The net as shown will, if the ‘A’ relay cells are stimulated by +1, -1, -1, +1 automatically deliver -1, -1, +1 & +1 where the synaptic values are thus set. ‘Rose seen’ will deliver ‘smell of rose’ (or vice versa). The net works just as well in reverse, of course. ‘Smelled rose’ will deliver ‘vision of rose’. And it gets better. The net may, in fact, be used to make more than one association. It may be used to associate more than one pair of items. It might also, for example, associate input from hearing with a memorised ‘fact’ - I might think ‘Beethoven’ (among other things) when I heard the Missa Solemnis. When you consider this multifunctionality and the fact that you have millions of neurons each capable of at least a couple of thousand synaptic connections it is easy to see where the mind’s colossal capacity comes from. It is also easy to see why behaviourist pair-associative ideas (e.g. stimulus-response bonds & reinforcement) made such sense (even while being so fundamentally ridiculous). And why associative notions still do make sense, though not in exactly that manner.

And it gets better still. There is one major and obvious aspect of human mental capacity which the net explains very gracefully. We saw (in chapter two) that our minds may recognise stuff by feature analysis but that, crucially, we do not have to have perfect input in order to make perfect decisions. We do not demand absolute specifications; we routinely get, and routinely accept, approximations. We think fuzzy and can manage fuzzy. Indeed, as we saw, the ability to use a degree of approximation is just about essential. We can, and regularly do, accept thoroughly imprecise data (as with the dog) and process it through to perfectly precise (and almost always correct) conclusions.

Pattern associator nets elegantly process imperfect inputs and reach perfect conclusions from them. To go back to our specific little net; if we saw a rose, but it was not a good example (if it was a ‘degraded stimulus’) - most of its petals gone and laying, somewhat squashed, on the floor - the input on the ‘A’ relay cells might degrade to only half their value (0.5 instead of 1) but would retain their direction (+ or -). Passed across the net these values would deliver exactly half the stimulus strength onto the ‘B’ relay cells but, again, in the same direction (+ or -). This output, particularly as still in the correct directional pattern, will be adequate to stimulate the relevant olfactory cortex and induce ‘smell of rose’. Faced with a very weedy ‘rose seen’ we still managed ‘smell of rose’ perfectly. Even if one ‘A’ relay cell were to be knocked out, to cease function, the net will still deliver a pattern across its remaining synapses perfectly adequate to stimulate association. In the words of Rumelhart & McClelland (1986 p. 36) the ‘...pattern retrieval performance of the model degrades gracefully both under degraded input and under damage.’

How does such a model ‘behave’? A computer simulation has been built to the specifications of a network of nets which build themselves up as information comes in. Each net associates one pair of bits of information, and the ‘synaptic’ connections set themselves according to an inbuilt Hebb-type rule. The simulation has been given various things to learn and has then been tested on its knowledge. Among these tests has been the formation of past tense of English verbs (no ‘rule’ about tense forming, just a lot of examples of past tense forms). After a number had been learned new ‘irregular’ verbs were presented (eg ‘purd’ & ‘gimp’). The program postulated regular past tense endings (‘purded’ & ‘gimped’ for example). The machine also ‘over-regularised’ some already learned irregular verbs, producing ‘camed’ and ‘wented’ just like children do. The machine behaved in an apparently rule-mediated way although it knew not a one.

The model learns to behave in accordance with the rule, not by explicitly noting that most words take -ed in the past tense and storing this rule away explicitly, but simply by building up a set of connections in a pattern associator through a long series of simple learning experiences. The same mechanisms of parallel distributed processing and connection modification ... serve, in this case, to produce implicit knowledge tantamount to a linguistic rule. The model also provides a fairly detailed account of a number of the specific aspects of the error patterns children made in learning the rule. In this sense it provides a richer and more detailed description of the acquisition process than any that falls naturally from the assumption that the child is building up a repertoire of explicit but inaccessible rules. (Rumelhart & McClelland 1986 p. 36)