The target paper of this session, Janet Wiles' "The Connectionist Modeller's Toolkit: A Review of Some Basic Processes Over Distributed Memories" is an attempt to tackle the question: What are the primitives - the basic building blocks - of cognitive models? The paper provides an important service. It not only highlights the need to make our "primitives" explicit, it also provides a taxonomy of connectionist primitives and compares these with the basic computational processes typically used in symbolic reasoning.
Although Wiles' paper begins the work of identifying the primitives needed to model human cognition, it does not say how to decide when a specific process is valuable for theorizing. There are a large number of potential computational processes. Indeed, we could probably develop these without end. How do we decide which ones are really used in human cognition? Are humans furnished with lots of potential computational processes or are we reduced instruction set computers? (Are we IBM mainframes or RISC machines?) Because the same cognitive tasks can be accomplished in many different ways, we need some way of deciding which way is the human way (assuming there is just one human way). Until we have such a descision rule, we run the risk of simply choosing our theoretical building blocks on some a priori basis or simply relying on those that are currently "in fashion".
The first discussant, Michael Johnson, recognizes that fashion plays a role in theory building and warns us against unthinking adoption of any basic process. The second discussant, Zoltan Schreter, discusses several properties of localized and distributed representations. Both discussants, but especially Johnson, suggest that non-primitive processes are important in theorizing. Like Wiles, he does not say how to determine which primitives are the "right" ones. However, this may not be as important for Johnson because he views theories as "metaphors". The word "metaphor" is used often in connectionist debate but not always with the same meaning.
When I say that brains are like connectionist networks, I am making an analogy. I am suggesting that there are some ways in which brains and networks are similar. When I say that a person acts like a computer, I am using a simile. I am suggesting the the person is cold, calculating, and unemotional. Expressions such as "data streams", "mental lexicons", and "memory stores" are metaphors. There is no water in a data stream, no book of words in the skull and no storehouses for memories. Metaphors are, by definition, not literally applicable to the phenomenon being described.
These differences are not just pedantic. Theorists who claim that networks are in some ways analogous to human brains are making rather different claims from those who merely express their ideas about cognition and neurology in vivid, metaphorical language. It would be useful indeed if scientists working in the field of connectionism decided whether they believe that networks are really analogous to brains or whether they are just colourful ways of describing cognitive processes with no literal resemblance to the processes they are meant to describe.