18.2.10
Associations
I was so fascinated by the associations that we made. Dan said he didn't think it was a big deal that we made "incorrect" (insane, irrational...) associations, but does DARCI make associations that are incorrect? Does anything she do come out of random associations or is all of her information supplied by us? What happens if the associations she makes are incorrect? Maybe nothing. Maybe it doesn't matter.
Subscribe to:
Post Comments (Atom)
DARCI is learning what combinations of features are associated with particular adjectives. These features are numeric measures of the image. In theory these features could be the RGB values of every pixel in the image. This would unfortunately yield a prohibitively large number of features. For this reason, we select more general features that have been proven in computer vision. These are features like image brightness, color counts, amount of noise, eccentricity (how circular the image is), how many repeating patterns there are, etc. There are 102 features that DARCI looks at right now.
ReplyDeleteWhen you label an image for DARCI, then you are telling her that the exact feature combination of that image correlates with the label you assigned. Using a neural network, she learns to extrapolate other images that will fit that label. The more examples she sees of a given label, the more accurate her extrapolation will be. To put it another way, DARCI is trying to build a general rule for what it means for an image to be a given adjective. The more examples she sees of the given adjective, the more accurate her rule will be.
So to answer your questions, all of the information she has is supplied by you guys; but, at the same time she’s filling in the blanks and generalizing on her own (i.e. using neural networks). If she is given “incorrect” associations by you guys, then those associations will indeed affect her generalization. But so what? If DARCI’s interpretation of what is scary is completely different from what most people think, then is she wrong? She would only come to that generalization if her trainers provided her with such examples. Thus there is a strong correlation between DARCI and her trainers. This is why Joe’s idea of using DARCI to study different cultures is so intriguing.
The associations being created are not always expected. They may not be the prophesied outcomes. But that does not make them incorrect. Each association has been built upon an individuals experiences. So for that individual the association is alive and correct. Even if it is not the association that was predicted.
ReplyDelete