[Korrelan]

@Art

I guess you just seem like a silver car… kind of person lol. 

@Lock

Soon it'll be time to give it a test here!

Nice resource. 

Using NAME is a human, humans are mammals, etc must be common testing solution, strange how close my test sentences where… and I’d not see this site.

 

[LOCKSUIT]

LOL "Harry's native language is "Commando""

That site is good resourse, it has CIA test, Harry, mind overview.
http://mindmap.iwarp.com/

lol one of his things says "Transform sentence."

[LOCKSUIT]

This is korrelan's expression when he saw that website:

[8pla.net]

Hey Korrelan,

I am enjoying your videos.

Thank you.

[Art]

Agreed! Keep it up, you're not that far away from entering those, "My Chatbot is smarter than your Chatbot" contests!! 

You can make it happen...

[8pla.net]

INTERFACE UPDATE:

http://www.chatbot.ml/MLP/

While it's not connected to the Deep Neural Network yet.
It does write an array of the training inputs and outputs
from the brand new training interface module.

______________________________________
« Edit Notes:  Updated broken link

[Art]

Your interface does not seem to be working properly. If one touches the colored area with their cursor and scrolls the mouse wheel, the colored area shifts upward as shown in the screen cap.

Probably just a slight fix but I thought you should be made aware of the snafu.

[8pla.net]

Thanks Art!

I can't explain that. It's like it has a mind of its own.
Maybe the Deep Neural Network was hungry
for those training inputs, and tried to escape. 

Apologies Korrelan for my deep neural network
running amuck on your thread with your readers. 

It's back in the cage now ( I think. )
I tightened the chains on the interface.

[Korrelan]

Apologies Korrelan for my deep neural network running amuck on your thread with your readers.

Not a problem… It’s always good to see other projects, no matter how unruly they are lol.

As I improve the Korrbot and it’s functions become more complex, I’m finding it difficult to track what it’s learned.

So I’ve wrote a little app to display it’s knowledge trees.  I can now easily visualise the data and add to it visually/ dynamically.



This is the first incarnation of the code; It will be improved along with Korrbot. The link colours will represent the predicate and the nodes the objects and subjects. It runs in 3D as well and has been designed for ease of use on my large touch screens.

If anyone has any ideas to improve the visualisation or vids of similar systems I can nick ideas from... speak up.

 

[LOCKSUIT]

I really like where this is going.

Can you make it form i.e. generate discoveries by first searching through the branches ex.
Hammers canBreak crystals, glass isA crystal, therefore hammers canBreak glass!
Hammers canBreak glass, windows areMadeOf glass, therefore hammers CanBreak windows!
Hammers canBreak windows, windows cost money, therefore hammers canBreak money!)

Is it possible to combine the 3D tree/net you have there with Parse Trees to make it stronger?

And what about the attachment to this post (see attachment).

Keep it simple so you can make new milestones. Do something new here.

[Korrelan]

Haha... highlighted a problem with double 's'... glass, I'm sure many peeps have had that problem before... lol

Hmmm... time to explain the transference/ conversion of energy types and material properties to the bot... lol

I'm glad this chatbot is just a bit of fun... and my AGI will solve all these problems.

 

[LOCKSUIT]

Hmm, it made a big connection loop I see there. So basically you said "Can hammers break windows?", then it says hmm maybe it's true, then it search-verifies the Transfer Learned (Acquisition) generation from you and calculates that it is true and says yes.

[LOCKSUIT]

Also, I realized something. It could learn "Iron-man loves crystals." (instead of "Windows areMadeOf glass."), and this presents a problem to which I have a solution. See, your algorithm found a in-direct connection loop that A=B B=C etc etc but....it can do that with anYthiNg, and easily become ultra pointless. Solution? You need confidences (weights!) of how much ex. B=C. Because if glass literally IS or "=" crystals, then hammers can break crystals and likely glass too, whereas if ironman loves crystals then it won't conclude hammers can break his iron-solid body if it really isn't connected or "=" crystals. Confidences (connection weights) also for how many facts support something (3 things weight-in one 1 node). There's also positive/neutral/negative integer ranks. Also another type of confidence for the probability of how true a fact is, ex. "Hammers canBreak iron man." has a probability of being 4% true.

[Korrelan]

Ed: Yes the chain event with using Iron man would have caused a problem... and yes I solved the problem by storing numerical values against the objects. So the tensile strength, brittleness and thermal properties for example can be applied to a set of functions.

I've just started writing the library of functions; so the bot can apply numerical judgements, formulas, rules, etc.

Still a bit rough around the edges but I'm very limited on time just lately.



 

[LOCKSUIT]

Hmm, if HUGE marshmallows (their tiny core) areMadeOf crystals, glass isMadeOf crystals, windows AreMadeOf glass, and windows cost money, then, if you ask "Can hugeMarshmallows break windows?", well, it can't begin the loop/connection around BY "predicate weight" because in the case it had learned that hammers canBreak crystals - the link predicate "canBreak" is not as strong as MadeOf (Madeof is stronger yet no proof hammers can break anything). And here with the hugeMarshmallows we have a predicate link "MadeOf" to crystals which is strong but doesn't say anything about being able to break anything. Hence this example loop would loop around and wrongly conclude hugeMarshmallows canBreak windows if asked so. SO, weight forwards a connection? No... So then is it predicate NAME? I guess so.