25.1.10
The student has become the teacher
The copy of a copy activity we did in class was fun and insightful. During the in class discussion and afterward I thought about how that activity could apply to what we are trying to do with DARCI. When we make photocopies of photocopies we loose quality each time. However, as discussed in class, this change isn't always bad, in fact in may be very good. It reminds me of a teacher to student relationship. When a student learns from a teacher it can be thought of as copying information from the teacher to the student. However, the student can understand and interpret the information differently and the information is then changed a little. That student can then teach another student and the information is then changed again. Teaching DARCI adjectives seems no different than this. We as the teachers have our own understanding of these adjectives and how they are associated with images. DARCI, however, will develop her own understanding and interpretation of these adjectives that will be a little different from our own. These differences could provide an argument that DARCI is not just an overall reflection of her teachers, but adds something of her own. I also think it would be very interesting to have DARCI teach another copy of DARCI and compare them.
Subscribe to:
Post Comments (Atom)
Having DARCI teach another DARCI would be really cool to try. And then what if the DARCI 2 taught DARCI 3 and so on. I speculate off the top of my head that one of two things would happen with the current implementation of DARCI. One, eventually the DARCI’s would stop changing significantly with each new iteration. This could happen because each iteration might be like taking the derivative of a derivative. Eventually you’re going to get a slope of zero, then no more changing.
ReplyDeleteThe other possibility is that the DARCI’s would become essentially random in behavior. This would happen because the neural net outputs might gradually dampen to approach 0.5, a completely neutral response. Then, the labeling decision would be random because tiny unpredictable fluctuations could throw DARCI’s output to either side of the threshold of 0.5. If we had a neutral band of output values, say between 0.4 and 0.6, then this instead would mean that DARCI X would eventually be unsure of everything.
The only way to know for sure what would happen, would be to actually test iterative DARCI’s. Either of the results I predict would not correlate with what would happen with humans. Humans, I would think, would always be different and never be purely random—no matter how many iterations occurred. They would probably become more and more “wrong” with each iteration (think the telephone game), but they would be confident with their knowledge. We should think about how to design DARCI so that this experiment would yield similar results to humans. This is all very speculative of course.