Neural networks represent the connectionist paradigm, which holds that the brain function representations are by connections, not symbols.  Neural networks have proved to be quicker and more accurate than conventional computer applications in many areas.

        One such area has been in language acquisition by children.  Language is considered one of the most complex rule-based operations performed by the brain.  Language is a function of all humans.  Noam Chomsky argued that language was innate to all humans, that all humans where born with some form of language instinct (cf. Pinker 1994).   It does not seem plausible that humans are born with all of the “worlds grammar books” already loaded into their brains.  If that were the case, children would not make errors when using irregular past-tense verbs, such as (sing, sang).  However, within a few short years children are able to master these complex language rules and their associated exceptions. [Spitzer,Pp.29-30]

        Around 1986 neural networks were employed to improve our understanding of past-tense verb form acquisition by Rumelhart and McClelland.  The neural network they used had 460 input-nodes and 460 output-nodes.  Every input-node was connected to every output-node, thus giving the network 211,600 synapses.  During the training phase the network was given a data set of 420 stem words.  Its task was to produce 420 past-tense forms of the stem words.  The network was allowed to initialize its own weights and to adjust them after each trial until the it reach 100% accuracy with the training data set. [Spitzer,Pp. 30-31]

        After 79,900 trials the network showed 92% accuracy with regular past-tense verbs forms and 84% accurate with irregular past-tense verb forms when given an unknown data set.  It is most notable that the network’s learning curve matched that of a child’s.  The network produced some of the same errors that children do when learning speech.  The network learned gradually how to yield correct past-tense verb forms, the same as a child does.  The fact that both the network and children produced similar sorts of errors shows strong support that children and the neural network learn the same way. [Spitzer,Pp. 31-32]

        Rumelhart and McClellend concluded that humans are not explicitly born with rules that govern language.  Language is learned through a series of resetting the threshold level of neural connections.   “Although a number of details of the model proposed by Rumelhart and McClellend have been subjected to criticism (cf. Marcus 1995, Pinker & Prince 1988, Plunkett 1995), the model is very plausible and has been confirmed by further simulation experiments (cf. Hoeffner 1992) [Spitzer,P.32]".  These simulations have shown that language behavior can be learned without any explicit internal representation of rules that govern language acquisition. [Spitzer,P.32]
 
        The fact that a certain task can be simulated on a computer running a neural network model does not mean that the task is implemented in the brain in that exact manner.  Further more, although a network simulation model, appears to behave the same way a real nervous system does, this can not be taken as proof that biological nervous systems operate in the same manner.  However, a neural network is able to demonstrate operating principles that might be found in real nature.  As recently heard in a class, "anytime a model can produce the same type of errors found within nature we are onto something [Graci]". [Spitzer,P.34]