However there are a large number of obstructions to overcome with the connectionist computational theory of mind, and although proponents of it want us to give up on rule-based systems of explanation such as the classicists version, these things need to be rectified before it can be considered a sufficient theory of mind.

Although the connectionist version has a viable approach to association processing it seems to fail on such tasks as language and reasoning. The hindrances range in magnitude from minute to monstrous and include quandaries such as problems with the concept of an individual, the problem of compositionality, the problem of quantification, recursive thoughts/propositional problems, trouble with commonsense questions and last but not least the problem of systematicity.

The problem with the concept of an individual is an painless point to understand but is not so easily solved. If we have a set of identical twins or even something a bit more general like two trees of the same species, height, age etc. the connectionist system is blind to the fact that they are actually separate entities. There are a number of proposed responses to this dilemma but none of them are adequate.

The archetypal property of all our languages is compositionality and this is where another predicament with the connectionist CTM occurs. Representations are built out of parts and have their meaning based on the meanings of all the parts and from the way they are combined. Subsequently, due to the fact that our thoughts are built out of concepts, and are not stored as whole entities within our mind, we encounter major problems and all the attempts to rectify them turn out to be substandard halfway measures. After all, babies eating slugs, and slugs eating babies are two very distinct ideas whose meanings are assembled on the fly using syntax, and even though you may never have seen either sentence before you can understand it with ease.

Systematicity on the other hand seems to be one of the more obtrusive problems the connectionist CTM must face, and was identified by Fodor & Pylyshyn (1988). It states that the ability to think/create/understand a sentence of a particular structure is intrinsically connected to the ability to think/create/understand a sentence that has a related structure, in that, there is no human that can understand the meaning of the sentence “Jason loves Ashley” but fails to understand the concept of “Ashley loves Jason”. Regrettably though, connectionist models, even once they have been trained to recognize one sentence of the previous example, still fail to recognize the second. Systematicity must be guaranteed to work in connectionist models for their theory of mind to be a viable option as this is a key component of human intelligence, and, according to Fodor & Pylyshyn, is a given in the classicists approach.

The connectionist computational theory of mind, although on distant inspection appears to be a much healthier option than other models, seems to have just as many drawbacks as them, and once again, cannot be considered a viable option for a complete theory of mind without at least some of these issues being overcome.

The ones we have discussed would be a good starting point but there are also other smaller, but still very relevant ones that must be addressed. Such as, the length of time it takes to train a connectionist model vs the length of time it takes a human to learn similar tasks, the fact that back propagation in connectionist models appears to ‘cheat’ as there is no evidence of our minds working in such ways and the fact that connectionist models fail to recognize our recursive thoughts i.e. a proposition embedded within another proposition. Until such times as these are solved, the connectionist CTM is promising, but still has a lot of work ahead of it.

The main problems that the classical CTM encounters are Searle’s Chinese Room argument, the problem of homunculus, how there seems to be syntax without any semantics, the frame problem of relevance and updating and the cost of representational atomism.

Probably one of the more famous arguments against the classical CTM is that of the Chinese Room argument conducted by John R. Searle. The overall point of this argument is that we have the syntax in place but it is lacking semantics. A computer can use rules to interpret symbols but this does not mean it understands what is going on. To put it another way, semantics is not intrinsic to syntax, just because a computer can understand a set of predefined rules or algorithms doesn’t mean it actually understands what its doing.

Another more recent argument, again by Searle, that by his own confession he should have seen ten years ago is that syntax seems to be imposed on the system by its observers. The fact that we can make a machine out of just about anything and have it perform computation means that the processes are not intrinsic to the system but are completely abstract.

This spawns the problem of homunculus in that if syntax is not intrinsic to the system, and computation is defined syntactically, then nothing can intrinsically be a digital computer based exclusively on its corporeal properties. The computational theory of mind, always in some way commits the homunculus fallacy wherein they treat the mind as a little man inside the brain, and this little man uses the brain to execute his computations. However proponents of the CTM make a reasonable attempt at eradicating such a problem by supposing recursive decomposition. This states that there are multiple levels of homunculus within any system such that at the lowest level there is a ‘stupid’ homunculus doing simple yes-no, 1-0 type computations. Regrettably though the problem still remains in that the syntax is not intrinsic to the system, and we must therefore have a homunculus that stands outside the system to provide the system with its operating syntax.

There are a number of other problems that the classical computational theory of mind encounters which we have not had enough time to examine in depth here. The cost of representational atomism, together with the frame problem of relevance and updating, but also the fact that syntax has no causal powers other than those of the implementing medium all need to be explained and overcome. But that isn’t to say that the brain isn’t a digital computer, on the contrary, some other form of the computational theory of mind might contain a better explanation of how things work.

Stay tuned tomorrow for a short paper on the Connectionist Computational Theory of Mind…