I work in information theoretic machine learning with Dr. Susanne Still. My interests are in predictive inference and evolutionary learning in robots and nature.
The goal of my PhD research is to gain a greater understanding of the role of information and prediction in self-organizing systems through information-theoretic machine learning methods.
I am developing algorithms for passive and active predictive model-building based
on related data-clustering algorithms and am investigating automatic parameter selection for parametric approximations.
Living organisms have evolved to predict future conditions by gathering and processing information while acting on the world. Information theory provides a quantitative basis for investigating the role of information in the emergence and evolution of complex systems such as biological life.
Information-theoretic machine learning methods operate on simple principles. Little prior knowledge of the input data is required and very few assumptions are made about the nature of the model being learned. Existing information-theoretic algorithms for learning predictive models with limited memory size, such as in living creatures, are optimal in an ideal case.
However, the most promising algorithms for building predictive models, with or without feedback from an acting agent, suffer from computational complexity issues. This limits our ability to apply these methods on real data. These issues also indicate that it may be natural to build approximate models from complex data if resources are limited or speed is a factor.
I am working on new algorithms to make more realistic applications and experiments possible. I plan to use them experimentally to build and evaluate predictive models of real time-series data and to examine ways in which information drives self-organization using real and simulated agents.
I expect these experiments to reveal limitations to existing approaches and will seek to develop improvements. The results of this work will be useful in several areas including time-series prediction of real-world data, learning methods for adaptable robotics, and may even shed light on the role of information in the rise and evolution of biological life.