Machine Learning When Training and Test Inputs Have Different Distributions

Abstract

Systems based on machine learning methods often suffer a major challenge when applied to the real-world datasets. The conditions under which the system was developed will differ from those in which we use the system. Few sophisticated examples could be email spam filtering, stock prediction, health diagnostic, and brain-computer interface (BCI) systems, that took a few years to develop. Will this system be usable, or will it need to be adapted because the distribution has changed since the system was first built? Apparently, any form of real-world data analysis is cursed with such problems, which arise for reasons varying from the sample selection bias or operating in non-stationary environments. This talk will focus on the issues of dataset shifts (e.g. covariate shift, prior-probability shift, and concept shift) in machine learning and managing to learn a satisfactory model.

Biography

Dr. Haider Raza is a post-doctoral researcher fellow at IADS (University of Essex) with a focus on learning methods for analyzing non-stationary data their applications in BCI, healthcare, and finance. Previously, he was post-doctoral research officer with the Farr Institute of health Informatics Research, Swansea University Medical School, U.K. He was advised by Dr. Girijesh Prasad and Dr. Hubert Cecotti to obtain his Ph.D. at Ulster University, UK in 2016 focused on adaptive learning for modeling non-stationarity in EEG-based BCI systems.

Tea/coffee will be provided.