Artificial Intelligence is defined as the science or technology of getting machines to do certain things that require intelligence and that were supposed to be performed by a human. However, there are different forms and definitions of natural intelligence and these forms are usually appropriate when developing systems that are effective in these areas. In relation to this, this paper explores the two models adopted in artificial intelligence namely the symbolic and the subsymbolic models.
This paper also tries to determine whether subsymbolic or connectionist and symbolic or rule-based models are competing or complementary approaches to artificial intelligence. Introduction Artificial Intelligence (AI) comprises tools, methods, and systems to generate solutions to problems that normally require human intelligence. The term “intelligence” is often defined as the capability of learning effectively, reacting adaptively, making proper decisions, communicating in images or languages in a sophisticated manner, and of understanding.
The main objectives of artificial intelligence are to develop systems and method to solve problems, usually through humans’ intellectual activity, such as language and speech processing, image recognition, prediction, and planning; thus, improving computer information systems; and to develop models for the simulation of the living organisms, particularly the human brain to enhance the understanding of how the human brain really works.
The main artificial intelligence directions of development are to develop systems and methods to solve problems without necessarily following the human’s way of doing so, but similar results must be provided, expert systems for example; and to develop systems and methods for solving artificial intelligence problems by modeling the human thinks or how the human brain works physically, artificial neural networks for example. In general, artificial intelligence is about the modeling of human intelligence. There are two main models adopted in artificial intelligence.
These are the symbolic and the subsymbolic models. The symbolic model is based on the manipulation of symbols while the subsymbolic model is based on neurocomputing. The Symbolic Model The symbolic model s based on the physical symbolic systems theory. A symbolic system is composed of two sets: (1) a set of symbols or elements which can be used for the construction or generation of more complex structures or elements; and (2) a set of rules and processes, which when applied to structures and symbols, produces new structures.
These symbols come with semantic meanings that represent objects and concepts. Predicate logic, propositional logic, and production systems facilitate dealing with a symbolic system. Some of the corresponding implementations of artificial intelligence are the logic programming, production languages, and the simple rule-based systems. Symbolic AI systems are also applied to expert systems, natural language processing, modeling cognitive processes, machine learning, and others.
Unfortunately, these do not really perform well in all cases especially when missing, uncertain, or inexact information is used, when parallel solution needs to be elaborated, or when only raw data is available and there is a need for a knowledge acquisition. However, these tasks are not really difficult for humans. The Subsymbolic Model The subsymbolic model asserts that intelligent behavior is carried out at a subsymbolic level higher or greater than the neuronal level. Knowledge processing is all about changing the states of networks made up of neurons while replicating or copying the analogy with the real neurons.
A single or a collection of neurons may represent a micro-feature of an object or a concept. It is possible to design intelligent systems that achieve the proper global behavior even if all the system components are simple and operate only on local information. The subsymbolic model does not only make the use of significant results in artificial neural networks area achieved in the last 2 decades in areas such as pattern recognition, and speech and image processing possible; it also allows the utilization of connectionist models for the processing of knowledge. Analysis of Symbolic and Subsymbolic Models
By their very nature, both the symbolic and subsymbolic models to artificial intelligence (AI) appear to be competing or incompatible (Taylor, 2005). The main difference is that the symbolic model representations are concatenative where they are accessible and changeable part by part. Conversely, the distributed representations cannot be modified or changed without affecting the other information in the network. This leads to different performance and learning properties for the two models. The symbolic model tends to be better in building abstractions and processing structure.
Neural networks on the other hand, discover the surface-level regularities naturally and perform under minor variations robustly. It is possible that eventually, all cognitions can be understood and interpreted in terms of the neural process, operating in the brain at the subsymbolic level. Nevertheless, this by no means would leave the symbolic approach irrelevant and unrelated at this point. Dror and Young noted that the symbolic and subsymbolic models may coexist for a period of time in the cognitive aspect of artificial intelligence provided they serve distinctly different roles (1994).
The analogy often used to describe this is the Newtonian physics and relativity where it is sometimes necessary to consider the low-level neural mechanisms to explain a certain phenomenon, while in some cases, a higher level of symbolic description is an adequate approximation and a more sophisticated, clear, and precise way to describe the process. There are many ways on how the symbolic and subsymbolic models of artificial intelligence (AI) may be complementary or may interact with each other.
One way is that they can be developed and utilized alternatively and separately. Another way is that hybrid systems incorporating the symbolic and subsymbolic systems may be developed and finally, subsymbolic systems may be used to model or display pure symbolic systems. However, reports claim that any idea coming from the symbolic model, even if it was implemented using the subsymbolic process, restricts the exploration and discovery of the new foundation of artificial intelligence on neural networks.
This may be true philosophically, in the practice of artificial intelligence, using the results of the symbolic method to guide the subsymbolic model is an effective and efficient way to make progress. Completely abandoning the old ideas about how artificial intelligence is put together and building everything using the connectionist foundations is such a mighty task. Several small experiments pertaining to artificial intelligence indicate that revolution might be in progress; however, it is far from being clear whether these completely distributed and unstructured systems will scale up.
When it comes to creating high level artificial intelligence, the “connectoplasm” approach may break down. The subsymbolic approach might be effective if it is possible to build a certain AI model, including the right structures and mechanisms, fed with the entire learning experience of human. Nevertheless, the computational task would still be prohibitive and it is probably takes a lot of years before a human could learn a language. Therefore, for AI, shortcuts are needed. Ideas on how a cognitive architecture of AI could be built are needed and how to break these into manageable parts must be learnt.
This is usually where the symbolic model comes in. Most of the researches about symbolic models have concentrated on taxonomy building and outlining the components of the artificial intelligence system, and it could also give valuable approach into how a certain connectionist model must be structured. Many of these symbolic models are supported strongly by psychological research, especially those at a high level. However, it is also right not to go too far in replicating or imitating symbolic models.
Blind implementation of symbolic architectures is unlikely or virtually impossible to lead to newer insights about artificial intelligence (Touretzky & Hinton, 1988). Using the concepts from symbolic models should be carried out in terms that are natural to subsymbolic systems. Moreover, as stronger algorithms become available, it is possible to create artificial intelligence models with fewer assumptions. In this sense, it is logical to think that the results of symbolic model researches are more like crutches instead of obstacles in the discovery of subsymbolic models.
Conclusion Based on the research presented in this paper, it is logical to conclude that subsymbolic or connectionist and symbolic or rule-based models are competing approaches to artificial intelligence (AI). Moreover, it has been concluded that a subsymbolic model may be used if we can solve artificial intelligence problems by ourselves but we can’t really explain precisely or clearly how. It is also possible to use this mode if the problems faces can’t be solved very well at all and if it is not important that the system can generate a verbal justification of decisions.
On the other hand, a symbolic or rule-based system may be used if the problem can be solved and can be explained how. This may also be applied if it is essential that a decision may be backtracked explicitly when a judgment is in question. With all these given cases, these two models do not really complement each other although they may interact in some points.
References
Dror, I. & Young, M. (1994). Evolution of Revolution. Psychology. 5:79. Taylor, J. G. (2005). Progress in Neurobiology. Paying Attention to Consciousness. 71, 305-335. Touretzky, D. S. & Hinton, G. E. (1988). Cognitive Science. 12:23, 466.