Speaker: Sunsern Cheamanunkul
Title: Co-adaptation for handwriting recognition
Human computer interaction tasks, such as handwriting recognition, pose
an interesting and challenging machine learning problem. There is no
single correct way to write a character. As a result, in addition to
the possibility of having the machine adapt to the handwriting of the
user, there is a viable possibility of having the user adapt their
handwriting so that it is easier to recognize.
In this talk, I will present a framework for handwriting recognition
that considers the writing trajectory as a continuous-time
communication channel with feedback. As the user is writing, the
recognition algorithm constantly re-evaluates the likelihoods of
different letters. The recognition system provides the user with
real-time feedback to help the user identify mistakes early. In
addition, offline analysis is used to guide the user how to make
adjustments to their writing that will improve the recognition accuracy.