Transcript Document

Achieving Advanced Machine
Consciousness via Artificial General
Intelligence in Virtual Worlds
Ben Goertzel, PhD
Contents
1. The Nature of Consciousness
2. Artificial General Intelligence versus Narrow AI
3. The Novamente and OpenCog AGI Projects
4. The Marriage of AGI and Virtual Worlds
5. Initial Application: Virtual Pet Brain
A Useful Philosophical Perspective
On Consciousness
In
Metaphysical Foundation:
Peircean/Jungian Categories
First: raw, unprocessed being … e.g. qualia
Second: reaction … e.g. pure physical reaction
Third: relationship
(beyond Peirce … “Fourth: synergy”, etc.)
Archetypal Perspectives
First person: firstness of X … the world as directly
experienced … the stream of qualia …
Third person: thirdness of X … the world as an
objective relational structure, a network of patterns
Fourth person (normally called “second person”):
fourthness of X … the synergy of relationships … the
Buber-ian I-You
The real second person: secondness of X …
experiencing the world as an automaton?
Inter-perspective correlations
Example of a hypothesis spanning perspectives:
The more intense qualia experienced by a system,
correspond to the more informationally significant
patterns detectable in that system by an intelligent,
well-informed observer
Reflective consciousness
and other emergent constructs
Hypothesis:
Among the more informationally significant patterns in
generally intelligent systems are:
•The phenomenal self
•Reflective consciousness
•The illusion of will
Modeling Reflective Consciousness,
Self and Will Using Hypersets
Hypothesis:
The qualia we humans describe as “reflective
awareness”, “self” and “will” correspond to patterns in
our brains that are conveniently expressible in terms of
hypersets (non-well-founded sets)
Modeling Reflective Consciousness,
Self and Will
Using Hypersets
“S is conscious of X" is defined as: The declarative
content that {"S is conscious of X" correlates with "X is a
pattern in S"}, where S is an intelligent system’s
phenomenal self
"S wills X" is defined as: The declarative content that
{"S wills X" causally implies "S does X”}, where S is an
intelligent system’s phenomenal self
"X is part of S's self" is defined as: The declarative
content that {"X is a part of S's self" correlates with "X is
a persistent pattern in S over time"}
Evaluating Hypersets as Patterns
in Dynamical Systems
The hyperset defined by X = F(X) may be evaluated as a
pattern in a system by comparing the iterates
A
F(A)
F(F(A))
…
to the system’s trajectory at various times for various A
Summary
•There are multiple archetypal perspectives: First,
Second, Third, Fourth person,…
•There are correlations between the different
perspectives (e.g. intense qualia correspond to
informational patterns)
•There are specific emergent structures (self, will,
reflection) that correlate with intense patterns/qualia in
generally intelligent systems
•It may be interesting to model these emergent structures
using hypersets
Artificial General Intelligence
versus Narrow AI
In
Artificial General Intelligence (AGI)
“The ability to achieve complex goals in complex
environments using limited computational resources”
• Autonomy
• Practical understanding of self and others
• Understanding “what the problem is” as opposed
to just solving problems posed explicitly by
programmers
•Solving problems that were not known to the
programmers
Narrow AI
The vast majority of AI research practiced in academia
and industry today fits into the “Narrow AI” category
Each “Narrow AI” program is (in the ideal case) highly
competent at carrying out certain complex goals in
certain environments
• Chess-playing, medical diagnosis, car-driving, etc.
Today, Narrow AI Dominates the AI Field
(in both academia and applications)
Deep Blue - whoops us pesky humans at
chess - but can’t learn to play a new game
based on a description of the game rules
DARPA Grand Challenge - a great leap
forward -- but it can’t learn to drive different
types of vehicles besides cars (trucks,
boats, motorcycles)
Google - fantastic service: but can’t
answer complex questions. Whatever
happened to AskJeeves?
2001
Artificial General Intelligence (AGI)
Hypothesis: Human-level general intelligence
naturally comes along with the emergence of
• Phenomenal self
• Reflective consciousness
• Illusion of free will
A Pragmatic, Integrative
Approach to Advanced AGI
In
Novamente Cognition Engine
The Novamente Cognition Engine (NCE) represents a serious
scientific/engineering effort to create powerful artificial general
intelligence, via an integrative, computer science based approach
While the NCE may be applied in many different domains, the most
natural way to develop and apply it, at the current stage, is in the
context of controlling physically and/or virtually embodied intelligent
agents
For more detail on the NCE, see novamente.net/papers
Open C ogni ti on Framew ork
The OpenCog project (opencog.org) is an open-source
offshoot of the Novamente project, which has been seeded in
2008 with significant AGI code donated by Novamente LLC
It includes the RelEx NL comprehension system, founded on
the CMU link parser plus additional rule-based and statistical
NLP methods
The essential dynamics of these AGI systems follows the
basic logic of animal behavior:
Enact a procedure so that
Context & Procedure ==> Goals
i.e.
at each moment, based on its observations and memories, the system
chooses to enact procedures that it estimates (based on the properties of
the current context) will enable it to achieve its goals, over the time-scales
these goals refer to
There is an important distinction between explicit goals and
implicit goals
 Explicit goals: the objective-functions the system explicitly chooses
actions in order to maximize
 Implicit goals: the objective-functions the system actually does
habitually maximize, in practice
For a system that is both rational, and capable with respect to its goals in
its environment, these will be basically the same. But in many real
cases, they may be radically different
Goal Dynamics

A sufficiently intelligent system is continually creating new
subgoals of its current goals

Some intelligent systems may be able to replace their toplevel supergoals with new ones, based on various dynamics

Goals may operate on radically different time-scales

Humans habitually experience “subgoal alienation” -- what
was once a subgoal of some other goal, becomes a top-level
goal in itself. AI’s need not be so prone to this phenomenon
Five key aspects of AGI design:
1.
2.
3.
4.
5.
Knowledge Representation
Cognitive Architecture
Knowledge Creation
Environment / Education (incl. physical &
virtual robotics)
Emergent Structures and Dynamics
There is no single, mechanism-level “magic trick” at the heart of general intelligence
… rather, intelligence arises in appropriately-constructed complex systems as an
emergent phenomenon.
The trick is to figure out what sorts of complex systems will give rise to general
intelligence as an emergent property.
There is unlikely to be “one correct answer” to this question … but all we need to build
the first thinking machine is one of the many correct answers.
The Novamente/OpenCog high-level cognitive architecture is based on the state of the art
in cognitive psychology and cognitive neuroscience. Most cognitive functions are
distributed across the whole system, yet principally guided by some particular module.
Unique hypergraph knowledge representation bridges the gap between
subsymbolic (neural net) and symbolic (logic / semantic net) representations,
achieving the advantages of both, and synergies resulting from their
combination.
Each cognitive processing machine, within each unit, contains an “Atom
Space” full of nodes and links representing knowledge, plus a set of cognitive
processes acting on this Atom Space, encapsulated in software objects called
MindAgents and scheduled by a Scheduler object.
Each box in the cognitive architecture diagram, corresponds at the software level to
a cluster of machines called a “unit”, containing a local persistent DB plus one or
more cognitive processing machines.
Algorithms for Procedural and Declarative
Knowledge Creation
MOSES Probabilistic
Evolutionary Learning
(for gaining procedural knowledge directly)
Combines the power of two leading AI
paradigms: evolutionary and probabilistic
learning
Extremely broad applicability. Successful
track record in bioinformatics, text and data
mining, and virtual agent control.
Probabilistic Logic Networks
(for gaining declarative knowledge directly)
The first general, practical integration of
probability theory and symbolic logic.
Extremely broad applicability. Successful track
record in bio text mining, virtual agent control.
Based on mathematics described in
Probabilistic Logic Networks, published by
Springer in 2008
Economic Attention Allocation
Each node or link in the knowledge network is tagged with a probabilistic truth
value, and also with an “attention value”, containing Short-Term Importance and
Long-Term Importance components.
An artificial-economics-based process is used to update these attention values
dynamically -- a complex, adaptive nonlinear process.
The system contains multiple heuristics for Atom creation, including
“blending” of existing Atoms
Atoms associated in a dynamic “map” may be grouped to form new
Atoms: the Atomspace hence explicitly representing patterns in itself
Hypothesis: Integrative Design
Can Allow Multiple AI Algorithms to Quell
Each Others’ Combinatorial Explosions
Pattern Mining
Probabilistic Evolutionary
Program Learning
Economic Attention
Allocation
Probabilistic
Logical Inference
Overall Philosophy
Algorithms for declarative and procedural knowledge
creation and attention allocation …
integrated with appropriate synergy and acting on an
appropriately powerful knoweldge representation …
used to control a system pursuing complex goals …
may lead to the emergence of system structures
characteristic of general intelligence
Why Do I Believe I Can Succeed When So Many Others
Have Failed?
•
Approach is based on a well-reasoned, comprehensive theory of
mind, which dictates a unified approach to the five key aspects
mentioned above






Knowledge representation
Learning/reasoning
Cognitive architecture
Embodiment / interaction
Emergent structures / dynamics
Cognitive Theory summarized in The Hidden Pattern (Ben
Goertzel, Brown Walker Press, 2006)
•
The specific algorithms and data structures chosen to implement this
theory of mind are efficient, robust and scalable and, so is the
software implementation
The Marriage of AGI
and Virtual Worlds
In
How Important Is Embodiment?

Some AI theorists believe that robotic embodiment is necessary for
the achievement of powerful AGI

Others believe embodiment is entirely unnecessary

We believe embodiment is extremely convenient for AGI though
perhaps not strictly necessary; and that virtual-world embodiment
is an important, pragmatic and scalable approach to pursue
alongside physical-robot embodiment
Public virtual worlds provide a wonderful opportunity for teaching baby
AI’s: not only the experience of embodiment, but the massive plus
of having hundreds of thousands or millions of teachers helping the
AI to learn
Current virtual world platforms have some fairly severe
limitations, which fortunately are fairly easily remedied
Object-object
interactions are
oversimplified, making
tool use difficult
Agent control relies on animations
and other simplified mechanisms,
rather than having virtual
servomotors associated with each
joint of an agent’s skeleton
Example solution: Integration of a
robot simulator with a virtual world
engine
+
Player / Gazebo: 3D
robot control + simulation
framework
RealXTend/OpenSim:
open-source virtual world
It seems feasible to replace OpenSim’s physics engine with appropriate components
of Player/Gazebo, and make coordinated OpenSim client modifications
Cognitive Control of agents in current virtual worlds -- e.g.
Second Life, Multiverse, HiPiHi
e.g. “take one step forward”
non-parametrized
behavior signals
Cognition
Engine
high-level perceptual
data
Coordinates of objects,
Labeled with type
Hybrid Generally-Intelligent Robot Brain Architecture, version 1
e.g. ”Force F exerted by
servomotor M in direction D”
action signals
e.g. “take one step forward,
using gait parameter vector V”
Behavioral
postprocessor
behavior signals
Behavioral
modules
Neural net
module evolver
Object
classification
modules
raw perceptions
e.g. video output of
camera eyes
Perceptual
preprocessor
mid-level perceptual
data
e.g. 3D polygonal mesh,
marked up with limited object
Identification information
Cognition
Engine
Application:
Novamente Pet Brain
In
Novamente Pet Brain
The Pet Brain utilizes a specialized version
of the Novamente Cognition Engine to
provide unprecedentedly intelligent virtual
pets with individual personalities, and the
ability to learn spontaneously and through
training.
Pets understand simple English; and future
versions to include language generation
The Pet Brain incorporates MOSES learning to
allow pets to learn tricks, and Probabilistic
Logic Networks (PLN) inference regulates
emotion-behavior interactions, and allows
generalization based on experience.
Demo Screenshots: Training
Novamente-powered smart pets can be taught to do simple or complex
tricks - from sitting to playing soccer or learning a dance - by learning from
a combination of encouragement, reinforcement and demonstration.
Teach
give “sit” command…
Imitate
show example…
Reinforce
Correct
successful sit, great… reinforce and/or correct.
Teaching with a Partner
In partner-based teaching, the pet understands that one avatar is the teacher
and the other is the student, whose interactions with the teacher the pet is
supposed to understand, abstract, and imitate
Next Step: Language Learning
Our initial virtual pets have robust but
simplistic language understanding,
sufficient to learn an unlimited variety of
commands
In the next version, integration of
Novamente’s RelEx language processing
system with the Novamente Pet Brain will
provide a more powerful approach to
embodied language learning
With human-controlled avatars as
language teachers, Novamente-controlled
virtual agents will be able to rapidly
improve their language comprehension
and generation via adaptive learning
Next-Gen Pet/Baby Brain Architecture
The next generation of the Avatar Brain will incorporate additional modules
allowing language processing and more advanced inference -- the next step on
the path from virtual dogs to human-level virtually-embodied AGIs
Stages of Development of an AGI
Full Self-Modification
Reflexive
Deep understanding and control of self structures
and dynamics
Formal
Abstract reasoning and hypothesizing. Objective
detachment from phenomenal self.
Concrete
Rich variety of learned mental representations and
operations thereon. Emergence of phenomenal self.
Infantile
Making sense of and achieving simple goals in
sensorimotor reality. No self yet.
Intelligence…
Intelligence…
Intelligence…
“Within thirty years, we will have the
technological means to create superhuman
intelligence. Shortly thereafter, the human era
will be ended…
When greater-than-human intelligence
drives progress, that progress will be much
more rapid”
The Coming Technological Singularity
Verner Vinge (1993)
“I set the date for the Singularityrepresenting a profound and disruptive
transformation in human capability- as 2045.
The nonbiological intelligence created in that
year will be one billion times more
powerful than all human intelligence
today."
The Singularity is Near,
When Humans Transcend
Biology - Ray Kurzweil (2005)