If you fill in the blanks with your own imagination, you are only seeing what you want to see. A reversed numeral 3 is clearly not a normal letter E. A reversed numeral 1 is simply not a normal letter L.
Robots don't care whether text is right-side up or right-side down. (Or they shouldn't.) Clever people don't, either, though it is easier the way we are taught.
So you're saying robots can't see ambigrams? Like they can't yet tell which boxes contain a bus? Good to know... If humans can just keep them away from training sites like this we have more time before the singularity.
Limitations in artificial perception. Inability to extract the signal from the background clutter (differentiate between them). Requires a generalized perceptual construct of "bus" as viewed from all directions. One word "bus" (a concept) relates to a huge perceptual construct.
True. We're leading AI down the garden path with the 2D captcha training. 5D thinking is required to pick that signal up off the noise floor. But 5 is just a few more D and should be well within the robots reach. How long before they see their mistake?
We don't know that much about perception, is the problem. Why is that striped animal a zebra and not a tiger? Is that a zebra or an okapi? What is the distinctive difference? (Answer: the zebra has single hooves, and the okapi has double hooves.) For us, that's easy. But notice that the discriminants are conceptual, not directly perceptual. For us, perception inherently involves a process of organizing the image that we see into a structural hierarchy. Things, and sub-things, and features, colors, and textures. 3-D perception is important to this process.
I did my work in target discrimination from decoy objects. Sometimes there were more objects than real because of overlapping signatures. How to deal with them? My answer was to track all targets, real or virtual. Eventually, the virtual targets would disappear or become intermittent.
If you fill in the blanks with your own imagination, you are only seeing what you want to see. A reversed numeral 3 is clearly not a normal letter E. A reversed numeral 1 is simply not a normal letter L.
Robots don't care whether text is right-side up or right-side down. (Or they shouldn't.) Clever people don't, either, though it is easier the way we are taught.
So you're saying robots can't see ambigrams? Like they can't yet tell which boxes contain a bus? Good to know... If humans can just keep them away from training sites like this we have more time before the singularity.
Limitations in artificial perception. Inability to extract the signal from the background clutter (differentiate between them). Requires a generalized perceptual construct of "bus" as viewed from all directions. One word "bus" (a concept) relates to a huge perceptual construct.
True. We're leading AI down the garden path with the 2D captcha training. 5D thinking is required to pick that signal up off the noise floor. But 5 is just a few more D and should be well within the robots reach. How long before they see their mistake?
We don't know that much about perception, is the problem. Why is that striped animal a zebra and not a tiger? Is that a zebra or an okapi? What is the distinctive difference? (Answer: the zebra has single hooves, and the okapi has double hooves.) For us, that's easy. But notice that the discriminants are conceptual, not directly perceptual. For us, perception inherently involves a process of organizing the image that we see into a structural hierarchy. Things, and sub-things, and features, colors, and textures. 3-D perception is important to this process.
I did my work in target discrimination from decoy objects. Sometimes there were more objects than real because of overlapping signatures. How to deal with them? My answer was to track all targets, real or virtual. Eventually, the virtual targets would disappear or become intermittent.