Ai Dreams Forum
Artificial Intelligence => General AI Discussion => Topic started by: LOCKSUIT on August 07, 2016, 09:44:06 pm
-
What is labeling for?
Do not use the word data.
Hint: Supervision (rewards) are needed. Un-supervision (for memory searching) is needed. Supervision is Learning, which comes first. Un-supervision is for Reasoning, which comes next. What's missing? - How do you know to do a memory selection or generate guesses?
-
What is labeling for?
Do not use the word data.
Hint: Supervision (rewards) are needed. Un-supervision (for memory searching) is needed. Supervision is Learning, which comes first. Un-supervision is Reasoning, which comes next. What's missing? - How do you know to do a memory selection or generate guesses?
Hmm, I'm not sure what labeling is for. Since I don't have that kind of thing in my algorithm.
From what I understand, machine learning comprised of many types, but it seems that I can only think of two like yours, maybe? Except my viewpoint of those are different; Supervised learning is like someone teaching you with proper data. Unsupervised learning is learning by themselves, all by their own. Well, I'm sure that's how we learn as a human being.
Without your label thing -- if it's related, I indeed do know how to do make a proper memory selection and memory guessing, and -- hm, wait, do you? Interesting... I wonder how is yours like.
-
Heh, the freaking current gen AI humans call memory assorting "labeling data". What it really is is storage of *senses into the outer-layer of the cortex and saving them near senses that look alike for *assorted memory i.e. KNN.
And this is all unsupervised can do really, sure it could make the most frequently sensed sense the reward, but that could be walls.
But this unsupervised memory assorting is important for selecting a memory by external/internal input entering.
What I call labels is giving each stored sense a either +/neutral/- label & rank by rewards, then when selection occurs by search&match or memory-self-ignition, + means do&tweak linked actions, n/-/0match means don't-do&guess.
Note that you can DO zero-action-actions, they are generated too hehe!
-
Learning could be seen as comparing newly constructed theories to reality. Until it matches, new different theories are being constructed. When a match is recognized, we memorize it by attaching a label to it and move on.
So we'd need:
1. semi-random theory builder, guided by some genetic algorithm
2. comparing theory to reality, returning a percentage of matching
I think that learning is just a form of recognizing.
-
Learning say actions is from guessing actions, and then ranking them by pre-installed rewards being matched to say lots/little.
When input matches rewards it adds a +/n/- label to it too. When input gets saved, it strengthens itself, and when it matches a memory, it strengthens it as long as attentively sensed.
-
+ and - 0ver many iteration and accumulated and remembered of each cycle.
MarI/O - Machine Learning for Video Games
https://www.youtube.com/watch?v=qv6UVOQ0F44 (https://www.youtube.com/watch?v=qv6UVOQ0F44)
-
Yes I have already seen this video a whole year ago when I first began, the rewards are, well, obvious, and obviously are not the rewards we have - food, sex, girl, etc. The way we can still be mario masters is because we make mini-Rewards by linking NON-rewarded senses to the rewarded. You can make "the time you stay alive" bad or good! Cancer good/bad! Find gold=bad!
-
How does your AI unlearn things ?
-
Think about it, any AI can guess actions, and reward them by rewarding the sense that follows next such as food and *when see a neutral NON-rewarded sense such as a violin, meaning when it sees the violin it will be like seeing food. If it finds actions with a incorrect cue like this, or finds the wrong actions such as crawling in a circle, then a negative reward must happen to stop the strengthening of these, such as getting dizzy or not seeing the violin anymore. For crawling, visual input matches the highest ranked sense as long as similar and does so to a certain vision because this one is linked to the stronger than positive negDizzyRank, negative label means guess not tweak. As it guesses, neither acceleration nor dizzy sense is gotten nor strengthens the 2 memories linked to vision, so they weaken, and does so fast since not strong enough yet, then on the next input search&match, the NEG to make it guess is GONE, and so is the accel, so when it gets a accel reward now it will be the only one currently saved since all others vanished as it guessed, then it gets matched and done. It may be a curved crawl, but since positive labeled/marked senses mean do - not guess - it tweaks it instead & improves+zigzags! I demonstrated this with a doll, that's how it unlearns things. Use weakens faster if less strong, not de-increment if no reward. Also strength/rank can get high! Ex 97. No 1-3 lol.
Now play a violin and re-read this, just don't cry.
-
How would you tell it to immediately unlearn something ?
-
Well with a adult, most civilized and speaking humans have a stored phrase that says that OR simply have the sound "BAD!!" stored and linked to bad bad neg big time, then when you say woh Joey those are not the right moves to do on the tight-rope, then the first part of that input told exactly matches this bad memory and makes it get linked to just tried actions.
Nextly, importantly, this bad linked to the good cancels eachother out, else the adult would guess :#LOL or worse cancel the next best OLD actions and do the 2cddd best lol, this canceling happens for my demo at the wall and turning, simply put, the crawl and wall when seen cancel when wall is matched as approach, and next best turn actions previously guessed are done SINCE when it had previously guessed them it got ranked&labeled+ since had turned and seen the good far wall only linking to +crawl actions.
-
Cool, okay last one for tonight. How would you get it to argue that what it thinks is correct when told otherwise ?
-
Because the actions it has stored for explaining its AI blueprint are ranked so high that it says them, and for another's explaining to be accepted would mean for it to save and rank this paragraph higher than its own, which will only happen if you understand it, which clearly works on the ranking of senses and the strengthening of them since actually seeing a full fledged human robot made would link to a massively ranked sense (the video of such stored in brain) that's positive labeled.
Also some funny pics I drew:
http://advancessss.deviantart.com/art/7867887777-626916200?q=in%3Ascraps%20sort%3Atime%20gallery%3AADVANCESSSS&qo=3 (http://advancessss.deviantart.com/art/7867887777-626916200?q=in%3Ascraps%20sort%3Atime%20gallery%3AADVANCESSSS&qo=3)
http://advancessss.deviantart.com/art/78687777-626916330?q=in%3Ascraps%20sort%3Atime%20gallery%3AADVANCESSSS&qo=2 (http://advancessss.deviantart.com/art/78687777-626916330?q=in%3Ascraps%20sort%3Atime%20gallery%3AADVANCESSSS&qo=2)
-
Really time for me to sleep so final thought.
Okay, but that's just parrot fashion - it's not actually arguing a point using evidence or structured thought.
Isn't your system a more elaborate pattern matcher at the end of the day ?
I don't mean that as a bad thing, because I like the idea in general.
-
@Freddy
That's exactly my thought as well. That's the flawed part of the A.I.
Edit: Additionally, it isn't really possible for a human to unlearn something, unless they have amnesia. Every memory within us becomes experience, and that is the important part what makes us, us.
Even memory priority declines by itself, and something about this must be extremely important. So which means something else is must be happening. With this idea of degenerative priority, emotion like negative and positive, does not have ranking. It does not have "linked to bad bad neg big time".
My idea of memory priority goes for the idea of how "fast" neurons transfer information through the axon terminals. Since the general idea of how our information are encoded, is by the formation of new connections between the axons and dendrites, which means the memory priority is the strengthening of these connections. At the same time, these connections, gets weak, and sometimes fails due obstruction, making us forget temporarily.
Perhaps it's a good thing, for us humans, for example, we get something called "Ear worm" when we listen to music. The music will get stuck and loop inside us for hours. Perhaps this is the faulty side of our brain attempting to strengthen the music. We get bored as too many recurred memories were formed, the ear worm within our head will finally drift away -- oh, wait, this isn't a perfect example, never mind...
The same music can't repeat easily anymore, and can be easily stopped, even after a long time -- well that's what happened to me. Except, certain new situations makes you want to play that music in your head, as it forms new type of memory -- exciting you. I've used this method to keep me awake.
Either way, I'm not sure what's up with you, but the two pictures you drew aren't funny.
-
My transcript explains that after all one's elaborate talk, memories (senses & linked actions) are only selected by either a match, or self-ignition, or if linked to one. The external world input or internal input from a igniting sense like ones just sensed then search memory and match a memory (highest rank as long as similar) and save near it if different. Also: Links allow a car to link to its sound, word vision and its sound, and write and speak actions. Subconscious is all of this processing done before selection. I too keghn called rewards sparks at first. Positive marked actions selected get tweaked.
-
Hi all...
Using random numbers to generate any kind of intelligent operation/ function is very impractical.
This shows a computer randomly generating an short eight digit number (black) and testing to see if it matches the given number in blue. The numbers could represent simple arm/ leg servo positions, etc.
After 1 minute its tried >140,000 match tests (red)... and that's just a eight digit number.
https://www.youtube.com/watch?v=NUV5x-u8ARQ (https://www.youtube.com/watch?v=NUV5x-u8ARQ)
I ran the program for 10 hours without a match before I gave up.
:)
-
What are you suggesting? The only way to find actions for food/win is to generate random actions and rank them by how close they match, then when selected - the highest rank is tweaked, even tweaked in the direction of a created 3D spatial graph cube of what areas look successful, even try all actions and then repeat trying them all.
Also by a match you could write out the exact same numbers~ 34 matches to vision of 34 and linked is hand-write actions.
Anyhow, Freddy, (and friends), so last night I replied back to your reply about "isn't your algorithm just a match machine"? My reply should stun you. It's just above korrelan's video.
-
Maybe I'm a bit jaded, I wasn't stunned.
Some of your ideas are like things I have thought of in the past. Especially, weighting, goals, ranking and things like that.
You use the word 'match' a heck of a lot. It's pattern matching, it's not a new species, sorry.
How old are you again ? Some of us have been here over ten years, we've seen a lot of ideas like this here and on other forums.
Having said that, I like some of your ideas so good luck to you O0
-
I'm 21. I'm trying to say to imagine our stored memories (ex. images/action) in the thin outer layer of 2 cortices, they are onnnnllly used ex. to talk to Freddy if a external or internal INPUT searches&matches a memory OR if a memory self-fires. Or if linked to one. All I say to you is from matches and links, and possibly their matches/links. No other way I talk with you.
-
If i may give my opinion.I think that there is no matching at all, there is no random sensing it is simple action and reaction.How.It is preprogramed interface to some point and form there it is just connecting others new reactions to same or new actions so it is more of physical connection than some energy matching.It is like plant that grows whit all new branches if there is a light.And if there is enough light , branches become interconnected sometimes in crazy ways.To prevent this crazy interconnections the light or signals are sparse, and this is accomplished by action potentials it was explained very well in some of videos i saw here.
Intelligence is not memory matching it is memory creating,as much as one creates adequate memory as much one accomplish further intelligent behavior,again not through matching but through simple reaction within the memory that was connected to some external action.
There is no randomness in intelligence systems it is biochemically,(what is life?) preprogramed.
-
My transcript says we are reactions. We are devices that are not bio-magic. We are destiny particles. Input Searches memory and then selects the match (best ranked as long as similar to input) and does the linked action on external/internal cue which had GOT the RANK for this baby. Memory creating can't happen unless you have rewarded senses and them being selected say by internal input to match or link to what has been saved and saved together, ELSE mud and nonsense would result and no actions for food would be found and no combinations would result either are the 2 points I'm saying in this sentence~..No creating. Random generator even for tweaking selected actions is needed (deaf people speak worse as live).
-
There is no real ranking it is only strengthening of connection by how many time is some reaction occurs.Sensing are going together so memory that emerge from it is complex and it is not ranked by reward,you could learn something even you are not get any reward through simple repetitive behavior.That means that you could be forced to learn something by outside action without getting any internal reward.
Rewards are there to initiate internal reaction, and there is no +- neutral rewards,it is a complex system for maintaining biochemical balance.You could get motivation from rewards whatever you perceive it as bad or good reward, motivation toward reaction is what is counting.Rewards are like some kind of modulators for input signals in that way you could perceive it as ranking system because they allow repetitive behavior toward those input signals that initiate rewards.They amplify action and reaction signals in some way.
Deaf people speak bad exactly because they cant reinforce connection for reaction on audio signals through repetitive behavior of speak.
So repetitive behavior is what create memory and that is old thing repetitio est mater studiorum.
Of course reward system is most important but is most complex too.And rewards system dont need to match memory it is just firing and memory that has strongest connection will excite.
No matching no ranking just preprogramed reward system.And how you preprogramed reward system, that is the same as question what is life.
-
And you can see something really hot but forget about if if not starred at long enough. There's strength and rank.
Deaf peoples actions would stay the same unless tweaked when initiated.
Rewards are preprogrammed, and so is the algorithm, or me making a sandwich, but the algorithm includes cue-matching. If I see a handle the action is matched to, and memories. Ignitions (self-fires) control most of our actions/senses though, so much that my vision is barely external...
Senses that get marked +/-/neutral (by +&- rewards) & ranked, can be linked to neutral low ranked senses to make ANYTHING motivative heheh...AI is like fries, well, it gonna get me eternity of it.
-
To presume that you make AI with this kind of reinforcement learning, but how long would take to get mental picture of out side world like humans have, and how you would say that you will teach it the right way, and what with all that has its senses capabilities because that would make impact on how will learn.
You just scratching the surface of AI understanding, it is not like fries it is very very complex.
-
A human talking about being aware is just a bot making sounds, which match and are understood by the selected actions/senses. And infact are those seconds we have per trillion years each second - selection of action and notices burglar's arm. We're "operatering". It will after it learns, if you don't trust it then you must read and look at my work, you will see for example Links allow a car to link to its sound, word vision and its sound, and write and speak actions. I have combos of things how we do some things that the algorithm allowwwws, therefore it will develop such~
Actually it is Test, Observe ~ Learn, Reason.
Lol humans call this Reasoning & Learning facility.
It's gotta Guess (generate random) motor actions first.