So the method of initial recognition that I am going to try with my (very early version of) image analysis will be to use multiple analysis algorithms on an image. Then I am going to use the results to output some kind of result that I choose to study. (The first result is going to be a modification of the original image which can be used for some purpose). Then I can select parts of the results to 'reinforce' as cases where the program achieved the desired goal. Then the test will be to see if the program can guess which parts should be analyzed by which algorithms to produce the desired goal. My, point however, is that it is not that simple. It won't work as described because initial image analysis is not strong enough to produce the desired results even when there is exhaustive training. I can predict pretty confidently that the initial analysis will need to taken through a kind of 'perceptual' stage where some higher level ideas about the image has to be projected onto the initial results of the analysis and compared to see if the analysis makes sense based on previous learning. Often times this comparison of the projection of high level insight onto the configuration of results of the initial analysis will not fit perfectly and some 'stretching' of the relations will need to be made.
One other thing. Although I think Mike Tintner's commentary has not very useful, I do feel that he has helped me to discover something important in the past. He tries to point out that there is no program that can produce every set of actions or algorithms. That seems reasonable. But most of us feel that he is missing the point. We can write programs that can learn new things without being programmed to be able to produce every possible variation. An AGI program can learn by interacting intelligently with the IO data world. However, when I once reacted to one of Mike's challenges and I tried to figure out a way to write an algorithm that could produce obvious variations of a typological character without being intelligent or reacting to input I was able to derive an insight about how such a thing might work. And I could use a variation of that idea on my image analyses (including analysis through modification) methods. So if I was able to get somewhere it is theoretically feasible to design a single algorithm that could create an immense number of image modification algorithms. The one thing that is left is to harness this process so that a useful analysis or modification algorithm could be chosen -without running them all- on the basis that it might produce a result that is required to produce an insightful result. By examining how the variations on the (single) image analysis super algorithm affects the result, it might be possible for the program to learn to 'project' families of characteristics of those variations and hence make educated guesses that resemble some of ours that go like, "If I could find this and that kind of thing and test it to see if it could be related to finding a solution then I might be able to get closer to figuring this out. Characteristic A and characteristic B are similar to the this and that therefore by trying variations on the super algorithm I might be able to detect the this and that event." Jim Bromer On Tue, Feb 5, 2013 at 9:07 AM, Jim Bromer <[email protected]> wrote: > Oh I forgot. The initial recognition dilemma is this. All good AI methods > work well in some situations but in others they don't work very well at > all. The problem is that the situations in which they work well do not > cover entire situations. So, even in a single particular situation, like a > scene in visual recognition, the AI algorithm might be spectacular with > some 'objects' of the scene but fail with others. Now if a programmer was > looking at the scene he can decide to use different algorithms that work > well with different parts of the scene and by this method get good initial > recognition coverage. But an automated system which does not have good > initial recognition (or understanding) of the scene is not going to be able > to choose which algorithms it should use for recognition. This assumes that > the recognition algorithms will both fail to recognize some parts and give > some false identifications of other parts. > So the programmer sees that there are some algorithms which work really > well, and he could select different algorithms to work on different parts > of a particular input object, but when you try to automate this the problem > becomes a dilemma. How can a program choose the right algorithm to evaluate > a part of a scene without first recognizing what parts cannot be identified > and what algorithms would be best for those parts. > Jim Bromer > ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
