[ExI] sciam blog article

[Rafal Smigrodzki]

On Tue, Mar 29, 2016 at 8:02 AM, Robin D Hanson <rhanson at gmu.edu> wrote:

>
> What is the evidence that one merely needs to model the cortex well to
> have a machine that can do most jobs as well or better than humans?
>
> ### You need more than the cortex model to do everything a human does but
a cortex model should suffice for all the intelligent, non-automatic
actions. Walking, simple emotions, basic sensor signal processing, manual
dexterity require a lot of non-cortical activity. The stuff of
intelligence, such as hierarchies of cognitive representations or libraries
of high-level behavior templates, are largely contained in the cortex.

I would list the following lines of evidence in favor of the cortex being
the seat of general human intelligence:

1. Lesion studies in humans and animals
2. Functional correlate studies in humans and animals
3. Comparative anatomy of humans and primates
4. Genetics of humans and primates
5. History of AI development

Ad. 1 - Generally, loss of parts of the cortex results in partial loss of
higher level functions. You can develop isolated acalculia, spatial
disorientation, various forms of agnosia, aphasias, apraxia, abulia,
emotional incontinence. There is a gradient of complexity of affected
functions along a few axes in the cortex - e.g. along the posterolateral
and posteroanterior occipital cortex there is a gradient towards more
abstract and non-localized visual function to be affected. On the other
hand, small lesions in the brainstem can take out a lot of motor or sensory
function while leaving higher representation unaffected. A sufficiently
large brainstem lesion affecting the ARAS will cause loss of consciousness,
so yes, you need subcortical structures for intelligence, in the same way
you need a power supply for your AlphaGo.

This line of evidence is extensive and detailed. A caveat is that large
cerebellar lesions in adults cause global cognitive dysfunction, however,
in primary cerebellar agenesis there is no significant cognitive
dysfunction, implying that the neocortex can adapt to function well without
cerebellar input, so cognition still remains a purely cortical process.

Ad. 2. Functional non-lesion work in humans and animals, using e.g. fMRI,
and electrophysiological approaches, finds neural correlates of cognition
in the cortex and only to a much lesser degree in subcortical areas. The
farther you go connection-wise from the cortex the less correlation to
cognition.

This is a massive body of work that I couldn't even begin to summarize here.

Ad. 3 - There is very little variation in the structure of the brainstem
between various primates (aside from size differences) but there are much
more pronounced differences in the structure of the cortex, with large
cortex and high metabolic cortical activity correlating with better
cognitive function.

This line of evidence is detailed but less so than lesion-derived
information.

Ad. 4 - There are significant genetic differences between humans and
non-human primates in genes that affect cortical development, more so than
in the genes affecting only the brainstem. The neocortex in primates is
modular at the small to medium level (cortical columns) but there is also a
component of non-modular, hierarchic structure (Brodmann areas) and the
shape and strength of connections at this hierarchic level is very
important for unique human functions.

This line of evidence is still work in progress.

Ad. 5 - It appears that robotics companies have already duplicated human
function at the lower level of walking, motor manipulation, simple visual
scene analysis. I do not know what is the structure of programs used for
these functions. I am assuming that these are task-specific programs that
do not need to embody extensive learning functions, and they are different
from the 200-layer neural networks used in modeling higher cognitive
function. If so, there would be an analogy to the division between the
task-specific, non-learning neural non-cortical networks and the general
function modular hierarchical cortical networks in the brain.

If you put a deep neural network with 2000 layers on top of whatever powers
ATLAS robots you could get a pretty close facsimile of a human mind in a
clumsy human body.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20160329/3fe06d3c/attachment.html>