Research Abstract
One-shot recognition refers to the special case of learning to classify an
object class when we are given just a single image to learn from. There are
numerous difficulties to this problem, as noise can manifest itself in the
system from many different aspects. For example, if we're given a very
unrepresentative image of a dog, it would be unreasonable to expect to be able
to identify other dogs based on just that one image.
This summer, we tackled the problem of one-shot recognition by utilizing the
idea of transfer learning. We formulate the transfer learning in two different
ways, both of which learn feature transformations to a different space, which we
term as attribute spaces. Our results show that both methods seem to learn
representative attribute spaces, as we show competitive one-shot recognition
results against a previous benchmark that utilizes a human-defined attribute
space.