My name is Eric Thomas Heim (commonly cited as Eric Heim), a recent PhD graduate of the Department of Computer Science at the University of Pittsburgh under advisement of Milos Hauskrecht. Broadly speaking, my research interests revolve around efficiently and effectively learning models of similarity and representations from different forms of human supervision. In the past I have been funded by teaching assistantantships, teaching fellowships, the NLM Predoctoral Training Fellowship, and the ASEE SMART Scholarship.
In short, the goal of my research is to create efficient methods that learn a model of similarity between objects from human feedback. Humans can provide unique insight into how objects relate that raw features cannot capture. Human intuition has been used to guide many object recognition, document retrieval, and medical diagnosis techniques, just to name a few. My research aims to effectively capture how humans think about the relationships among objects in a model that intelligent systems can use to perform various tasks. This has multiple inherent challenges:
The bulk of my previous work has focused on learning a nonparametric positive semidefinite kernel matrix from relative comparison feedback. Many machine learning methods are kernelized, thus a kernel matrix can be easily utilized by such kernelized methods. Relative comparisons (which we define as comparisons of the form "Object A is more similar to object B than object C") are intuitive for humans to provide, and can be extracted from many human/computer interactions, thus making them an effective means for obtaining similarity from humans. I have works detailing methods that aim to learn accurate models with little human feedback in an efficient manner.
My current work has two main goals. First, I want to create methods that learn accurate models of similarity among a large number of objects for application to web-scale data. Learning similarity amongst an enormous number of objects poses an obvious challenge due to scale, but also due to the potential variation in such a large collection of objects. Second, I wish to create efficient active learning techniques that can be used in conjunction with methods that learn a model of similarity. By doing so, a similarity model-learning method can make the most out of human feedback, which is often time-consuming and costly to obtain.
|[C15b]||"Sparse Multidimensional Patient Modeling using Auxiliary Confidence Labels" To appear in The Proceedings of the 2015 IEEE International Conference on Bioinformatics and Biomedicine, 2015. [*Extended* arXiv],|
|[C15a]||"Efficient Online Relative Comparison Kernel Learning," Proceedings of the 2015 SIAM International Conference on Data Mining (SDM15), 2015. [link] [arXiv],|
"Relative Comparison Kernel Learning with Auxiliary Kernels," in
Machine Learning and Knowledge Discovery in Databases, volume
Lecture Notes in Computer Science, pages
Springer Berlin Heidelberg,
Presented at the 2014 European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD 2014)