When learning works better than machine learning: Recovering damaged QR-codes with manual choice of image features to recognize

Speaker: Michael Raskin, Department of Computer Science, Aarhus University

Time, Friday, December 2nd, 14:00-15:00

Place: Peter Bøgh Andersen auditorium

Home page:

http://mccme.ru/~raskin

Abstract:

This talk decribes a real-world situation where a large number of damaged QR-codes needed to be recognized.  I explain why machine learning is not always a correct choice in such a situation.  The talk covers the process of learning about the details of the input data which can lead to quite a simple algorithm.  This algorithm turned out to be sufficient in practice.  I also highlight the ways in which the real world is `noisy'.

This noisiness and its interaction with the software are something one should consider when making decisions.

Biosketch:

Michael Raskin has obtained his MSc and PhD degrees at Moscow State University (Faculty of Mechanics and Mathematics, Department of Mathematical Logic and Theory of Algorithms).  He also has some applied experience developing software systems, including, but not limited to, education-related workflow support systems.  Currently he is a postdoc here at the Department of Computer Science.