Ok, so you have a
correspondence theory. Not terribly novel or specific, but definitely a right 
focus.
That’s correspondence between model and accessible environment, and the only
way to quantify it is comparison between the two. Both are supposed to expand 
with an indefinite input stream to learn / build a model from, incrementally.


 


The only thing we can
define is a starting point. Which is pixels, the limit of resolution. Start
from the first pixel: initial model to predict from, and compare it to adjacent
ones: immediate environment. Center-to-rim comparison in kernels, to compute 
your
correspondence. That’s what edge detection operators (Sobel) do, they just use
perverse terminology to describe it.


 


And then you have to
cluster these initial inputs according to resulting correspondence, because
that’s compression: the only way to manage search (comparisons) across your 
ever-expanding
model. The clusters are patterns: groups of input elements that can achieve
above-average compression, or internal correspondence.


 


There are
fundamentally two types of clustering: centroid-based or vertical and
connectivity based or lateral. The former is summation-first, forming a 
centroid,
then comparison of elements to that centroid to define weights: grey-scale 
inclusion
/ exclusion future inputs. That’s how all statistical learning operates,
inluding neural nets, think Perceptron. So you are not bringing anything new to
the table here.


 


But that primary
summation is horribly lossy, leading to ridiculous number of backprop cycles
needed to get anything meaningful. We and evolution hit on that mindless
destruction of data because it’s easy. Compared to immediate cross-comparison 
and
connectivity clustering, from flood-fill to graphs. Which would be a 
conceptually
consitent and ultimately efficient way to maximize your correspondence, but
whole lot more complex to do right.


 


Anyway that’s my spiel.
If you haven’t had enough, lots of updates since you last looked at it:


http://www.cognitivealgorithm.info.



------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T30e487c54062d930-Md862e3a2579c3e1405f17a63
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to