I'm curious about the idea that was alluded to in the Rosch paper of the importance of functional attributes in categorization. While we can certainly use solely visual cues in the categorization task, there is not necessarily any single set of shape or visual cues that defines an object as belonging to a particular category.
Instead, it seems likely that there are a whole host of other ideas about an object's use and our experience/memories with it that go into categorization. With a sufficiently large set of memories and associations, categorization can be as simple as a nearest-neighbor to some other object we have dealt with in our lives, then related to a category by function. Also, there may be complex semantic associations between certain shape features and possible functional use that feed into categorization in a significant way.
Traditional categorization has attracted vision researchers to language processing. Now I wonder, if we start considering a different model for categorizing what cameras see, then what other areas of knowledge would become critical? In other words, what else do we need to study to understand (for example) how to categorize based on functionality? Is machine learning enough?
As Peter said, experience seems very important. But experience for us is not just about seeing things and learning how they are named. There is touch, for example, that may help us understand objects resistance and might constraint their use. Without extending too much, my point here is that I feel like vision is going beyond just processing images. In the case of categorization, for example, it seems like there is still more to learn from other areas of research to make it work.
I'm curious about the idea that was alluded to in the Rosch paper of the importance of functional attributes in categorization. While we can certainly use solely visual cues in the categorization task, there is not necessarily any single set of shape or visual cues that defines an object as belonging to a particular category.
ReplyDeleteInstead, it seems likely that there are a whole host of other ideas about an object's use and our experience/memories with it that go into categorization. With a sufficiently large set of memories and associations, categorization can be as simple as a nearest-neighbor to some other object we have dealt with in our lives, then related to a category by function. Also, there may be complex semantic associations between certain shape features and possible functional use that feed into categorization in a significant way.
Traditional categorization has attracted vision researchers to language processing. Now I wonder, if we start considering a different model for categorizing what cameras see, then what other areas of knowledge would become critical? In other words, what else do we need to study to understand (for example) how to categorize based on functionality? Is machine learning enough?
ReplyDeleteAs Peter said, experience seems very important. But experience for us is not just about seeing things and learning how they are named. There is touch, for example, that may help us understand objects resistance and might constraint their use. Without extending too much, my point here is that I feel like vision is going beyond just processing images. In the case of categorization, for example, it seems like there is still more to learn from other areas of research to make it work.