Chapter 6 briefly introduces connectionist models and their functioning. The purpose of this appendix is to go into more detail about how these systems work, and to give the unfamiliar reader a feel for their various capabilities. As discussed in Chapter 6, the connectionist approach is motivated by the low-level architecture of the brain – whence the term “artificial neural networks” applied to connectionist models. The first essential idea about artificial neural networks is that, like the brain, which is a network of a huge number (1010) of tightly connected neurons, artificial networks are composed of a large number (typically ranging from tens to hundreds) of interconnected “nodes.” Far from being an architectural analogue of a biological neuron, however, the node of connectionist networks is “an abstract neuron” – that is, it lacks many of the structural and functional details of a real neuron such as the large number of synapses (104 on the average), differentiated input and output channels, distinct electrical and chemical processes, three-dimensional space distribution of neurons and distal connections among them, the time lag involved in interneuronal communication, and so on (Rumelhart 1989: 207). It is a simple processing element that communicates by “sending numbers along the lines that connect it [to other] processing elements” (the equivalent of firing rate in brain neurons) (ibid.). This, according to Rumelhart, is what accounts for the similarity of artificial neural networks to their natural counterparts.