Human Intelligence Models
In the past, there have been several attempts to define a benchmark against which artificial intelligence could be judged, e.g. the Turing test. In retrospect, this often leads to proclaimed AI systems that only mimicked intelligence rather than replicate it. While no inference is being made that only the human brain model can lead to true intelligence, it is a model that appears, from our perspective, to work and has stood the test of evolution. Therefore, it is seen as a positive development that new lines of AI research have been expanded to include research into human or biological intelligence, even though this approach is also turning out to be far more complex than originally perceived. However, the study of human intelligence has already highlighted, what might prove to be, a number of conceptual errors in some of the original AI models:
- Change Models:
Humans appear not to build complete models of the world they see. Experiments have shown that when the environment around us is changed, we only notice the more drastic changes. As such, we do not keep all the information in our head, but instead refer back to the environment for more detail. Equally, humans tend to only represent what is immediately relevant from their environment, and somewhat surprisingly some of these representations are even independent of one another.
- Distributed Control:
In the early days of computing, most systems were based on a central processing unit and initially the brain was seen as the equivalent central controller. However, the human brain is actually comprised of distributed control systems, rather than a single central one. Again, this conclusion has been derived from numerous experiments on subjects suffering from a range of brain disorders or abnormalities.
- General Purpose Processing:
While humans are adept at solving a range of different problems, our brains are not general-purpose machines. Again, experiments have demonstrated the specialised nature of human computational processes that evolved to solve the problem of survival. Further more, humans are often more emotional than rational, and there is evidence that our emotions are an important aspect of our decision making.
In order to simplify the building of intelligent systems, one initial AI approach was to simply ignore or avoid many of the more complex aspects of human intelligence. However, these elements are undoubtedly essential to human intelligence and so considerable research is now going into understanding their role in AI:
- Learning and development:
Humans are not born complete and our reasoning, motor and sensory systems all have to undergo a considerable learning process, as we develop towards adulthood. Development involves incrementally mastering ever more difficult tasks in ever more complex environments. So development is an incremental process with learned skills laying the foundations for the further acquisition of more advanced skills. As such, newborn infants do not have full control over their limbs, but through a gradual process of learning, their motor control develops to allow extremely complex movements of their bodies to be coordinated. A process, in which the control of both the sensory and motor systems is gradually increased, is believed to significantly reduce the difficulty of learning such complexity in one step. In turn, this process increases our ability to solve more complex problems. In effect, humans are adaptive, self-learning systems rather than pre-programmed mechanisms
- Social interaction:
Human babies enter the world with their brains still far from fully developed for various physiological reasons. Of course, while the human brain is capable of incredible learning and development after its birth, this takes time, during which the infant is extremely vulnerable. As a consequence, humans have evolved complex social environments, not only to protect their young, but also to provide for their long-term intellectual development. However, without the ability to interact socially, an infant may be severely limited in the range of new skills it can learn from those around it. Many of the basic physical coordination skills are learnt from the social environment through play, involving both mimicry and imitation. The quality of play is often dependent on the level of social interaction with other children. However, the mastering of a language skill (vocal and visual) is the critical development to deeper social interaction and knowledge acquisition skills. Initially, building social skills into AI may start off being little more than defining some mechanical rules of etiquette. However, the issue of the development of more advanced social interaction will be more complex and involve the development and association of AI emotions to learnt experiences. It is said that children reach half their IQ levels by the age of four; however an adult entering a pre-school playgroup is not struck by structured social interactions and learning, but rather something closer to chaos. In emotional terms, this chaos is actually children being motivated to learn through fun. As humans develop, other emotions provide additional motivation to learn and develop new skills.
- Physical interaction:
Although it was not directly addressed, social interaction also requires physical interaction. However, this aspect of the development of human intelligence is often overlooked. The physical embodiment of the human brain provides the ability to move and sense its surroundings. This ability appears to be directly linked to the way the human brain then comprehends its surroundings. For example, the human brain does not form a complete picture of it surroundings, but rather continuously monitors for important changes through its physical senses. This approach to real-time data processing is very efficient and is emulated in many computer systems. Differential systems are based on building larger models from small pieces of incremental data. Equally, most computer systems have a hierarchy of interrupt signals with each signal having a priority with respect to another. Low-priority actions can either be deferred or handed off to sub-system processors, while high-priority signals demand immediate attention by the central management system.
- Sensory Integration:
Humans receive an enormous amount of information from the world via their senses. These senses are processed to provide us with a view of the surrounding world. It may not be too surprising that many of the sensory inputs are integrated together to provide a more coherent perception of our surrounding reality. Sensory integration can also simplify the computation of any given task. Attempting to perform the task using only one sense may be either awkward or impossible and invariably requires more concentration, i.e. processing power. As in the case of differential processes, the integration of smaller bits of individual sensory information can allow a more complex perception model to be efficiently created. This multi-sensory model can also facilitate new types of associative learning that may otherwise not occur. For example, objects that make noise often move, the speed in which the brain can make this perceptual link could mean the difference between life and death.
If these are some of the central elements of human intelligence it is probably a reasonable assumption to say that the solution is not confined to computer science. Therefore, the corollary to this argument is that if human intelligence goes beyond computer science, AI may also have to go beyond computers in their current form. However, the idea to create artificial life has been around for centuries, but only with the development of the modern computer, in the 1950’s, did this become a tangible possibility. In fact, it appears that people started to extrapolate their ideas about the possibility of strong AI, even before weak AI. With hindsight, this initial exuberance has been overtaken by pragmatism, as the real difficulty of strong and weak AI has become better understood. Human intelligence has both physical and meta-physical structures, but at this time, we barely have little more than a rudimentary understanding of the former, and the latter is still a matter of philosophical debate. There is absolutely no evidence that artificial intelligence is going to be any simpler than biological intelligence, unless it is inferior. Of course, there is every possibility that AI will initially be an inferior intelligence, which is something to be considered, as we hand-over more and more control of our society’s critical systems to AI.
We now know that the structure of human intelligence is a result
of a very lengthy process of evolution. In addition, the process of
brain development continues after birth with neural pathways adapting
to assimilate new knowledge throughout a person’s lifetime. The challenge
facing AI research is what aspects of the human model are essential
and can any of it be reproduced and ultimately improved. History suggests
that some of the early pioneers in AI were too optimistic about future
progress and the blunt truth is that despite over 50 years of incredible
technical advancement; AI is still just an elusive concept. This statement
is not a denial of the progress now being made, but simply seeks to
keep this progress in perspective when looking at the long road ahead.