| |
FAQ - Frequently Asked Questions
|
|
|
What is 'General Intelligence'?
General Intelligence refers to the basic skills needed to acquire a wide range
of domain-specific knowledge (data and skills) - i.e., to learn. More
specifically, this learning ability needs to be autonomous, goal-directed, and
adaptive:
-
Autonomous - Learning is both automatic, through exposure to sense
data (unsupervised), and through bi-directional interaction with the
environment, including exploration/ experimentation (self-supervised).
-
Goal-directed - Learning is directed (autonomously) towards
achieving varying and novel goals and subgoals - be they 'hard-wired',
externally specified, or self-generated. Goal-directedness also implies very
selective learning and data acquisition (from a massively data-rich, noisy,
complex environment).
-
Adaptive - Learning is cumulative, integrative, contextual and
adjusts to changing goals and environments
General Intelligence is fundamentally different from domain-specific
'intelligence' such as speech- or face-recognition, knowledge databases/
ontologies, expert systems, or search, regression or optimization algorithms.
Why is General
Intelligence so important?
General intelligence can ultimately learn any specialized knowledge and skills
possible - human intelligence is the proof! The opposite is obviously not true.
Once general artificial intelligence is developed, it can acquire an enormous
range of specific knowledge and skills - limited only by processing and storage
capacity. What is more, much of this learning will be autonomous - without a
teacher, and certainly without explicit programming. This is a crucial aspect
of Seed AI.
What is 'Seed AI'?
Seed AI is the concept of building an AI system with a limited, but carefully
chosen, set of capabilities that allow the system to dramatically increase its
knowledge and skills through self-directed learning and adaptation. One
concentrates on carefully designing the seed of intelligence, and then nurses
it to maturity - bootstrapping intelligence. In my AGI design this has two
distinct forms/ phases:
-
Coding the basic skills that allow to system to acquire a large amount of
specific knowledge
-
The system reaching sufficient intelligence, and conceptual understanding of
its own design, that it can deliberately improve its own design.
Will
achieving AGI not require a multi-billion dollar budget?
A key element of 'Real AI' theory (and its implementation) is to identify the
essentials of intelligence. Seed AI becomes a manageable problem - in some
respects much simpler than other mainstream AI goals - by eliminating huge
areas of difficult, but inessential AI complexity. Once we get the crucial
fundamental functionality working, much of the additional 'intelligence'
(ability) required is taught or learnt, not programmed. Having said this, I do
believe that very substantial resources will be required to scale up the system
to human level storage and processing capacity. However, the far more moderate
initial prototypes will serve as proof-of-concept for AGI, and provide a huge
number of practical new applications.
Why does the 'Real AI'
approach favored by Ben Goertzel,
Eliezer Yudkowsky, and myself find virtually no support in the
mainstream AI community?
The main reason, I believe, is that very few researchers have actually studied
the problem from our (the general intelligence/ seed AI) perspective - they
simply don't have a good feel for it. With no clear model of a developmental
path to human-level intelligence they are unable to see the merit of our
designs. A vision of how to get from 'here' to 'there' is crucial for lending
credibility to our work.
*Why are so few
researchers pursuing Real AI?
Basic division -
More specific reasons for not focusing on 'Real AI' -
-
They don't believe that 'general intelligence' is a valid
concept
-
They don't believe that AGI can be achieved within their lifetime
- the time is not ripe!
-
They don't believe that 'general intelligence' is the best
approach to achieving 'AI'
-
They don't see why it's so important - don't consider Seed AI
benefits
-
They don't know how to do it (no model) - intimidated!
-
They are trying to reverse engineer the brain - one function at a
time
-
They are focusing all of their attention on one (or a few) aspects
of intelligence - not the whole picture
-
They tried in their youth (15 - 40 years ago) and failed - now,
'conclude' that it can't be done
-
They can get quicker results (financial and other) pursuing
specialized AI
-
They get little academic respect/ support/ funding
-
They are afraid of it
All of the above combine to create a dynamic where Real AI is not 'fashionable',
further reducing the number of people drawn into it!
Why is there so
little progress in (workable) AGI models and systems?
-
See: 'Why are so few researchers pursuing Real AI?'
-
The importance of general intelligence is not appreciated
-
Poor epistemology - a proper understanding of
grounding, focus and selection, concepts, context, etc. is needed to appreciate
embedding, vector encoding, activation spreading, multi-sense, etc.
-
Poor understanding of key concepts - getting stuck on:
consciousness, volition, meaning, representation/ world-model, emotions,
common-sense, 'qualia
-
Poor understanding of 'intelligence': Knowledge (CYC)
versus two-way interactive, adaptive learning. Also, missing the importance of
abstract cognition
-
Not appreciating the central importance of patterns -
especially dynamic ones (entities, attributes, concepts, actions, thoughts,
etc.)
-
Too little emphasis on adaptive output/ actuation
capability
-
Too much focus on copying/ reverse-engineering the brain -
biological feasibility
-
Too much focus on high-level abilities (logic, language,
creativity, etc.)
-
Performance expectations too high for any specific
functionality (vision, speech, etc.)
-
Schism between traditionalists and connectionists - False
dichotomy: symbol/ schema Vs pattern/ incomprehensible
-
Undue focus on evolutionary and agent systems (Society of
Mind)
-
Getting stuck on custom (NN and/ or robotics) hardware in
the early stages (CAM, Cog)
-
Over-estimate of hardware and software needed (Seed AI,
minimal AGI, limited motor-sensory, NN without paying NN price)
Under-researched areas: Incremental, real-time,
unsupervised/ self-supervised learning (Vs backprop!) - Self-tuning: bottom up
and top-down (data and emotions/ goal/ meta-cognitive driven) - Dynamic NN
topologies - Dynamical, interactive, adaptive AGI systems - Combining the best
from NN, traditional AI, fuzzy, etc.
What is
different in your AGI approach?
1) The focus is more on: selective, dynamic, incremental, and interactive -
-
Importance of selective inputs/ senses
-
Dynamic patterns (temporal/ sequences) an integral part of design
-
Importance of adaptive actuation (output) circuitry for interacting with
environment, and for modifying internal states.
-
Data that is fuzzy (overlapping), noisy, complex, analog
-
Integrated entity/ attribute/ category data network (fuzzy entities)
-
Importance of concepts (with and without labels)
-
Importance of context (relationships and activation spreading)
-
Unsupervised learning (data driven)
-
Interaction with overwhelmingly complex, dynamic environment
-
Self-supervised learning (adaptive output patterns provide feedback loop)
-
Implement functionality as NN data structures, not coding - Self-programming,
less brittle
2) Concentrate less on: knowledge (data), high-level/ specialized skills,
complex senses and locomotion -
-
Knowledge (should be learnt, not coded)
-
Logic (should be learnt, not coded)
-
Language (should be learnt, not coded)
-
Schemas (too brittle)
-
Batch optimization (learning should be incremental, dynamic, and interactive)
-
XML (too hierarchical - not a network - not fine grained - not unlabeled,
data-driven structure)
-
Evolutionary programming (too hit and miss - insufficient data/ trials -
unnecessary)
-
Agent programming (too fluid - too much overhead - unnecessary)
-
Reverse engineering (or copying) the brain - aiming for biological plausibility
-
Complex vision, speech recognition and production, locomotion
What is your working definition
of the following terms: Mind, Intelligence, Consciousness?
Mind - The totality of all activity
and data of a functioning brain.
Intelligence - The adaptive ability to
form and achieve varying and novel \ goals in a complex, dynamic environment.
Consciousness - This is a 'suitcase'
word; it has many different meanings. The most consistent meaning is
'awareness' (i.e. 'has knowledge of'). In the context of AI this covers a range
of "aware of features/ properties of the environment", to "aware of its own
'body'", to "aware of its own mental processes".
How does your AGI
represent/ encode data and skills and what are the learning mechanisms?
All data is stored in one common, integrated
neural-nettish structure that has a dynamic, adaptive topology (Neural Gas
like)
There are various learning mechanisms - though all
major systems are real-time and incremental. Learning is both bottom-up (data
driven) and top-down (goal driven, emotion, and meta-cognition). Learning modes
are unsupervised, self-supervised, supervised (plus aping, coaching, coaxing).
(Further details: discussion only)
Minimizing the
AGI task - Maximizing results
-
Eliminate function known to be nonessential: complex
mobility, hearing/ vision/ speech (Helen Keller)
-
Concentrate on the functionality of a crippled baby
animal: primarily the ability to learn adaptively.
-
Eliminate function that can be bootstrapped
(learnt/taught): knowledge, structured language, logic, etc.
-
-
Use existing technology and research when possible
-
Be prepared to implement/try poorly understood designs
-
Use 'dirty tricks': be prepared to use
non-traditional, non-biological-feasible algorithms
-
Try to bootstrap/self-design/self-tune parameters and
functionality (feedback loops)
-
Selection of early products: capitalize on prototype
strength, minimize effect of limitations.
Practical issues -
How do you get from a
basic perceptual/ animal 'AGI' to human-level intelligence?
High-level functionality is elaboration/ adaptation of
basic animal-level pattern processing capability.
(Further details: discussion only)
-
More effective control of 'emotions'
-
Restore from backup
-
Freeze knowledge with high degree of certainty
-
Unlearn on demand
-
Reverse activation
-
Nodes and links can easily be tagged (labeled) and
categorized.
-
Node, link, and internal parameter data is available
as 'input data' (full introspection)
-
Enforce logical thinking (when appropriate)
-
Rapid loading, copying, and merging of knowledge
(also: one-shot learning and photographic memory)
-
Simple interface with other databases and machines
-
Design information available (to designer and itself!)
-
Patterns, Thinking, and Cognition -
Howard Margolis
-
The Engine of Reason, Seat of the
Soul - Paul Churchland
-
Affective Computing - Rosalind
Picard
-
AGIRI:
(especially: 'Webmind - A True Digital Mind in the Making')
|
|
|
|
|
|
|
