amsler@FLASH.BELLCORE.COM (Robert Amsler) (07/29/87)
Another division of information which I find significant is that of visual vs. the combined structural and functional descriptions. While a visual description might be termed `structural' I think there is a significant difference. Visual information, i.e. information obtained from looking at a visual still or moving image of an object, is often not available in pre-recorded structural form. It `may' be possible to describe visual information in symbolic text, but it can prove very hard to extract it from existing descriptions because there is so much visual information to represent and often the description doesn't contain the key element needed to answer a question. I first encountered this when looking at the information dictionaries present for a word such as `horse'. They give definitions of all the parts of a horse, but you cannot assemble a horse from these part definitions accurately enough to answer a simple question such as whether the horse's head is higher than its tail? (Dictionaries almost universally have an illustration for a horse, which suggests they know something about how hard it is to describe one by definitions only). Initially I saw this as demonstrating the complimentarity of visual and definitional information, much in the same manner that Minsky sees the complimentarity of the structural and functional descriptions. But now, it looks to be a more basic problem. Even if you could assemble a horse from the definition plus the static visual knowledge (e.g. add coordinates and a wire frame model of a horse to the description), I can't animate it well enough to answer questions (Are all the feet ever off the ground simultaneously while running?) This probably suggests a simulation as the correct representation, but often a simulation is really just a means of displaying the visual representation of the object so you can perform the observation needed on the simulated entity rather than on the real entity. What this seems to imply is that ultimately the `description' of an object should be a simulation accurate enough to permit direct observation and generation of the functional and structural information we know about the object?