Towards Computational Models of Artificial Cognitive Systems that

Report
Monkey Before the
Skeleton (Ecce simia),
Gabriel von Max, Prague
painter (1840-1915)
Towards Computational Models
of Artificial Cognitive Systems
That Can, in Principle,
Pass the Turing test
Jiri Wiedermann
Institute of Computer Science, Prague
Academy of Sciences of the Czech Republic
SOFSEM 2012 January 21-27, 2012
Spindleruv Mlyn
Partially supported GA CR
grant No. P202/10/1333
``I believe that in about fifty years' time it will be
possible, to program computers, with a storage capacity
of about 100 kB, to make them play the imitation game
so well that an average interrogator will not have more
than 70 % chance of making the right identification
after five minutes of questioning.
The original question, "Can machines think?" I believe to
be too meaningless to deserve discussion. Nevertheless
I believe that at the end of the century the use of
words and general educated opinion will have altered so
much that one will be able to speak of machines thinking
without expecting to be contradicted."
From the discussion between Turing and one of his colleagues (M. H. A. Newman, professor of
mathematics at the Manchester University):
Newman: I should like to be there when your match between a man and a machine takes place,
and perhaps to try my hand at making up some of the questions. But that will be a long time from
now, if the machine is to stand any chance with no questions barred?
Turing: Oh yes, at least 100 years, I should say.
Three heretic ideas:
We already have a sufficient knowledge to understand the working of
interesting minds achieving a high-level cognition
Achieving a higher-level AI is not a matter of a fundamental scientific
breakthrough but rather a matter of exploiting our best theories of
artificial minds, and a matter of scale, speed and technological achievements
It is unlikely that thinking machines will be developed by purely academic
research since it is beyond its power to concentrate the necessary amount of
man power and technology.
Approaches to mind understanding:
Understanding by philosophying
Understanding by designing (specifying)
Understanding by constructing
Outline
Current state: Watson the Computer vs. humanoid robotic
systems
Winds of Change
1.
2.
–
–
–
–
–
–
–
–
–
3.
4.
Escaping the Turing test
Escaping Biologism
Internal World Models
Mirror neurons
Global Workspace Theory
(Dis)solving the Hard Problem of Consciousness
Episodic Memories
Real Time Massive Data Processing
Comprehensive and Up-To-Date Models of Cognitive Systems
HUGO: A Non-Biological Model of a Conscious Agent
System
Conclusions – lessons from what we have seen
Watson - an AI system capable to answer
the questions stated in natural language
Jeopardy! (in the CR – the TV game „Riskuj!“) – given an
answer one has to guess the question.
E.g.: 5280 (how many feets has a mile), or
79 Wistful Vista (address of Fibber and Molly McGee)
Category: General Science
Clue: When hit by electrons, a phosphor gives off
electromagnetic energy in this form.
Answer: Light (or Photons)
Category: Rhyme Time
Clue: It’s where Pele
stores his ball.
Subclue 1: Pele ball
(soccer)
Subclue 2: where store
(cabinet, drawer, locker,
and so on)
Answer: soccer locker
Category: Lincoln Blogs
Clue: Secretary Chase just submitted this to me for the third
time; guess what, pal. This time I’m accepting it.
Answer: his resignation
Category: Head North
Clue: They’re the two states you could be reentering if you’re
crossing Florida’s northern border.
Answer: Georgia and Alabama
Source: AI Magazine, Fall 2010
Winds of Change
New trends in theory:
• escaping biologism
• escaping the Turing Test
• strengthening the position of embodiment: a common sensorimotor
basis for phenomenal and functional consciousness
• evolutionary priority of phenomenal consciousness over functional one
• internal world models, mirror neurons
• global workspace theory
• episodic memory
Technological progress:
• maintenance of supercritical volumes of data, and
• searching and retrieval of data by supercritical speed
A shift in popular thinking about artificial minds
- people generally accept that computers can think (albeit in a different
sense than some philosophers of mind would like to see)
John Searle: “Watson Doesn't Know It Won on 'Jeopardy!'
IBM invented an ingenious program—not a computer that can
think.”
Noam Chomsky: “Watson understands nothing. It’s a bigger
steamroller. Actually, I work in AI, and a lot of what is done
impresses me, but not these devices to sell computers.”
What these gentlemen failed to see
is the giant leap
from the formal rules of the chess playing to informality of Jeopardy! rules…
J.R. Lipton: Big insight – a program can be immensely powerful even if it is
imperfect.
A new trend: escaping biologism
Why should we only think about human brain when
designing artificial minds?
Rodolfo Llinas (a prominent neuroscientist):
“I must tell you one of the most alarming
experiences I've had in pondering brain function....
that the octopus is capable of truly extraordinary
feats of intelligence… most remarkable is the
report that octopi may learn from observing other
octopi at work. The alarming fact here is that the
organization of the nervous system of this animal
is totally different from the organization we have
learned is capable of supporting this type of
activity in the vertebrate brain.... there may well
be a large number of possible architectures that
could provide the basis of what we consider
necessary for cognition and qualia....
Many possible
architectures
for cognition
A new trend: escaping the Turing test
Turing test is explicitly anthropomorphic.
Russell and Norvig: "aeronautical engineering texts do not
define the goal of their field as 'making machines that fly so exactly
like pigeons that they can fool other pigeons’”.
All minds
Human
mind
Animal
minds
Artificial
minds
Alien
minds
A new trend: Internal World Models
Mechanisms situating an
agent in its environment ;
they determine the syntax
and the semantic of agent
behavior and perception in its
environment
Finite control
(Infinite) stream
of inputs generated
by sensory-motor
interaction
IWMs capture a “description”
of that (finite) part of the
world and that part of the
self which has been “learned”
by agent’s sensori-motor
activities. An IWM is fully
determined by the agent’s
embodiment and is
World model
Sensory-motor
units
automatically built during
The body
agent’s interaction with the
real world.
A virtual inner world in which an agent can think
A new trend: Mirror neurons –
a mechanism for “mind reading” of other subjects
Mirror neurons: are active when a subject performs a specific action
as well as when the subject observes an other or a similar subject
performing a similar action (Rizzolatti, 199x)
“the discovery of mirror neurons in
the frontal lobes of monkeys, and
their potential relevance to human
brain evolution is the single most
important ``unreported“ (or at least,
unpublicized) story of the decade.
I predict that mirror neurons will do
for psychology what DNA did for
biology: they will provide a unifying
framework and help explain a host
of mental abilities that have hitherto remained mysterious and
inaccessible to experiments“
V.S. Ramachandran
A new trend: Global Workspace Theory
a simplistic, very high-level cognitive architecture that has been
developed by B. J. Baars by the end of the last century to explain
emergence of a conscious process from large sets of unconscious
processes in the human brain.
The GWT can
successfully model a
number of
characteristics of
consciousness, such as its
role in handling novel
situations, its limited
capacity, its sequential
nature, and its ability to
trigger a vast range of
unconscious brain
processes.
Interesting: Watson the Computer works according to the GWT
A new trend: evolutionary approach to phenomenal
consciousness (Inman Harvey)
A naive “incremental”
approach to create
phenomenal consciousness:
1. Create a “zombie” with
functional consciousness
(the easy problem)
2. Add the extra
ingredient to give it a
phenomenal
consciousness
(the hard problem)
“evolutionary approach allows emulation without comprehension”
A new trend:
a common sensorimotor basis for
phenomenal and functional consciousness
A sensorimotor interaction with the
environment involving corporality,
alerting capacity, richness,
insubordinateness, and the self
Instead of thinking of the brain as the
generator of feel, feel is considered
as a way of interacting with the world
Source: How to build a robot that feels. J.Kevin
O'Regan ,Talk given at CogSys 2010 at ETH Zurich
(Drawing by Ruth Tulving)
A new trend: Episodic Memory
is what people ``remember", i.e., the
contextualized information about
autobiographical events (times, places,
associated emotions), and other contextual
knowledge that can be explicitly stated.
An agent without episodic memory is like
a person with amnesia
Episodic memory systems allow “mental time travel” and can support a
vast number of cognitive capabilities based on inspecting memories
from the past that are ``similar" to the present situation, such as
• noticing novel situations,
• detecting repetitions,
• virtual sensing (reminded by some recall),
• future action modeling,
• planning ahead,
• environment modeling,
• predicting success/failure,
• managing long term goals, etc.
Efficient management and
retrieval from episodic
memories is a case for
real-time massive data
processing technologies.
A new trend: intelligence might be a matter of
scale and speed: maintaining supercritical volumes of data and
their searching and retrieval by supercritical speed (cf. episodic memories).
A lesson from Watson the Computer: intelligence
might not only be a matter of suitable algorithms, but also, and mainly so,
of the ability to accumulate (e.g., via learning and episodic memories
storing), organize, and exploit large data volumes representing knowledge
at a speed matching the timescale of the environmental requirements (real
time data processing).
Element
Number
of cores
Time to answer one
Jeopardy! question
Single core
1
2 hours
Single IBM Power 750
server
32
<4 min
Single rack (10 servers)
320
<30 seconds
IBM Watson (90 servers)
2 880
<3 seconds
~1 000 000
million
lines of code
5 years
development
(20 men)
Memory:
20 TB
200 million
pages
(~1 000 000
books)
A new trend: Comprehensive and up-to-date models
of cognitive systems
An urgent need of situatedness via embodiment
An embodied cognitive agent
is a robot i.e., an embodied
computer, which is a computer equipped by sensors by which it “perceives”
its environment and by effectors by which it interacts with its environment
(from J. A. Comenius, Orbis pictus, 1658)
Nuremberg funnel, Harsdörffer, Georg Philipp:
Poetischer Trichter, Nuremberg 1648-1653
HUGO: a Non-Biological Model of an Embodied
Conscious Agent
Semantic world model
Global workspace
Syntactic world model
Mirror net
Episodic
memory
From: J.
Wiedermann:
A High Level
Model of an
Embodied
Conscious
Agent,
IJSSCI, 2,
2010
A high-level schema of a robot:
The body
Finite control (a computer)
World model
(Infinite) stream of inputs
generated by sensory-motor
interaction
Sensory-motor units
Mechanisms situating the agent in its environment
must be considered: internal world models
Real world
The central idea:
Educating and Teaching a Robot
The purpose of
educating and teaching
an agent is to build its
internal world model
The internal world model gives a “description” of that (finite) part of
the world (inclusively of agent’s (it)self) which has been “learned” by
agent’s S-M activities.
The model is fully determined by the agent’s embodiment and is
automatically built during agent’s interaction with the real world
The idea of two cooperating world
models in cognitive systems
Dynamic world model:
sequences of sensorimotor information
Controls the agent’s behavior
Motor instructions
“action”
“cognition”
Real world
Static world model:
perception
Elements of a coupled
sensory-motor
information; responsible for situating the agent
Sensory-motor units
An architecture of an embodied cognitive agent
Control unit
G
r
o
u
Abstract
concepts
n
d
i
n
g
Multimodal
information
Embodied
concepts
Units of S-M
information
(World’s
“syntax”)
Mirror net
Symbolic level
Sub-symbolic level
Motor instructions
Motor
instructions
Perception
Body
Environment
S-M
units
The task of the syntactic world model:



Coupling the motor instructions with the perception information
into so-called multimodal information;
Learning frequently occurring multimodal information from the
coupled input streams (one coming from the dynamic model and
one from the S-M units)
Associative retrieval: a partial, or “damaged”, or previously
“unseen” incoming multimodal information gets completed so
that it corresponds to the “most similar” previously learned
information; the result captures the instantaneous agent’s
situation
The task of the semantic world model:


Learning (mining) and maintaining the knowledge from the datastream of multimodal information delivered by static
(syntactic) world model
Realizing the intentionality: with each unit of multimodal
information a sequence of actions (motor commands) – habits gets associated which can be realized in the given context;
Implementing the syntactic world model:
Mirror neurons: are active when a subject performs a specific action
as well as when the subject observes an other or a similar subject
performing a similar action (Rizzolatti, 199x)
A generalization: … a set of neurons which are active when a subject
performs any frequent action as well as when only partial information
related to that action is available to the subject at hand
Visual inf.
Aural inf.
Haptic
Propriocept.
Multimodal
information
• Learns frequently occurring conjunctions
of related input information
• It gets activated when only partially
excited (by one or several of its inputs)
• Works as associative memory, completing
the missing input information
• Mirror net forms and stores (pointers
to) episodic memories
The basis for understanding imitation learning, language acquisition,
thinking, consciousness.
What knowledge is mined and maintained in a dynamic
world model:
• often occurring concepts
• resemblance of concepts
• contiguity in time or place
• cause and effect
An algebra of thoughts…
Cognitive tasks:
1.
2.
3.
4.
5.
6.
David Hume 1711-1766
Simple conditioning
Learning of sequences
Operand conditioning (by rewards and punishment)
Imitation learning
Abstraction forming
Habits formation, etc. “Hume’s test” for intelligence
Implementing the dynamic world model
A cogitoid: an algorithm building a neural net for
knowledge-mining from the flow of multi-modal information
Excitatory and
inhibitory links
Multimodal
information
Passive
concepts
Currently
activated
Previously
concepts
activated
concepts
Habits: often followed
chains of concepts
Newly
activated
concepts
affect
aaaa
Emotions
Wiedermann 1999
What both world models jointly do for an agent:
A mechanism enabling imitation of activities of other
agents (without understanding)
 A germ of awareness – a mechanism for distinguishing
between one’s own action, and that of an observed agent
 A mechanism of empathy
 A substrate for a mechanism for predicting the results of
an agent’s own or observed actions via their “simulation” in
the virtual model of the known part of the real world
 Understanding: an agent “understands” its actions in terms
of their embodiment in terms of habits (and thus: of S-M
actions plus associated emotions)
 Phenomenal consciousness (according to
O’Regan) as a habit of conscious awareness of
performing one’s own skills

Humanoid Robot Mahru Mimics a Person's Movements in Real Time
A person wears the motion tracking suit while performing various tasks. The
movements are recorded and the robot is then programmed to reproduce the
tasks while adapting to changes in the space, such as a displaced objects.
The birth of communication and speaking
•
•
•
•
•
By indicating a certain
action an agent broadcasts
a visual information which
is completed by the
empathy and prediction
mechanism of an observing
agent into the intended
action
Formation of the self
concept
Possibility for emotions to
enter the game
The birth of the body
language
Adding of articulation
(vocalization) and
gesticulation tempering
•
•
The verbal component of the
language gets associated
with the motor of speech
organs and
prevails over
gesticulation
Development
of
episodic
memory management and
retrieval mechanisms
The birth of thinking
•
Beginning of thinking as a
habit of speaking to oneself
•
Multimodal
information
•
•
cogitoid
Motor
instructions
Mirror
neurons
Wiedermann 2004
Subsequent decay of whatever
motor activity (of vocal
organs)
Perception suppressing
Switching-off motor
instruction realization
Mirror neurons complete
motor instructions by missing
perception learned by
experience
An agent operates similarly
as before, albeit it
processes “virtual” data.
It works in an „off-line“
mode, it is virtually
situated
The birth of functional consciousness
The agents are said to possess artificial functional
consciousness iff their communication abilities reach such a
level that the agents are able to fable on a given theme.
More precisely, the conscious agents can
 Communicate in a high-level language
 Verbally describe past and present experience, and
expected consequences of future actions, of self or of other
agents
 Realize a certain activity given its verbal high-level
description
 Explain the meaning of notions
 Learn new notions and new languages
Consciousness is a
big suitcase
M. Minsky
A sketch of the evolutionary development of
cognitive abilities, consciousness included
P
h
e
n
o
m
e
n
a
l
F
u
n
c
t.
c
o
n
s
c.
c
o
n
s
c.
From: J. Wiedermann: A High Level Model of an Embodied Conscious Agent, IJSSCI, 2, 2010
A thinking machine: a de-embodied robot
cogitoid
Mirror
neurons
A robot’s thinking mechanism
in a computer
A brain in a vat
Lessons from what we have seen
• Achieving a higher-level artificial intelligence no longer seems to be a matter of
a fundamental scientific breakthrough but rather a matter of exploiting our best
algorithmic theories of thinking machines supported by our most advanced
robotic and real time data processing technologies.
• An artificial cognitive system is quite a complex system with only a few
components none of which could work alone and none of them could be developed
separately;
• It is unlikely that thinking machines will be developed by purely academic
research since it is beyond its power to concentrate the necessary amount of man
power and technology.
• This cannot be accomplished by large international research programs either
since a dedicated long-term open-ended effort of many researchers
concentrated on a single practically non-decomposable task is needed.
• It seems to be a unique strategic opportunity for giant IT corporations.
• The road towards thinking machines glimpses ahead of us and it only is a matter
of money whether we set off for a journey along this road.
Caspar David Friedrich, Giant Mountains, cca 1830

similar documents