Abstract:
The talk will focus on connectivity inference with asynchronously
updated Ising models. We have developed two sets of learning rules for
the couplings with the tools of statistical physics. They are
approximate and exact versions respectively. The approximate ones are
based on different levels of mean-field approximations, while the exact
ones are based on maximization of the log-likelihood of the data set.
Both of them produce densely connected networks. However, the weak and
spurious links can be eliminated by regularization in a controllable
way. The developed algorithms are firstly tested on synthetic data, and
then applied to real recordered data from experiments.
About the speaker:
Hongli Zeng defended her doctoral dissertation at the Department of
Applied Physics of Aalto University on 15 August 2014, and will receive
her PhD degree this September. She is also a member of the COMP Center
of Excellence at the Applied Physics Department and the COIN Center of
Excellence at the Information and Computer Science Department. Her
research interests include Machine Learning, Pattern or Data Analysis,
Computational Neuroscience, Regularization, etc.
Last updated on 2 Sep 2014 by Antti Ukkonen - Page created on 2 Sep 2014 by Antti Ukkonen