The document summarizes a presentation on artificial general intelligence (AGI) given at the IntelliFest 2012 conference. It discusses the limitations of narrow AI and the constructivist approach needed for AGI. This involves self-constructing systems that can learn new tasks and adapt. The presentation highlights the HUMANOBS project, which uses a new architecture and programming language called Replicode to develop humanoid robots that can learn social skills through observation. Attention and temporal grounding are also identified as important issues for developing practical AGI systems.
2. Helgi Páll Helgason
helgi@perseptio.com
In collaboration with: Dr. Kristinn R. Thórisson & Eric Nivel
Intellifest 2012
AI researcher, Ph.D. candidate
Center for Analysis and Design
of Intelligent Agents,
Reykjavik University
Founder / CEO,
Perseptio
4. Displaying human-like behavior?
Solving computationally complex problems (in
some unspecified amount of time)?
Performing isolated tasks that have
conventionally required humans?
Adapting to a complex, dynamic environment
with insufficient knowledge and resources?
Intellifest 2012
5. Intellifest 2012
Nature’s way of dealing with complexity
under resource and time constraints
No real-world intelligence exists that does
not address the passage of time head on
6. Intellifest 2012
“Intelligence is the capacity of a system
to adapt to its environment while
operating with insufficient
knowledge and resources.”
- Pei Wang
(Rigid flexibility: The Logic of Intelligence. Springer 2006)
7. Intellifest 2012
“If either time or computational
resources are infinite, intelligence is
irrelevant .”
- Dr. Kristinn R. Thórisson
10. “Narrow” (classical) AI:
• Systems explicitly designed to solve
specific, reasonably well-defined problems
E.g. Deep Blue, Watson, etc.
Artificial General Intelligence:
• Systems designed to autonomously learn new
tasks and adapt to changing environments
Intellifest 2012
11. Neural-network based virus detection system
Personalized movie recommendation system
Software for prediction, classification and
pattern recognition
NLP-based personalized news delivery
software
Development and implementation of
algorithmic trading strategies using AI
techniques
Intellifest 2012
12. Constitutes vast majority of work in field of AI to-date
While proven useful in industry beyond doubt, by
definition unlikely to lead to human-level intelligence
• Highly unlikely that isolated bits and pieces of the “intelligence
puzzle” can somehow be fused into holistic intelligence
Difficult to generalize for new problems
Task environments are typically limited or simplified
representations of the real world
Substantial adaption to tasks not required
Intellifest 2012
13. AGI (Artificial General Intelligence):
• Relatively small group of researchers (so far)
refocusing on an original idea behind AI:
human-level AI
• First AGI conference: 2008
• Systems explicitly designed to autonomously learn
novel tasks and adapt to changing environments
• Ultimately targets human-level intelligence (and
beyond) in real world environments
Intellifest 2012
14. AGI (Artificial General Intelligence):
• Long-term research effort
• Challenging funding situation
• Results are not guaranteed
Intellifest 2012
18. Benefits
• Versatile, highly adaptive and autonomous systems
• Reduced design/development cost due to high reusability
Downsides
• Predictability and determinism are sacrificed to some
degree
• Require learning phase before coming practically useful
Intellifest 2012
19. Intellifest 2012
Software systems approaching human-level
intelligence will be massive and complex
• Significantly greater complexity than exists in current
software systems
• Cognitive limitations of designers/programmers
Realistic to believe we will build such systems
using current software methodologies?
• Manual construction
• Coarse grained modular systems
• Divide-and-conquer
20. Intellifest 2012
Con - struct - ionist A.I.: Manually-built artificial
intelligence systems; learning restricted to combining
predefined situations and tasks, from detailed
specifications provided by a human programmer.
“The programmer as construction worker”
21. Intellifest 2012
Underlying assumptions:
• Systems of reasonable intelligence can be built
with an architecture of a few thousand manually
constructed modules
• Such a system could automatically:
Tune parameters as needed
Route information and control among the modules
22. Intellifest 2012
Reality:
• Intelligent systems are (functionally)
Heterogeneous
Large
Densely-coupled
Self-reflective
• Intelligence is the product of the operation of a
system
23. Intellifest 2012
Reality:
• Massive, complex dependencies exclude:
manual construction of modules
modular construction
piecewise composition where each piece built in isolation
Exceedingly large functional state-space
Subdivision hides important interconnections
• Divide-and-conquer fails
26. Intellifest 2012
Available evidence strongly indicates that the
power of general intelligence, arising from a
high degree of architectural plasticity, is of a
complexity well beyond the maximum reach
of traditional software methodologies.
28. Intellifest 2012
Con - struct - ivist A.I.: Self-constructive artificial
intelligence systems with general knowledge
acquisition skills; systems develop from a seed
specification; capable of learning to perceive and act
in a wide range of novel tasks, situations and
domains.
Thórisson, K. R. (2009). From Constructionist to Constructivist A.I. Keynote, AAAI Fall
Symposium Series: Biologically Inspired Cognitive Architectures, Washington D.C., Nov. 5-
7, 175-183. AAAI Tech Report FS-09-01, AAAI press, Menlo Park, CA.
29. Intellifest 2012
Developmental approach
Targets a ratio of hand-coded to auto-
generated programs of magnitude 1:1000000
and up
Requires:
• Small building blocks (peewee-size)
Facilitates transfer and re-use between tasks and
domains
Fast, (temporally) predictable execution
30. Intellifest 2012
Standard programming languages
• Designed for humans
• Complex operational semantics
• Not suited for automatic self-programming
• No (explicit) temporal grounding
Constructivist AI needs a new paradigm:
• Transparent operational semantics
• Machine-understandable language
• Explicit temporal grounding
• Self-organizing management
• Uniform representation
31. Intellifest 2012
Fundamental to AGI systems
• All cognitive actions take time, impacting the system’s
place in the context of the real world
Resource management
• Constant operating scenario: abundant information,
limited resources
• Range of time-constraints
Many posed by the environment
Temporal dimension of knowledge is
important
• Past events
• (Expected) future events
32. Intellifest 2012
System must be able to
• Predict the effects and side-effects of its actions in
the world
• Predict the effects and side-effects of its own
internal operations
Requires uniform representation
• That includes time at atomic operation level
33. Intellifest 2012
Humanoids that Learn Socio-Comunnicative Skills Through Observation
Funded by European Union 7th Framework Programme
Coordinator / Principal Investigator: Dr. Kristinn R. Thórisson
34. Intellifest 2012
Target domain: TV-style interview
Two roles: Interviewee, Interviewer
Scenario:
• Two humanoid avatars and props in a virtual 3D
environment
• Multiple modalities involved in perception and
control of an avatar
Speech, intonation, gaze, head movements, hand
movements
35. Intellifest 2012
Developmental operation:
• 1. Motor babbling phase
System tries out its actuators and builds models of how
it can impact the environment
• 2. Observation phase:
Humans control both avatars and perform interview
while system observes
• 3. Operation phase:
System takes over one of the roles
36. Intellifest 2012
New AGI architecture developed for this
project:
Autocatalytic
Endogenous
Reflective
Architecture
37. Intellifest 2012
Broad-scope, general-purpose architecture
addressing:
• Perception, decision, motor control
• General-purpose action learning in dynamic worlds
Tiny construction components
• Relative to size of architecture
Support of transversal cognitive skills, at multiple
levels of granularity and abstraction
• System-wide learning
• Temporal grounding
• Observation and imitation of complex realtime events
• Inference, abstraction, prediction, simulation ... and more
40. Intellifest 2012
Replicode: New programming language
• Developed specifically for HUMANOBS/AERA
Open source
• Designed to build model-based, model-driven
production systems that can modify their own code
Containing a (very) large number of
concurrent, interacting programs
41. Intellifest 2012
New programming language
• Encodes short parallel programs and executable models
• Explicit temporal grounding
Soft realtime
• Data-driven execution model
Computation based on pattern matching
• No explicit conditional statements (if-then) or loops
• All executable code runs concurrently
42. Intellifest 2012
New programming language
• Code can be active or inactive
• Code can be input for some other code
• Dynamic code production
• Execution feedback
• Supports distribution of computation and knowledge across
clusters of computing nodes
43. Intellifest 2012
[FACT@T0: BOX AT POS (0,0)]
[CMD@T0: MOVE BOX (1, 0)]
[PRED@T1: BOX AT POS (1,0)]
[GOAL@T1: BOX AT POS (0,0)]
[FACT@T0: BOX AT POS (1,0)]
[CMD@T1: MOVE BOX (-1,0)]
PREDICTION
Model
PRESCRIBE ACTION
47. Intellifest 2012
In the domain of (generally) intelligent
systems, the management of system
resources is typically called “attention”
Critical (and neglected) issue for AGI
• Systems constantly working with limited resources
under time constraints in environments providing
abundant information
48. Intellifest 2012
Design of AGI systems needs to address
practical limitations from the outset
AGI systems will face time-constraints and
need to be reactive and interruptible, yet
capable of planning
Retrofitting AGI systems with resource
management highly challenging
• Duration of atomic operations becomes important
49. Intellifest 2012
Control mechanism responsible for prioritizing data
and processess
Targets equally
• External information (from the environment)
• Internal information (from within the system)
General, no assumptions about
• Tasks
• Environments
• Modalities / Embodiment
Adaptive
• Learns to improve itself based on experience
51. Intellifest 2012
Potential:
• Knowledge sharing
Systems learning not just from their own experience, but
from the experience of other identical (or similar) systems
On-demand access to knowledge bases & services
• Distributed resources
Systems using computational resources they do not
physically contain
• Remote on-demand sensing
52. Intellifest 2012
Possible limitations:
• Communication latency
Operations involving network communication may introduce
time delays, which may be significant in terms of operation
• Communication bandwidth
Sensory information can be a significant amount of data (100
MB+/sec)
• Cloud server load
Response times for cloud-based operations less predictable
53. Intellifest 2012
Interesting directions:
• Augment system knowledge from cloud during idle
time
• Compress and/or pre-process sensory information and
run cognitive processes in the cloud
Latency still an issue
• Cloud-based services used as specialized tools
When allowed for by temporal constraints
54. Intellifest 2012
Practically difficult:
• Distributing cognitive processes between system’s
own hardware and the cloud
AGI systems likely to have a very large number of
components with rich, complex interconnections and
interactions
Communication latency becomes a major issue