing a text on the basis of context, the
appearance of individual words, the network will learn the semantics of
many ideas.
Decision-making and task execution:
A model of the prefrontal cortex will be
able to keep attention on performed
tasks in spite of hindering variables.
Development of the representation of
the motor and somatosensory cortex:
through learning and controlled selforganization;
21
EE141
Advantages of model simulations
Models help to understand phenomena:
 enable new inspirations, perspectives on a problem
 allow the simulation of effects of damages and disorders (drugs,
poisoning).
 help to understand behavior,
 models can be formulated on various levels of complexity,
 models of phenomena overlapping in a continuous fashion (e.g. motion
or perception),
 models allow detailed control of experimental conditions and an exact
analysis of the results
Models require exact specification of underlying assumptions:
 allow for new predictions
 perform deconstructions of psychological concepts (working memory?)
 allow us to understand the complexity of a problem
 allow for simplifications enabling analysis of a complex system
 provide a uniform, cohesive plan of action
22
EE141
Disadvantages of simulations








Models are often too simple, they should contain many levels.
Models can be too complex, sometimes theory allows for simpler
explanations (why are there no hurricanes on the equator?).
It’s not always known what to provide for in a model.
Even if models work, that doesn’t mean that we understand the
mechanisms
Many alternative yet very different models can explain the same
phenomenon.
What’s important are general rules, parameters are limited by
neurobiology on various levels; the more phenomena a model
explains, the more plausible and universal it is.
Allowing for interactions and emergences (construction) is very
important.
Knowledge acquired from models should undergo accumulation.
23
EE141
Cognitive motivation
 Although
the thinking process seems to be sequential information
processing, more detailed models predict parallel processing
 Gradual transition between conscious and subconscious processes
 Parallel processing of sensory-motor signals by tens of millions of
neurons

Specialized areas of memory responsible for various representations
e.g. shape, color, space, time
 Levels of symbolic representation
 More diffuse than binary logic

Learning mechanisms as a foundation for cognitive science
 When you learn, you change the method of information processing in
your brain
Resonance between ”bottom-up” representation and ”top-down”
understanding
 Prediction and competition of ideas

24
EE141