[LOCKSUIT]

You know how they say ANNs are DARK magic, a black box, uncontrollable, etc. Well I have a few questions now about that.

Few let's rule out the components in ANNs. I figured I'd keep that type (typo! lol 2nd typo!) because it's interesting.

Correct me if I'm wrong. ANNs are simply just node entries of data/representations, links between nodes, and of which with weight strengths. Right? Again: Nodes, links, strengths! Right? I know there's feature detectors that fly around the net but my concern is the net; accessing nodes and reading them, links between them, link strengths.

So, for our AGI projects, if the master symbolic-level Artificial Neural Network abover object recognizer nets learns data entry nodes, why is it having such a hard time keeping those nodes then, or reading them!? What if, we disallow representations in this 2nd level network and only use references ex. layer 3's node 4 refers to both layer 2's nodes 4 (cats love) and 5 (good food) which makes "cats love good food" 'node' in layer 3's node 4? And I can't see why it'd delete any node entries if it seen them enough in text or found one once that ranked high (understood by the ANN and understood to be very good/convinced).

Same for links. If it creates links from A to B D and H, then they do, all the links should stay...just the weight strengths can change, so why delete the links? Why is it hard to travel them?

As for weights, that may be not problematic then if data and links stay but weights improve by reading or testing skills.

[LOCKSUIT]

Now that I think about it, the inputs and outputs are labeled aren't they in ANNs, but not hidden layers are they. Yet I said the ANN should be able to say hey I see a nose, no face but I see a nose! "Outputs "nose" ". Well, I'll leave that flaw in my upcoming PICTURE YAYYY but I'll explain that is wrong.

[ranch vermin]

the master symbolic unit is a good idea,  its response is more likely than parts of the ann that respond only to specific circumstances.    these high level symbolic areas are more suited for the planning of things, where the details all end up confusing everything and making it look hard.   if small environmental changes happen (and especially its own artificial form) then the symbolic area of its mind is still valid,  just the details need replacing.

[LOCKSUIT]

Ohhh....

ranch

you mean the artificial environment is its internal environment where it discovers in also right?

Like, words describing real objects, like "hammer smashes glass", right?

[ranch vermin]

yes,  but when i think about "hammer smashes glass"  i mean more "hammer SMASHING glass"  as in a report from the external environment sorrounding the droid turning into quasi text in its master symbolic section. 

Its not quite NLP tho,  its a bit of a haphazzard pattern recording.  Because im hopeless at NLP.

[LOCKSUIT]

Yeah but once you record it in as "hamer smash glass", you can compare it to other recorded ones and make internal discoveries...

Like:

hammer canBreak glass......knife can break brittel things.....glass is brittle....therefore knfes can also break glass!

[ranch vermin]

You can also search (like a chess engine) through your collected data with a goal in mind,  the system can maximize the goal inside its given set of linked information.  (the model.)

Thats what im planning on doing,  gives you a primitive yet unbiased result of a datastructure.    yet can be costly to do,   and its hard to think of.

[AgentSmith]

Now that I think about it, the inputs and outputs are labeled aren't they in ANNs, but not hidden layers are they.

Yes, hidden units are not labeled. Their weights are adjusted such that their impact increases the quality of the output produced by the ANN for some input. In supervised learning this quality is expressed by some similarity measure that tells how similar the produced output was to the output label for the input label.

[LOCKSUIT]

OH YEAH I MADE THIS THREAD

ANNs are tangled. There's no point in looking there, really. But do look there. It's where you begin your smarty pants. Now you on top.

[AgentSmith]

ANNs are tangled. There's no point in looking there, really.

Looking at the weights of an ANN is pointless, very right. And nobody does so, because in general no human is able to predict the outcome of an ANN (and yes, this includes the "experts"). What you look at is the result your ANN produces after you let it run. Quite simple and intuitive.

[OMITS]

Definitions! ANN?