In article <2f9h8j$pfe at wp-sp.nba.trw.com>,
Harry Erwin <erwin at trwacs.fp.trw.com> wrote:
>This speculative posting attempts to define a possible engineering concept
>for the architecture of the neocortex.
This posting adds 3-4 points I've worked out since...
>>I posit that the neocortex is made up of modules ('loops') consisting of
>pairs of KII systems arranged to form feedback systems. The general
>structure is as follows:
>>e_n mixes the efferent signal from Kx_n to produce the afferent signal to
>Kx_n+1. r_n necessarily inverts that mixing to transform a pattern
>attended to in the n+1th level into its constituent features in the nth
This is actually fairly easy to set up if it is allowed to develop from
the core outwards. Stimulate a column in Kx_n+1 so that its signal passes
around the loop and then through the e_n path. Kx_n+1 will automatically
condition the connections with signals and habituate the connections
without signals. Since the coreward loops are known to be able to do the
same processing as the loops on the periphery, there must be afferent
paths parallel to the 'e_n's, but that don't mix. In the olfactory system,
this seems to involve the AON. This allows the coreward loops to
'understand' the meaning of the mixed data based on the 'invariant'
version of the same data.
>>The weakness of a system designed this way is that attention in each layer
>is controlled by the layer below it. Eventually, this regression of
>control peters out, and you get a hard-coded attention function.
This statement led to some differences of opinion. The problem with a
finite number of layers is that the hierarchy of patterns then has a finite
depth. If I want to play a bluffing game, I have to allow for recursive
patterns. At some point, I have to have self-reference in some layer to
get that. I've never heard of neurons connecting to themselves, but I can
get around that by having two modules, A and B, with A connecting to B and
B connecting to A. I probably need a second module, C, which is used with
A in a similar manner to B, to create the inverse to the transforms via B.
Once I have that, A can evolve in the forward direction via B and in the
backward direction via C (to establish patterns for matching that imply
the current pattern being sought). I might be able to control this entire
system via simple, genetically/environmentally-established preferences
implemented via some of the nuclei.
Also, based on some of the discussions of what distinguishes modern H.
sapiens, it may be the increased efficiency of this system that was the major
evolutionary innovation that is believed to have occurred 50-100,000 years
Internet: herwin at cs.gmu.edu or erwin at trwacs.fp.trw.com
Working on Freeman nets....