 |
|
|
 |
|
|
|
|
Mark
[Guest] |
| | | | | | posted 9/24/2007 00:28 |  | | | | | | | | | I haven't seen any major discussions yet on AI that uses error bounding for its thinking processes. There are several tremendous advantages to error bounded AI. An example would be face recognition.
Instead of remembering everything exactly, the recognition of a face, could be as simple as bounding or contrasting specific areas of a human like face. The bounded areas of the face indicate where eyes or the nose or whatever, should be in there, but it doesn't exactly say where in the area. So, an error is allowed. In other words, you get a face map that consists of bound areas, that can be placed over any face. The system will recognize any arbitrary object as a face, if the face map corresponds to the object's appearance.
Another example would be memory. If you construct an error correction system that uses a local time error, and then tries to adapt the system to fade out the error. In other words, every time the system is adapted, it carries on the information of the previous errors. That way, the system could gain a sense of logic, due to overlapping input with memorized errors.
Another example is decision making. Bounding the range of decisions an AI can make by using a certain memory-based error function, and finding the optimum, allows the AI to actually chose. While at the same time, it maintains the ability to chose something else, something arbitrary, that may be wrong, or may be tremendously right.
There is an interesting philosophy behind Error Bounded AI. It assumes that the AI can never obtain an exact abstraction of anything, only the absolute minimum of the error range.
My personal interest with AI, is that I am trying to understand how humans learn and how to adapt that to AI.
One of the more interesting revelations I had, was recalling a story about a girl who at young age, was abandoned in some African Jungle, but survived purely instinctively. When she was rediscovered, she lost all her abilities of being human. She adapted to her environment, and learnt what was necessary, but behaved more like a monkey than a human.
If you look at society, and how we learn, we learn by the passing of understanding that we do something wrong. But in general sense, we don't always understand what we are doing wrong, just that we know that we are doing something wrong and that we have to change to in the end do things that will cancel that. In other words, we don't always understand that we do something wrong, but we were simply told that we were doing it wrong. And we listen and obey, because the ones that tell us that we are doing it wrong, provide us with means of survival. We evolved to listen and it's what makes humans intelligent. Therefor, I firmly believe in the Error Bounded AI approach, because it's the only one that allows to make perfect logical sense, in scientific and mathematical terms. It also allows for more firm radicalization of common perceptions of human thinking, without falling into the endless debate black holes like about what consciousness is.
| | |
| | |
hunt |
| | | | | | posted 3/12/2008 20:40 |  | | | | | | | | | I think error bounding is necessary for so many AI applications. I'm currently working on a parser for my own bot that uses a Bayesian probability scheme to fit sentences to grammars that it has seen before. When it sees a similar grammar, what is the threshhold for using it as a framework for parsing the new sentence?
Error bounding is such a subjective thing in practice though. How do you define to what degree a partial grammar is "like" a known grammar? Do you weight certain matching components, or the relative position of matching components? To what degree?
The same with facial recognition, I imagine. How do you set the bounds? You could, I suppose, take an active teacher approach and have your algorithm learn what probabilities often correspond to the correct answer, using active user feedback. But this can be time consuming, and not enough if your original probability calculating algorithm is insufficient.
| | |
|
|
|
|
|
