Electronic Version of Poster Presented At Accelerating Change 2005
Link Back to AdaptiveAI Links Page
Link to Accelerating Change 2005
AGI Artificial General
Intelligence
The Return of Real AI
The nascent field of AGI
can best be seen as a re-birth of AI a return to the original goal of
artificial intelligence: the vision of building systems that learn and think
like humans computers that interact with, and understand the world.
There are several key
differences between AGI and current mainstream AI. Among the most important are
1. A focus on general intelligence (ie.
domain-independent abilities) as opposed to application specific knowledge and
skills.
2. The view that the ability to learn is
fundamental to real intelligence; acquisition of knowledge of skills is
much more important than simple possession.
3. Autonomous, real-time interactive agents
that both sense the world and act on it, versus manually fed batch systems and
specialized sub-modules (like expert systems or optimizers).
4. Ongoing goal-directed adaptation, compared to one-time
coding of algorithms, rules, and parameters.
General versus Domain-Specific Ability
General cognitive ability stands in sharp contrast to inherent specializations such as speech- or face-recognition, knowledge databases/ ontologies, expert systems, or search, regression or optimization algorithms. It allows an entity to acquire a virtually unlimited range of new specialized abilities. The mark of a generally intelligent system is not having a lot of knowledge and skills, but being able to acquire and improve them and to be able to appropriately apply them. Furthermore, knowledge must be acquired and stored in ways appropriate both to the nature of the data, and to the goals and tasks at hand ie. appropriate to its operating environments. These requirements imply that AGI systems operate interactively in real-time, and that they can learn autonomously.
General intelligence, as described above, demands a number of irreducible features and capabilities. In order to proactively accumulate knowledge from various changing environments, it requires:
1. Senses to obtain features from the world (virtual or actual),
2. Cognitive and meta-cognitive subsystems to learn and perform ongoing goal-directed learning and action,
3. A coherent, integrated means for storing knowledge and skills obtained this way, and
4. Adaptive output/ actuation mechanisms (both static and dynamic).
Conceptual Learning
versus Encoded Rules
AGIs usefulness and power
derives not only from the fact that they learn, but just as importantly,
that they learn conceptually and contextually. Conceptual
learning implies that knowledge is assimilated in a suitably generalized and
abstract form: Skills acquired for one task are available for similar, but
non-identical tasks; which makes these systems much more useful and robust when
coping with environmental changes. Context, on the other hand, allows the
system to utilize relevant background information to appropriately tailor
its responses to each specific situation. It can take into account such crucial
factors as recent actions and events, current goals and priorities, who it is
communicating with, and anything else that affects its current actions.
Traditional AI systems
obtain most of their abilities from initial encoding be it in the form of
programming, databases, decisions trees, or supervised training. AGIs, in
contrast, learn on an ongoing basis, and in variety of ways. These include
unsupervised and self-supervised learning, learning by doing (including from
their mistakes), and learning implicitly and explicitly from teachers.
Ongoing, interactive,
conceptual learning overcomes many of the problems of existing computer
technology, such as brittleness, lack of grounding/ understanding, and an
inability to automatically adjust to changing circumstances and requirements.
Other central AGI features include an ability to anticipate events and outcomes,
and the ability to introspect to be aware of its own cognitive states (such
as novelty, confusion, certainty, its level of ability, etc). These design
features, combined with the fact that AGIs directly perceive their environments
via built-in senses, endow them with human-like understanding of facts and
situations.
The Power of AGI
AGI technology promises to overcome current systems problems of brittleness, lack of grounded and contextual understanding, and their inability to adapt they will be truly intelligent. AGIs open-ended intelligence will not only match human levels, but will be further enhanced by computers natural strength. These include photographic memory, high-speed accuracy, upgradeability, seamless interfacing and communicating with other systems, the ability to be copied, etc.
The AGI approach inherently aims at having to solve the design, programming and engineering problems only once for a vast number of applications specific knowledge and skills will be acquired through learning and self-improvement. Users will not have scrap or re-design systems as their needs change.
Given its enormous potential, one may wonder why AGI isnt a
well-known, well-funded area of research and development.
Why such a Dearth of AGI Projects?
Several contributing factors seem to be accidents of history. Firstly, we now find ourselves in the depth of the AI winter a period of deep pessimism and lethargy towards AGI ambitions following the spectacular failure of early AI promises. In backlash to unfulfilled expectations of 30 and 40 years ago, Artificial Intelligence is still a swearword to many. Without delving into detailed analysis of these early failures, suffice it to say that hardware and software technologies and cognitive theories had simply not advanced sufficiently to enable the creation of human-level artificial intelligence.
However, while limitations of early technology were a definite handicap, several other theoretical and practical limitations, errors, and blind spots were and are even bigger impediments. These include the following:
Human-level AI is impossible At the most basic level, this is usually caused by remnants of an ancient philosophical position called Dualism. This long-since-discredited idea that there is an inherent dichotomy between mind and body leads many AI researchers to reject even the theoretical possibility of AGI. Thus they dont even try to solve the problem.
Not in my lifetime Of those who do not in principle object to the possibility of AGI, many do not believe that it can happen in their lifetime, if ever. Some hold this position because they themselves tried and failed in their youth. Others believe that AGI is not the best or fastest approach to achieving AI, or are at a total loss on how to go about it. One popular idea is that we need to reverse-engineer the human brain one function at a time in order to create intelligent machines.
There is no such thing as general intelligence - A great percentage of researchers reject the validity or importance of general intelligence. For many, controversies in psychology (such as those stoked by The Bell Curve) make this an unpopular, if not taboo subject. Others, conditioned by decades of domain-specific work, simply do not see the benefits of AGI of having intelligent systems with general learning ability.
We should not try to create AGI Several groups oppose AGI development on moral grounds, or because they fear it.
We dont know how to do it Many potential AGI entrepreneurs and researchers simply dont enter our field, because they lack crucial insights on how to achieve real artificial intelligence. There are many ways to be misdirected, and academia, if anything, hinders in that regard. To name just one of the most common errors entrenched in conventional AI thinking: the mistaken belief that intelligence is primarily about having knowledge. We see the ability to acquire knowledge i.e. to learn as far more fundamental.
Poor AI theory There are a many theories of artificial intelligence. Most of them will not lead to practical systems possessing general intelligence. Several theoretical errors and blind spots have already been mentioned here a few more common traps: The belief that AI can be solved by language alone, or conversely, that they require full embodiment (robotics); approaches that focus unduly on vision (or any other single aspect); overly abstract mathematical or philosophical theories that lack real-world grounding (universal Turing Machines, quantum consciousness and qualia); rigid rule-based designs, and statistical models that require infinite processing power.
Short-term academic and commercial pressure - Today, the bulk of AI research and development focuses on narrow applications that are quite domain specific. From a competitive point-of-view it doesnt really matter whether this results from a theoretical rejection of general intelligence, or simply from practical, short-term commercial or academic pressures; it is a lot quicker and cheaper to solve specific problems one at a time than to develop general learning. Of course, many are so focused on particular, narrow aspects of intelligence that they simply dont get around to looking at the big picture they leave it to others to make it happen. It is also important to note that there are often strong financial and institutional pressures to pursue specialized AI.
Loss of project focus The few projects that do pursue AGI based on relatively sound models run yet another risk: they can easily lose focus. Sometimes commercial considerations hijack a projects direction, while others get sidetracked by (relatively) irrelevant technical issues, such as trying to match an unrealistically high level of performance, fixating on biological feasibility of design, or attempting to implement high-level functions prematurely.
Peter Voss Adaptive A.I.Inc. September 2005
References -