 |
 |
|
|
|
|
shakir1311 |
| | | | | | posted 1/6/2012 21:53 |   | | | | | | | | | Image 1:
http://i42.tinypic.com/fuqb9u.jpg
Image 2:
http://i40.tinypic.com/nw3t0.jpg
Problem: If a human looks at Image 1 then we can instantly identify that the Bulb is not Lit and the the Light being shown on the Bulb Holder is from a source other than the bulb itself.
Image 2 makes this vision based perception easier for Humans.
Now what AI algo OR Type of Vision Percepttion can answer this question:
- Take Image 1 as input and answer the question:
- Is the bulb On?
- Take Image 2 as input and answer the question:
- Is the bulb on?
Technically if the AI can answer the question with Image 1 then it should be able to answer the question with Image 2.
We can also create an ALGO or kind of AI which takes both images as input and tells us whether the bulb is on or not.
| | |
|
|
|
Vandrian |
| | | | | | posted 1/7/2012 07:47 | | | | | | | | | |
Not sure if I am answering your question (sincer there doesn't seem to be a question???
I think this question requires two pre known things
- the ability to identify bulbs
- to know what is meant by the light being on or off.
You can model shade on a surface with RGB adjustments I did this a few times in photoshop to see if different colors had a similar change function. Kind of like dampening a piano with the soft pedal. I mean if you have tungsten lights then its going to be different than flourescent lights, but the shift is going to be the same in histogram of objects.
Most of the time is spent on identifying the objects in the image. Whether the light is on is pretty easy if you can create an accurate model of the scene. Doing that can take several different directions on how general or specific you make the system.
I included a link that looks like they really complicated the whole thing :)
| | |  http://en.wikipedia.org/wiki/Phong_reflection_model | | |
|
|
|
shakir1311 |
| | | | | | posted 1/7/2012 08:01 | | | | | | | | | | Ok. Thats great. So it means we need to have a Model of the scene.
But this is ordinary AI practice. What my mind is telling me:
- Maybe we are not approaching AI in the best possible way.
Now consider this:
Person A sees these two images (posted in the actual question). And you ask the person: Is the bulb on. The person will answer: No
Ok.
Now take a computer and provide these two images and ask the question: Is the bulb on? The computer may get help from modelling and maybe ultimately (but with a greater delay than a human) it can answer NO.
Now what I am suggesting is this:
- Lets forget about models
- Lets forget about computer vision
- Lets not go into RGB details of the image
- Lets find a way where we can create a computer who is as Intelligent as a Human (in solving this single problem).
Now the question is:
- Suppose a boy is born on Mars (where there is no electricity). He is 30 years old and then you send these images to him and ask him the same question. He wont be able to answer it. Because he never learnt it.
We should be (after around 50 years into applied AI) able to device a mechanism where we actually make machine learn (just like kids learn).
Ok. Now lets approach this light bulb issue in my way:
- Find out all the input a 30 years old human brain will utilize to answer this single question. (is the bulb lit or is it sunlight on it).
- Create a computer porgram who can parse all that input (and the computer will take no more than 2 days to consume all that data).
- The program must follow the same learning pattern as a human brain does in solving this specific problem.
And bingo..... We have a computer program which is as good as a human barin (for this single question). Now consider the limitless intelligence we can achieve this way.
Infact I am not good at AI. But sometimes i do realize that the present AI practices are lacking something. Maybe the discussion above is just a dumb idea. But who knows. Maybe that is the ultimate answer to all AI problems.
| | |
|
|
|
Vandrian |
| | | | | | posted 1/7/2012 10:01 | | | | | | | | | | ug...
my favorite part was the "bingo" I really wish we were at "bingo"
It's like seeing the other side of the lake and saying we will just follow the coast and "bingo" we will be over there. We have no idea how long or where this coast goes, we can guess sometimes, but the people getting the farthest are the ones that actually enjoy the coast rather than point at the other side of the lake.
We can see a light bulb we can't see hoe we see light bulbs. This is the exploring part and its fun :)
I think I agree with you about having general heurisitcs rather than just a simple specialized system that says light bulb or no light bulb but who knows what it will say when it sees a duck.
| | |
|
|
|
shakir1311 |
| | | | | | posted 1/7/2012 10:31 | | | | | | | | | | Well. First of all. I really appreciate you talking to me about this. And secondly please dont mind me if i sound stupid. Sometimes stupid ideas change the world.
Well for the duck here is what we will do:
- Copy all information from a human brain that is relevant to identifying the duck. This info includes images of ducks, sounds of ducks, people (mom and dad) saying "Hey look thats a duck. Quack quack quack"
- Train the program the same way human has trained in 30 years (with much faster speeds than human)
- And there you go
All we need to do is:
- Find a way of gathering each and every bit of info a brain has (about a particular domain or topic or object of the world)
- Find the way how brain process that info over time and converts that to lifelong knowledge and make a program which can do the same training with the same info provided to it
Once we have such a working model then we can train it on limitless things. Remember once we achieve this then we get rid of the human problems (i.e ageing, loss of memory, limited memory)
In other words lets say: We want to create a AI program which has intelligence of a 1 day old kid. But it has the same training features as the kid. Then put that program into using all the info a kid recieves in his life time.
Why shouldn't this approach work. and what name should we give to it? Maybe some work is already in progress based on these lines. I heard of some Japanese project where they intend to record a 30 years life span of a kid.
Do i sound interesting or just dumb :(
| | |
|
|
|
tkorrovi |
| | | | | | posted 1/8/2012 03:12 | | | | | | | | | | | |
shakir1311 wrote @ 1/7/2012 10:31:00 AM:
- Copy all information from a human brain that is relevant to identifying the duck. This info includes images of ducks, sounds of ducks, people (mom and dad) saying "Hey look thats a duck. Quack quack quack"
- Train the program the same way human has trained in 30 years (with much faster speeds than human)
- And there you go
| |
I don't see how this makes sense. And about copying the brain there are many problems which make the exact copying impossible. There may be no distinct separation between one information and the other, without information somewhere else the information may even make no sense, the information is fastly changing so it may not be possible to take a snapshot of the system, the system copied should start to work in another environment so the consistency of the system would be lost.
This light bulb thing reminds me one joke. A long time ago the police cars had a light on the roof which blinked. So one policeman said to the other policeman, see whether the light on the car roof is on. The other policeman stood outside the car and said "Light is on, light is off, light is on, light is off, light is on, light is off".
| | | Artificial Consciousness ADS-AC project | | | Last edited by tkorrovi @ 1/8/2012 3:25:00 AM | | |
|
|
|
shakir1311 |
| | | | | | posted 1/9/2012 23:13 | | | | | | | | | | Ok.
So we have a new definition for AI:
- Try to imitate human intelligence with only "Well known" techniques and with using "Over the counter" tools and methods.
Come on. I am talking about creating replicas of human intelligence.
What has AI done so far:
- In 1950 we thought we would be having robots as servants at home by 2000
- In 1950 we realized fuzzy logic can actually wash dishes just like a maid does
- We also assumed we would be having a software by year 2000 which could actually talk (as in conversation)
But what we have:
- We do not even have a close to human OCR
- We dont even have Real Voice controlled systems which can take commands from any person in this world
- Even AI gets beaten at chess
So what is that?
Do we have a major flaw in the basic principles we have established so far?
If yes then lets find out. Lets think different. Lets not call it Artificial Intelligence. Lets call it Intelligence.
What I see the problem is:
- We always try to have mathematical models
Why dont we invent Chemical Models for solving problems.
Does our brain always make decisions based on algorithms?
Maybe it uses a Chemical method for distinguishing a Light Bulb's Light from Sun's Light.
| | |
|
|
|
tkorrovi |
| | | | | | posted 1/10/2012 01:08 | | | | | | | | | | | |
shakir1311 wrote @ 1/9/2012 11:13:00 PM:
So what is that?
| |
If my answer is good enough for you because i only deal with True AI. This is a part of these "same old things" which i have been talking here for many years, yet again and again people come who have never read them, so i have to tell them again and again. If your question was not about True AI, then i'm really sorry and you may ignore my answer.
Why i think not much is achieved in more than 50 years is that one field is completely not researched. And this is unrestricted self-developing systems. And the primary reason why this field is not researched at all is that no one is interested. Like my project below, many people have said that they are interested but they are not interested to the extent that they would support it, not to talk about helping in development and research, which finally still amounts to not interested. The other reasons are that the whole AI field is defined as an "applied science", but especially in unrestricted systems it is not possible to go ahead without research, and the third reason is that AI has meant to be like a showcase of behaviorism, but research of extremely flexible and changing systems does not correspond to that philosophy.
What concerns AI other than True AI, then for theoretical reasons it can only succeed when it is made for some restricted task because every conventional AI is inherently at least somehow restricted. Like artificial neural networks are restricted in that they can model statical events but not processes. My joke about policemen was also partly about that.
| | | Artificial Consciousness ADS-AC project | | |
|
|
|
shakir1311 |
| | | | | | posted 1/10/2012 09:29 | | | | | | | | | | | |
Why i think not much is achieved in more than 50 years is that one field is completely not researched. And this is unrestricted self-developing systems
| |
Yes. There are the words I was starving for. We need to have unrestricted self-developing systems which are independent of domain restrictions (or whose domain = whole world).
Thanks for showing me the right path.
| | |
|
|
