Creating New Informational Primitives in Minds and Machines - Computers and Creativity - page 398

mysteries, and the present situation in the neurosciences is not unlike biology before
DNA-based mechanisms of inheritance were understood.
Despite resurgent interest in neuronal temporal synchronies and oscillations,
mainstream opinion in the neurosciences still heavily favours neural firing rate
codes and their connectionist architectures over temporal codes and timing architec-
tures. For introductory overviews of how connectionist networks operate, see (Arbib
1989 ; 2003 , Horgan and Tienson 1996 , Churchland and Sejnowski 1992 , Anderson
et al. 1988 , Boden 2006 , Marcus 2001 ,Rose 2006 ). Although strictly connection-
ist schemes can be shown to work in principle for simple tasks, there are still few
concrete, neurally-grounded demonstrations of how connectionist networks in real
brains might flexibly and reliably function to carry out complex tasks, such as the
parsing of visual and auditory scenes, or to integrate novel, multimodal informa-
tion. We have yet to develop robust machine vision and listening systems that can
perform in real world contexts on par with many animals.
In the late 1980s a “connectionism-computationalism” debate ensued about
whether connectionist networks are at least theoretically capable of the kinds of
combinatorial creativities we humans produce when we form novel, meaningful
sentences out of pre-existing lexical and conceptual primitives (Marcus 2001 ,Hor-
gan and Tienson 1996 , Boden 2006 ,Rose 2006 ). Proponents of computationalism
argued that the discrete symbols and explicit computations of classical logics are
needed in order to flexibly handle arbitrary combinations of primitives. On the other
hand, the brain appears to operate as a distributed network that functions through
the mass statistics of ensembles of adaptive neuronal elements, where the discrete
symbols and computational operations of classical logics are nowhere yet to be seen.
But difficulties arise when one attempts to use subsymbolic processing in connec-
tionist nets to implement simple conceptual operations that any child can do. It's
not necessarily impossible to get some of these operations to work in modified
connectionist nets, but the implementations generally do not appear to be robust,
flexible, scalable, or neurally plausible (Marcus 2001 ). It is possible, however, that
fundamentally different kinds of neural networks with different types of signals and
informational topologies can support classical logics using distributed elements and
operations.
Because we have the strong and persistent feeling that we do not yet understand
even the basics of how the brain operates, the alternative view of the brain out-
lined here, which instead is based on multidimensional temporal codes, should be
regarded as highly provisional and speculative in nature, more rudimentary heuristic
than refined model.
15.5.2 Brains as Networks of Adaptive Pattern-Resonances
Brains are simultaneously communications networks, anticipatory correlation ma-
chines, and purposive, semantic engines that analyse their sensory inputs in light
of previous experience to organise, direct, and coordinate effective action. Animals