Abstract

Machine learning has enabled and improved many real-life applications. However, when humans interact with machines an estimate of model uncertainty is required, especially in applications such as health care. My research focuses on designing probabilistic models and advancing variational inference for this kind of domain.

In this talk, I will mainly focus on two parts of my work. Firstly, I will present probabilistic latent variable models and their applications in health care. I investigate the question of diagnostic prediction from various sources as a multi-view learning problem setting. I design a general framework consisting of a factorized multi-view topic model for image annotation and diagnostic prediction.

Secondly, I present my recent work on accelerating stochastic approximate inference. We employ determinantal point processes for variance reduction and data re-balancing in stochastic gradient methods. Moreover, we propose a unified view of black-box variational inference and importance sampling, and further introduce perturbative variational inference that can have a mass covering effect but at the same time maintain a low variance.

In the end, I conclude the talk with possible future work in terms of both machine learning theory and its application to health care.

Speaker Bio: Cheng Zhang is a postdoctoral research associate in the machine learning group at Disney Research Pittsburgh. She has received her Ph.D. at the Department of Robotics, Perception and Learning (RPL/ former CVAP ), School of Computer Science and Communication, KTH Royal Institute of Technology, Sep, 2016. Her thesis was on Structured Representation Using Latent Variable Models. She has been a postdoctoral researcher in the same group till the end of 2016. She obtained her master’s degree in System Control and Robotics from KTH in fall 2011. During her PhD studies, she visited Prof. Neil Lawrence’ group at the University of Sheffield in 2013 and was an intern at Microsoft Research Cambridge in the Infer.Net group in summer 2014. She has received two grants from Stiftelsen Promobilia (a Swedish research foundation) for research on robust inference for computer vision tasks where she is the PI. Her joint project with Karolinska Institutet (KI) on patient-centered decision support is funded by a Vinnova (Sweden’s Innovation Agency) UDI step 1 grant. She is active in the areas of machine learning and computer vision, and with a strong interest in machine learning in health care.