Ai Dreams Forum
Artificial Intelligence => General AI Discussion => Topic started by: LOCKSUIT on December 17, 2016, 04:49:41 pm
-
I've only been on AI for over a year now.
And how many times have I read discussions/articles/Wikipedia on Generalizing and Fitness?
Generalizing
Overfit
Underfit
What do they mean in REAL world terms? I don't want you to walk up to me holding a musical-sheet and say a bunch of convexing vector terms.
Does generalize mean that the actions to pick up a fork also pick up a knife? Or that the sensory input is able to search memory and still make a decision based on what you see?
-
Fitness is usually a measure of how well a trained system accomplishes the designed/ desired functions. This term is used a lot when referring to Genetic Algorithm based systems.
Not sure about overfit… unless you are referring to the over training of common neural nets to the point they become ineffective at recognising/ filtering the desired results.
Generalisation is usually the ability of a trained system to apply its skills to other problems/ actions sets, etc. So yes… if it can hold a fork it should be able to hold a knife. The system can use what it’s learned about the problem space to generalize, and solve/ accomplish similar problems/ functions with the same skill set/ knowledge.
:)
-
Not sure about overfit… unless you are referring to the over training of common neural nets to the point they become ineffective at recognising/ filtering the desired results.
The system (at least mine) can't become ineffective at recognising (only better) because the sensory input that enters will "still" have a match just for it and what it does is choose the match that "is a match" but is the highest rewarded (by rewards).
So fitness is just it improving. Note that's just what I described above.
Generalizing is not even have to be added it does dis too. Though note you want it to get the best fork actions, and the best knife actions, and the best paper-thin actions, etc.
-
What's the difference between these answers...
1+1=2
3-1=2
5-3=2
1*2=2
or...
Answer one
Answer one
Answer one
Answer one
:)
-
Each is a cue to different actions.
All 4 are the same and is a cue to make me stare at the above cues. (these will match to different actions.)
Cue search > actions out.
French Fry Nugget time soon.
-
Learning to See [Part 5: To Learn is to Generalize:
https://www.youtube.com/watch?v=efR8ybG7Ihs&list=PLiaHhY2iBX9ihLasvE8BKnS2Xg8AhY6iV&index=5 (https://www.youtube.com/watch?v=efR8ybG7Ihs&list=PLiaHhY2iBX9ihLasvE8BKnS2Xg8AhY6iV&index=5)
https://www.youtube.com/user/Taylorns34/playlists (https://www.youtube.com/user/Taylorns34/playlists)
-
@lock
First set…
Different number of pixels in each equation.
The order of the equations are written.
The average horizontal graphical pace they require.
The numbers and the order of the numbers.
The function symbols (+-*).
Each equation gives a different light level over an average area.
It gives a pleasing pattern (too me).
Did I use * on the last equation because I’d used + and – too much?
Etc…
Second set…
Besides the obvious order in which they are listed if you highlight with your mouse you will see each has a different number of spaces post text.
How does this make you feel? Can you imagine being in ‘my shoes’ when I wrote this? What do you think I'm trying to convey? How much experience and types of learning do you think you brought to bear on reading, understanding and replying to my question?
The system (at least mine) can't become ineffective at recognising (only better) because the sensory input that enters will "still" have a match just for it and what it does is choose the match that "is a match" but is the highest rewarded (by rewards).
How is saving images and linking rewards going to give results like this?
I'm not saying your approach is incorrect; who am I to say that, but rather than quoting the same ‘formula’ rewards to images, you might want to consider exactly what our brains accomplish and how you can adapt your system to convey this.
:)
-
The system (at least mine) can't become ineffective at recognising (only better) because the sensory input that enters will "still" have a match just for it and what it does is choose the match that "is a match" but is the highest rewarded (by rewards).
How is saving images and linking rewards going to give results like this?
Because once a image is saved, then any input later afterwards that CAN match (ex. a needle or toothpick will match a string) then they will 2 years later. But when it does match (as long as similar) it will select the highest-ranked sense. ~~~ Highest ranked as long as similar.
First set…
Different number of pixels in each equation.
The order of the equations are written.
The average horizontal graphical pace they require.
The numbers and the order of the numbers.
The function symbols (+-*).
Each equation gives a different light level over an average area.
It gives a pleasing pattern (too me).
Did I use * on the last equation because I’d used + and – too much?
Etc…
"orange voice"
noooooooo o o o o
The visual image (area of image as focus*) of those math equations searches memory and links to the speech/write actions just-like-that. 0 calculation needed. Plus can't calculate cus names of numbers don't set calculate calibration it is just sensory linking to actionory.
You can't make your brain make "five" send 5 sparks. Talk to a chinese, they say 7 + 7, you no even understand. Understand is sense>action selection.
Ah, korrelan, the CNN sense selection isn't just by scoring up the most features, it's by the reward sum of senses added-in TOO. Similar but highest ranking WINS. But not too non-similar, it'd blow system and never Guess again. Lol>Oh ya - fireplace - does eat actions.
-
Perfect is the enemy of good. ~ Voltaire
-
Deep learning is trained on so many images That the K mean border are really good of getting all generalized images
with in.
if you lack data to train your NN detector then you will need a transformation algorithm and a image back
propagation algorithm.
The idea is to transform one object to another. The back propagation algorithm will let you find the most direct route, by
letting you find the weight you need to mess with. This recorded linear path of change will allow you to over shoot the
targets and will allow to see other possibilities beyond. So when a new object comes around it might lay on a exciting
transformation path? Or you may have to make a new path connection with others.
Transformation algorithm:
https://www.youtube.com/watch?v=XtC1M7nrDk0 (https://www.youtube.com/watch?v=XtC1M7nrDk0)
https://www.youtube.com/watch?v=Te-iGDX34Co (https://www.youtube.com/watch?v=Te-iGDX34Co)
back propagation algorithms:
http://www.benfrederickson.com/numerical-optimization/ (http://www.benfrederickson.com/numerical-optimization/)
Recorded format:
http://www.google.com/search?q=images+clustering&espv=2&biw=1195&bih=615&tbm=isch&tbo=u&source=univ&sa=X&ved=0ahUKEwin7pyuvP7QAhViw1QKHX79BIcQsAQIIw#imgrc=0SBcdpjt4tr2YM%3A (http://www.google.com/search?q=images+clustering&espv=2&biw=1195&bih=615&tbm=isch&tbo=u&source=univ&sa=X&ved=0ahUKEwin7pyuvP7QAhViw1QKHX79BIcQsAQIIw#imgrc=0SBcdpjt4tr2YM%3A)
-
Gooooooooooooooooooooood d
D
I don't think that belongs on this thread keghn. . .
-
if it uses less pixel sample data is it also generalizing ?
-
QUESTIONNNNNNNNNNNN
If the triangle in the link knows which of the 3 dots must flipflop, then we already know say in 3D where the dark zones/mesh are and could just point your finger at them. Why must we watch it flipflop when I see the dam target!? We already have the 3D bunny rabbit mesh! (above vid)
At the video's end (above) it shows a 3D-ized person walking (by multiple cameras I suppose) (hence the dots wrap arounddd him)...but this is the same as the thing keghn showed us before - dots stuck on a 3D-ized mesh in the pc. Therefore the later adding of the green gas dots to the already 3D-ized guy has no purpose...
http://www.benfrederickson.com/numerical-optimization/ (http://www.benfrederickson.com/numerical-optimization/)
-
Gonna hound both keghn & korrelan on the above.....I gotta know
-
Kehgn
+
Korrelan
Remember the cameras circling around an object and engram 3D object into the computer? With dots on the surface walls too. Multicolor faces of/on object.
And remember GAN? GAN can make a 2D picture into a 3D (2D>3D) model mesh object into the computer, just from looking at the image by itself.
Question --- Please explain what does the "Growing Neural Gas" do different. Why use it? I need it? The upper above 2 videos you posted show it, and it can't be wraping itself onto/around a image (in 3D sense) all on its own, so, what is it being used for???
Question #2 --- And if I'm in a 3D editor such as Blender and import a 3D mesh object, then why would I want the "Growing Neural Gas" to give me a less-quality duplicate/copy of the bunny rabbit animal shown above? WHY? I don't see any use/purpose/capability in/of/for this "Growing Neural Gas" NOR do I see how it is being used when GAN/circling-cameras are needed as said/explained remember remember.
-
https://en.wikipedia.org/wiki/Low_poly (https://en.wikipedia.org/wiki/Low_poly)
https://www.google.com/search?q=low+poly+art&espv=2&biw=1214&bih=611&tbm=isch&tbo=u&source=univ&sa=X&ved=0ahUKEwj7nb681IrRAhUU4mMKHW1pDKUQsAQIIg#imgrc=Tzu8BhT8R0WhhM%3A (https://www.google.com/search?q=low+poly+art&espv=2&biw=1214&bih=611&tbm=isch&tbo=u&source=univ&sa=X&ved=0ahUKEwj7nb681IrRAhUU4mMKHW1pDKUQsAQIIg#imgrc=Tzu8BhT8R0WhhM%3A)
https://www.youtube.com/watch?v=q_6ytN1ybmI (https://www.youtube.com/watch?v=q_6ytN1ybmI)
-
oh I didn't see the 2cd page, k wait I must watch/etc wait
-
.........................that doesn't answer any of my questions...............................
-
What is the difference between:
"Growing Neural Gas"
and
"GANs"
and
"cameras going around/scanning an object"