Transcript DOWN - Ubiquitous Computing Lab

Soft Computing
Lection 2
General characteristics of minds/brains that
contemporary researchers in AI and cognitive
science are trying to understand and replicate:
Perception - manipulation, integration, and interpretation of data provided by sensors (in the context of the internal state of the system |
including purposeful, goal-directed, active perception).
Action - coordination, control, and use of effectors to accomplish a
variety of tasks including exploration and manipulation of the
environment,
including design and construction of tools towards this end.
Reasoning - deductive (logical) inference, inductive inference,
analogical inference | including reasoning in the face of uncertainty and
incomplete information, hypothetical reasoning, justication and
explanation of inferences, evaluation of explanations, adapting
explanations in the light of falsied assumptions or changing world
states.
Adaptation and Learning - adapting behavior to better cope with
changing environmental demands, discovery of regularities,
explanation of observations in terms of known facts and hypotheses,
construction of task-specic internal representations of the
environment, discovery of procedures, learning to differentiate despite
similarities and generalize despite differences, learning to describe
specific domains in terms of abstract theories and concepts, learning
to use, adapt, and extend language, learning to reason, plan, and act.
Communication - with other intelligent agents including humans
using signals, signs, icons, symbols, sound, pictures, touch, language
and other communication media | including communication of goals,
desires, beliefs, narratives of real and imaginary episodes, explanation
of actions and events.
Planning and goal-directed problem-solving - Formulation of
plans |sequences or agenda of actions to accomplish externally or
internally determined goals, evaluating and choosing among
alternative plans, adapting plans in the face of unexpected changes
in the environment, explaining and justifying plans, modifying old
plans to t new tasks, handling complexity by abstraction and
simplification.
Autonomy - Setting of goals, deciding on the appropriate course of
actions to take in order to accomplish the goals or directives (without
explicit instructions from another entity), executing the actions to
satisfy the goals, adapting the actions and/or goals as necessary to
deal with any unforeseen circumstances (to the extent permitted by
the agent's physical capabilities and the environmental constraints).
Creativity - exploration, modification, and extension of domains (e.g.,
language, mathematics, music) by manipulation of domain-specific
constraints, or by other means.
Reflection and awareness - of internal processes (e.g., reasoning, goals,
etc.) of self as well as other agents. Aesthetics | articulation and use of
aesthetic principles.
Organization - into social groups based on shared objectives, development
of shared conventions to facilitate orderly interaction, culture.
Mitchell (Carnegie Mellon
University):
The synergy between AI and Brain Sciences will yield
profound advances in our understanding of intelligence
over the coming decade, fundamentally changing
the nature of our field
The synergy between AI and Brain Sciences will
yield profound advances in our understanding of
intelligence over the coming decade (said in 2002).
1. Common goal: understand intelligence
2. Significant correspondences between AI
methods and brain organization
3. New instrumentation is causing a
revolution
Human brain
Two-level model of mind
“All models are wrong, but some are useful” (George Box, 1979).
Logical (verbal) thinking
F A(D,D)
Signs (symbols)
Control of associations
by consciousness
F F(K,K)
Forming of signs from images by
classification and recognition
Associative (creative)
thinking
f(K,A)
Images
Environment
D - signs, K – images, A - actions
Consciousness and
subconsciousness
actions
Visual images
consciousness
Internal images
i. g. pain
Subconsciousness
Sound images
Tactile images
Taste images
Smell images
Basic tasks of associative level
• Recognition – relating of image (pattern) to any
determined class
• Classification - The process of learning to relate
of image (pattern) to one of set of determined
classes.
• Clustering - The process of grouping similar
images (patterns) together in cluster (may be
named as class) and forming set of classes
during learning
• Forming of associative links between images
and between classes
• Associative search images or classes similar to
any input image (pattern)
Basic tasks of logical level
• Forming of signs (words, symbols,
formulas and so on) and links between it
and any class
• Forming of structures consists of signs
(Trees, lists, formulas, sentences and so
on) may be named as concepts
• Search signs connected start sign
(inference)
• Here the concept of context appears
Process of thinking
Division natural mind into two levels is relative.
Concept, class (may be using for reasoning
and to be on different “length” from sensors)
Influence
of context
Process of
classificationrecognition
Primary features from different
sensors
Process of thinking
Process of thinking may be viewed as sequence of firing of set
of neurons – on associative level power of set is larger than on
logical level
Reasoning
on logical level
Neuron
Associative search
on image level
Associations, classification and
fuzzy analogy
• Association – link created when any
different images were firing together
during process of thinking,
• Could say that between these images
exist fuzzy analogy (or similarity) (different
from analogy in knowledge engineering
based on formalized relation of similarity),
• Couple of fuzzy similar Images may be
recognized as related to same class
Examples of similar images
Class
“face of
woman”
Class “face of man”
All images relates
to class “faces”
Forming of mean of word or name
(sign) of class
Associative link
Classification
(rocognition) of
visual images
face
Classification
(recognition) of
acoustic images
Any formal definitions
Set of features K={pi}| i=1,Np, describing state of environment
and self intelligent system in time t, where Np – the number
of features,
Set of combination of values of features on set
K ={Pj} | Pj={pij} | j=1,No, i=1,Np, describing concrete images,
where No – number of images,
Set of real images (it not includes full set of features)
Ψ={Pkj} | j=1,No and k is integer from (1,Np),
Query (image, initializing associative search) P Ψ,
Image-result of associative search R Ψ.
Any formal definitions
May be two different processes:
1) The process of restoration of image by partially determined
features. Usually this process is simulated in different models of
associative memory from memory based on Hopfield model
to memory based on spike neurons;
2) The process of searching of associatively connected images
linked with different moments of time.
These images mean reasons or consequences of initial
image.
First variant is implemented in natural intelligent
systems in sensor subsystems of brain.
Second – in neocortex and one is main for
forecasting and thinking of animal or man.
Any formal definitions
The pair of images (P,R) may be called
an association A or A(P,R)
Set of associations A={Ai(Pi,Ri)} | i(1,M)
forms memory or knowledge base of intelligent system.
Predicate (Pa,Ra,Ta), describing process of restoring
of Ra | Ra  R by Pa | Pa  P, is called as associative search,
Pa – initial image of associative search and
Ra - final image of associative search, Ta – duration of
associative search
Such associative search (Pa,Ra,Ta), as it use only one
association from memory A=(P,R) | PaP, RaR,,
may be called elementary associative search.
Process of associative search
Result - image
Used association
Initial image
Models of logical (symbol) level in
knowledge engineering and simulation of
mind top-down
• 1-order logic
• Other logics based on boolean logic
• Rules
• Semantic nets
• Frames
Attempts to include in these models fuzziness:
Fuzzy logic,
Linguistic variables,
Probabilistic reasoning.
Models of associative (image) level by
neural networks and simulation of mind
bottom-up
• Different model of neural networks
Attempts to include in neural network
forming of signs (concepts, words):
• Semantic neural networks
• Fuzzy neural networks
• Ensemble neural networks
Usual performance about correlation
between features of brain and
consciousness
(1) patterns of neural activity correlate with
mental states;
(2) synchronous network oscillations of
neuronal circuits in the thalamus and
cerebral cortex temporarily binds
information;
(3) consciousness emerges as a novel
property of computational complexity
among neurons.
Other performance about brain
Stuart Hameroff, Roger Penrouse
However, these approaches appear to fall short in fully explaining
certain enigmatic features of consciousness, such as
• the nature of subjective experience, or “qualia”—our “inner life”
(Chalmers’ “hard problem,” 1996)
• the binding of spatially distributed brain activities into unitary
objects in vision, and a coherent sense of self, or “oneness”
• the transition from preconscious processes to consciousness itself
• noncomputability, or the notion that consciousness involves a factor
that is neither random nor algorithmic, and that consciousness
cannot be simulated (Penrose, 1989, 1994, 1997)
• free will
• subjective time flow.
However, in fitting the brain to a computational
view, such explanations omit incompatible
neurophysiological details, for example:
• widespread apparent randomness at all levels of
neural processes (is it noise or underlying levels
of complexity?)
• glial cells (which accounts for some 80 percent of
the brain)
• dendritic-dendritic processing
• cytoplasmic/cytoskeletal activities
Quantum theory of mind
Activities within cells ranging from single-celled
organisms to the brain’s neurons are organized by a
dynamic scaffolding called the cytoskeleton. A major
component of the cytoskeleton is the microtubule, a
hollow, crystalline cylinder 25 nm in diameter.
Microtubules are, in turn, composed of hexagonal
lattices of proteins, known as tubulin.
Quantum theory of mind
Microtubule automaton switching offers a potentially vast
increase in the computational capacity of the brain. While
conventional approaches focus on synaptic switching at the
neural level, which optimally yields about 1018 operations per
second in human brains (~1011 neurons per brain, with ~104
synapses per neuron, switching at ~103 sec–1), microtubule
automata switching can explain some 1027 operations per
second (~1011 neurons with ~107 tubulins per neuron,
switching at ~109 sec–1). Indeed, the fact that all biological
cells typically contain approximately 107 tubulins could
account for the adaptive behaviors of single-celled
organisms, which have no nervous system or synapses.
Rather than simple switches, then, it seems that neurons are
actually complex computers.