Let's take a deep layer architecture. Every node is connected to every node in
every layer. Nodes are representations that can be activated by multiple very
different inputs, after training. Bottom nodes are a-z alphabet inputs+outputs.
So what happens if you hear 'hi there'? Well these 7 nodes better not activate
all nodes the same amount! It's ok it's fully connected, the weights aren't all
100%, they are initialized random at first 23% 89% 12% etc. So here we want the
activated input nodes to flow certain places, then the next certain places, and
so on. We want multiple different but still similar inputs to activate the same
places/nodes. If the exact same input is heard later 'hi there' it flows t the
same place it stored yesterday, and so do similar phrases based on delay and
context handles! If we take my simple hierarchy, we can too remove nodes to
make it not so huge and end up with representations that can be activated by
many (more than it [already] can!) ex. we have 'the cat ate food' 'this dog
loves kibble' 'his cat ate the old dinner' 'cats loves food' and remove all but
one and end up with 'cats food love' because the word rearrangement etc enable
it to be triggered by them all the most, if 'cat' is in all of them at x
position relative then we want to average the position. So as far as I'm aware
this net of mine is perfectly able to handle everything thrown at it. If you
want a node that's activated by both hi and hello and welcome, first of all
these words all do exist, but anyway you'd combine them into a node that says
'weheliom', however I don't remember every re-generating such a deformity so, I
think this only happens at word level as shown but only slightly. I also
generate word by word when generate a story, not Next Bit or Next Letter. I can
though, I can though t, I can though tr, I can though try, but it's not most
consciously (focus) used clearly.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T409fc28ec41e6e3a-M9f9ec8cdb0797d7e792f6337
Delivery options: https://agi.topicbox.com/groups/agi/subscription
