While connectionism is a theory of neural architecture that provides a description of cognition at the physical level, the classical theory of mind is one of cognitive architecture that depicts cognition at the computational level (Litch, 1997). The connectionist model is process orientated as all cognitive processes are the result of the group of processes that take place in the brain. In contrast, classical is representation orientated (Stillings et al., 1995). The classicist models the mind as a symbolic processor. This notion postulates that mental items are symbols that have semantics (meanings) and syntax (rules). Thus, all cognitive processes are rule governed manipulations of symbols (Betchel, 1988).
The two theories have further differences. Firstly, the classicist considers symbols as both causal and meaningful entities. As a result, the classical theory of mind vindicates Folk Psychology, regarding it as a rational and sound archetype (Stillings et al., 1995). Moreover, this maintains that central characteristics of the mind are higher order properties, such as beliefs and desires. On the other hand, connectionism has varying implications for folk psychology; later discussed.
Another distinction is what is constituted as learning. The classicist describes learning as being able to use programmed symbols to present the world in the mind, which necessitates that each rule and symbol are explicitly programmed into the system; thus instructed that an input produces a distinct output (Poersch, 2005). Whereas, the connectionist classifies learning as the ability to modify the weights of connections in order to attain a “desired” output. The connectionist model is advantageous as ANNs have shown the ability to learn (Garson, 2008). At first the weights of the model are set randomly, where an input produces an incorrect output. This output is then compared to the “desired” output, and the weights are adjusted accordingly. This process is known as back propagation and is repeated thousands of times using a set of training inputs (ibid), however, is a minor flaw, as back propagation learning is very slow. The test for successful training, thus learning, is for the model to correctly classify items outside the training set.
Also, the way in which each theory states rules are acquired differs. The classical theory of mind insinuates there are innate syntactic rules which are initially programmed into the system. In contrast, the connectionist declares there are no innate rules for intellectual abilities, as rules are acquired through experience (Seidenberg & MacDonalds, 1999), and therefore are learnt. The classical model faces a problem, as when presented with a new input, the model will stop, until the input is explicitly programmed into the system. Instead, the connectionist model, as a result of automatic generalization, can recognize familiar aspects of a new input and continue to function (Garson, 2008). Automatically generalizing an input produces a stronger input-output relationship, consequently making quicker computations when dealing with familiar inputs. Furthermore, automatic generalization bares another advantage as ANNs can fill in missing parts of an input, whereas the classical model cannot (Poersch, 2005).
Rumelhart and McClelland (1986) investigated whether the connectionist model could account for language by examining the acquisition of past tense English verbs, both regular (e.g. walk/walked) and irregular (go/went). Rules were not programmed into the network. The authors found that the network’s development mimicked that of children, as both acquire and use language in a rule governed way. When children acquire the past tense suffix “-ed”, it is applied to all verbs (Burton, 2009), and the network mirrored this outcome. This functions as psychological validity for the connectionist, a noteworthy advantage over the classicist.
However, there were some limitations to the study. Primarily, Rumelhart and McClelland’s explanation of ANNs as a definitive mimic of child acquisition of language is somewhat incorrect. When children acquire the “-ed” suffix they return to the present form and apply the suffix; e.g. eat -> ate -> eated. Children also make doubly errors, e.g. ated. Rummelhart and McClelland inaccurately postulated that children make this error due to following the wrong root to produce words, whereas ANNs make this error due to ‘blending’. Furthermore, the network made errors on a large percentage of words tested on; e.g. taught grow -> grew, but when tested, blow produced blowed instead of blew. The study also found an underlying issue of the connectionist model, as computations are relatively simple tasks, and whether the model can complete more complex task is of question. Despite these issues, Rumelhart and McClelland concluded that ANNs can provide a satisfactory account of language. Furthermore, progress in ANNs and the development of a more complex model may only require implementing a symbolic rule-based approach.
In addition to demonstrating human intellectual abilities, the connectionist model has further distinct strengths. The model is robust, which is beneficial as ANNs sustain resistance to noise and gracefully degrade when damaged (Garson, 2008). Thus, unlike the classical model, the connectionist model could continue functioning when some of its components are damaged.
Yet, the connectionist paradigm is not without flaws. Neuroscience has revealed that the brain is much more complex than current ANNs (Litch, 1997). Also, ANNs are rather limited as they do not account for the hormonal activity which is controlled by, and takes place in the brain (Mader, 2006). One could conclude that it is relatively more plausible than previous models, however further progress and developments are required to make ANNs neurologically realistic. Another limitation to ANNs is that their connectivity is limited, as nodes have multiple inputs for only one output. Also, Nodes are limited as they only discriminate the sum of all the inputs, rather than examining each input individually. Additionally, ANNs have a finite space, restricting the amount of information they can maintain.
Another weakness of connectionism is the objection put forward by Fodor and Pylyshyn. Fodor and Pylyshyn (1988) argued that connectionism fails as a theory because it cannot account for the pervasive psychological phenomenon they refer to as “systematicity”. They reasoned that an explanation of “systematicity” requires classical architecture. Though, if connectionism could account for systematicity, then connectionism is implementing higher order entities, and therefore instantiates classical theory. As a result, it is classical theory that bears the explanation of mind, not connectionism.
Connectionism has been an incrementally controversial topic of philosophical debate, from which two distinct strains have emanated. Each strain has varying implications for the classical theory of mind and accordingly, folk psychology. The focal point of the debate derives from whether one views beliefs, and ultimately folk psychology, as a symbolic or non-symbolic entity. The view that declares connectionism as a distinct alternative to classical theory is known as radical connectionism. Radical connectionism can further be subdivided into, one which contests symbolic processing is eliminated but beliefs stand; or eliminative connectionism, where beliefs are eliminated. The forefront of radical connectionism is, as Pinker and Prince (1988) termed, “eliminative connectionism”. Pinker and Prince asserted connectionism and viewed folk psychology as symbolic, consequently eliminating folk psychology. Likewise, Ramsey, Stitch and Garon (1991) suggested the entire operation of processing cognition is characterized only by the physical processes that occur, thus eliminating beliefs and ultimately, folk psychology.
The other distinct strain of connectionism is somewhat of a combined theory of classical and connectionism, called implementational connectionism. Implementational connectionism postulates that symbol manipulations could be put into practice using ANNs (Stillings et al., 1995). Garson (2008) states that “the mind is a neural net; but it is also a symbolic processor at a higher and more abstract level of description”. Accordingly, symbolism accounts for higher order properties, hence folk psychology is maintained. Similarly, Bechtel (1988) claimed that connectionism provides the underlying framework in which a rule and representational, i.e. classical, model may be implemented. By maintaining this link to the symbolic entities, folk psychology still stands.
Overall, it is apparent that connectionism provides a relatively strong conceptual basis of the mind. Both connectionism and the classical theory of mind have definite merits and limitations. However, connectionism appears to some extent, provide answers to the challenges that face the classical theory of mind. Further developments to ANNs must be made in order for connectionism to situate itself as a distinct alternative to the classical theory of mind. One may conclude that the implementation of classical theory into the connectionist network appears to provide a plausible foundation for conceptualizing the mind.
References
Betchel, W. (1988). Connectionism and Rules and Representation Systems: are they compatible? Philosophical Psychology , 1 (1), 5-16.
Burton, L. W. (2009). Psychology (2nd Ed.). Hoboken, NJ: Wiley.
Garson, J. (2008). Connectionism. In E. N. Zalta (Ed.), The Stanford Encyclopedia of Philosophy (Winter 2008 Edition).
Kashima, Y., Gurumurthy, A. K., Ouschan, L., Chong, T., & Mattingley, J. (2007). Connectionism and Self: James, Mead, and the Stream of Enculturated Consciousness. Psychological Inquiry , 18 (2), 73 - 96.
Litch, M. (1997). Computation, connectionism and modelling the mind. Philosophical Psychology , 10 (3), 357 - 364.
Mader, S. S. (2006). Inquiry into life (11th ed.). The New York: McGraw Hill Companies, Inc.
Pinker, S., & Prince, A. (1988). On Language and Connectionism: Analysis of a Parallel Distributed Processing Model of Language Acquisition. Cognition , 23, 73-193.
Plunkett, K. (2000). O conexionismo hoje [connectionism today]. In J. M. Poersch, Psicolinguistica, ciencia e arte [Psycholinguistics, science and art] (pp. 109-122). Porto Alegre: EDIPUCRS.
Poersch, J. M. (2005). Um novo paradigma para a aprendizagem da linguagem: inteligencia Artificial conexionista [A new paradigm for learning language: Connectionist artificial inteligence]. Linguagem & Ensino , 8 (1), 161-183.
Ramsey, W., Stitch, S., & Garon. (1991). Connectionism, eliminativism, and the future of folk psychology. In W. Ramsey, S. Stich, & D. R. (Eds.), Philosophy and Connectionist Theory. Hilsdale, N. J.: Erlbaum.
Rumelhart, D., & McClelland, J. (1986). On Learning the Past Tenses of English Verbs. In D. Rumelhart, & J. a. McClelland, Parallel distributed processing: Explorations in the microstructure of cognition. (Vol. II, pp. 216-271). Cambridge, Massechusetts: MIT Press.
Seidenberg, M. S., & MacDonalds, M. C. (1999). A probalistic constraints approach to languae acquisition and processing. Cognitive Science (23), 569-588.
Stillings, N. A., Weisler, S. W., Chase, C. H., Feinstein, M. H., Garfield, J. L., & Rissland, E. L. (1995). Cognitive Science: An Introduction. Massechusetts: MIT Press.
Appendix 1
How neurons communicate (From http://users.rcn.com/jkimball.ma.ultranet/ BiologyPages/~/Neurons.html)
Appendix 2
The connectionist model (ANNs) (From http://plato.stanford.edu/entries/connectionism/) - Garson, J. (2008).