"Attentional" neural network modeling is based on assemblies of identical
neurons doing whatever it is that such assemblies might do, in a largely mathematical sense. Although there are many
wrinkles of interest within this research landscape (see work by J. Jose, R. Deth, J. Freedman; all at NU), the bottom
line is that none of this will ever explain thought or consciousness, because I maintain that the mammalian CNS is far more
complex and that mental language is an evolutionary by product of NPD -- neuronal phenotypic
The simplest vertebrate animals (e.g. larval zebrafish) are likely to have thousands
of distinct neuronal cell types. Humans are likely to have tens of thousands or millions or more distinct cell types.
Why does this matter? Each cell type constitutes a distinct computation. If our "mysterious flame" blossoms each
morning as a result of peculiar patterns of neural activity, is not the neural basis of this flame-igniting process important?
One point of Kurzweil's that stuck with me (AoSM) was a conversation where humans
were surprised to learn that cortex performed many hundreds of distinct neural computations. This is in sharp contrast
to the view of Mountcastle and Hawkins (and perhaps Markram) that we have one basic cortical module and that it will operate
on light or sound or language or motor control or whatever inputs it happens to be given. Chuck Stevens calculated NPD
in a cortical module at 300 distinct cell types, including a built-in reduction by a factor of ten, perhaps because the number
3000 might be ridiculed (my spec here). Great NPD yield great computational diversity, by any reasonable analysis, I
do expect. But the big question is this: does consciousness have anything to do with NPD? If it is a function
of computational complexity, then it might depend directly on NPD and this might be something important: something that distinguishes
us from 500-node clusters of Xeon processors.
This reintroduces a parallel wrinkle: compaction of information into the neural
code, aka Eric Baum. Great NPD allows one to compress large amounts of information into our DNA, i.e. into the
process that orchestrates the developmental appearance and wiring of diverse cell types into discrete computational structures. The
larval zebrafish with its 150,000 neurons contains such compact descriptions of the world, as Baum might say. The
first time it sees a paramecium, it knows to track it and eat it and it knows how to track it
and eat it, employing a very sophisticated sequence of locomotor maneuvers (Borla et al., 2002; McElligott and O;Malley,
2005). The much larger adult zebrafish brain enables a vastly more acrobatic display of locomotor behaviors, as
well as sleeping and waking. There is no reason to suppose that
it lacks consciousness, i.e. conscious experience of its world, with the sensory stream, prior experience and instinct guiding
its behavior in exactly the same way as these 3 items control our behaviors.
From this 2 opposing ideas emerge: (1) consciousness does not require great
complexity or great NPD (consistent with zebrafish consciousness) vs. (2) a big requirement for massive computational
complexity (in e.g. humans and chimps), such that lower vertebrates do not have "the right stuff" to ignite the mysterious
flame. This is a critical question in regards to the emergence of CDEs (conscious digital entities), because the less
stringent the requirements for consciousness, the sooner we expect to see it attained in silico. On the other hand, if we
can define a modest, precise set of architectures and layouts that are essential for mammalian consciousness, that will provide
a guide for constructing CDEs such that they will be here soon.
I have seen the problem and it is We.
extreme npd requires extreme genetic control, if the dvlpmntl process is to produce both distinct cell types and mechanisms
needed to assemble them into precise neural architectures that operate in an extremely adaptive, i.e. fxlly advantageous fashion.
This all fits very well with Baum's Gaiae-computational mother hypothesis where the information from trillions of trillions
of natural selection events is compacted into a small string of