Intro Factorized representation

I have been working on represent information that is relevant for the task and information that is irrelevant for the given task in a separate manner since the data is in general noisy for the given task. I am interested in both single modal and multi-modal seniors with different type of data. Papers that I have contributed under this topic include:
C. Zhang, C.H. Ek, A. Damianou, H. Kjellström, Factorized Topic Models.ICLR 2013
C. Zhang and H. Kjellström, How to Supervise Topic Models, ECCV WS on Graphical Models in Computer Vision,2014



Bayesian Inference

Efficient and robust inference method is the key for the model performance. I have been using both Gibbs sampling and variational inference for my research. My current focus is on robust inference methods based on variational inference for big data. I have been recently granted 1 200 000 SEK for 2 years for research on this topic.


PD Diagnostic Prediction from Discomfort Drawing

A discomfort drawing is a drawing on a body picture where a patient may shade all areas of discomfort in preparation for a visit with health care personal. The drawing has been shown to be able to make diagnostic pre- dictions – especially to discern neuropathic from nociceptive and psychiatric diseases. We believe that applying machine learning for diagnostic prediction on patient made discomfort drawings may have significant impact on the health care systems. It may lead to decision support systems that can help health care staff to increase effectiveness and precision in diagnosing and treat patients. Papers that I have contributed under this topic include:
C. Zhang, H. Kjellström and Bo Christer Bertilson, Diagnostic Prediction Using Discomfort Drawing with Multimodal Topic Model;UAI workshop on Machine Leaning for Health, 2016


IROSContextual Modeling for Robotic Vision

I am interested in various machine learning applications.  In robotics, we want the robot not only recognizing  the objects but also be able to manipulate with them. This benefits from contextual modeling and can be performed using a multi-modal generative model. Details can be found in the following paper.
C. Zhang, D. Song and H. Kjellström, Contextual Modeling with Labeled Multi-LDA, IROS 2013



Recommendation System

I worked on recommendation system for groups users during my internship in Microsoft Research Cambridge. Generative model was used for the application. Details can be found in the report MSR-TR-2015-61. After the internship, I still keep an interest in recommendation system in general.