The Philosophy of Psychology

in terms of Marr’s (1982) three levels of analysis:Wrst there is the (top)
functionallevel, at which mental-state transitions are characterisedas such,
and at which intentional laws are discovered and formulated; then there is
the (middle)algorithmiclevel, at which the rules for transforming Men-
talese sentences may be speciWed; and thenWnally there is the (bottom)
implementationallevel, where physical processes suYcient to execute those
algorithmic steps may be described.
This classical picture of the way in which cognitive processes are realised
in the brain has been challenged by recent progress in connectionism. But
it is important to distinguish between connectionism as a claim about the
mere lower-levelimplementationof cognitive processes (Marr’s bottom
level), on the one hand, and attempts to use connectionism to usurp
altogether either the level of algorithmic explanation (middle) or the level
of psychological description (top), on the other. The former is no threat to
the classical computational account of cognition: it is possible for a sym-
bol-crunching program to run on a connectionist machine, just as connec-
tionist networks are in fact modelled by programs running on orthodox
digital computers. It is the extension of connectionism into the cognitive –
algorithmic – domain which is controversial, whether it be to propose
type-identities with psychological properties, or to substitute replacements
for them.

2.2 Some common (mostly bad) arguments for connectionism

As we explained in chapter 1 (section 2.5), connectionism is often defended
on grounds of neurological plausibility. It is said that the fact that nerve
cells in the cortex each connect with many hundreds of other such cells,
often with extensive ‘feed-back’ as well as ‘feed-forward’ connections,
suggests that representations are likely to bedistributedacross such neural
networks. But this is not so: there is simply no relationship between the one
idea and the other. And from what little is known of representation within
actual neural systems it seems that it is generally local, rather than dis-
tributed.
For example, the area of primary visual cortex known as ‘V1’ (the area
which is distinctively damaged in the case of blindsight; seeWgure 3.1)
appears to be a retinotopic map, in such a way that stimulation of diVerent
areas of the retina maps directly onto stimulation of similarly spatially-
related areas of the cortex – so if you could see the pattern of stimulation in
V1, you could, quite literally, see what the subject was seeing. And al-
though the visual system then bifurcates into a number of diVerent streams
of processing – of which the ‘what-stream’ and the ‘where-stream’ are the
most salient – it is still possible toWnd individual cells, or small groups of

Mentalese versus connectionism 197