Jaime Shinsuke Ide

Random Generation of Bayesian Networks

This paper presents new methods for generation of random Bayesian networks. Such methods can be used to test inference and learning algorithms for Bayesian networks, and to obtain insights on average properties of such networks. Any method that generates Bayesian networks must first generate directed acyclic graphs (the structure of the network) and then, for the generated graph, conditional probability distributions. No algorithm in the literature currently offers guarantees concerning the distribution of generated Bayesian networks. Using tools from the theory of Markov chains, we propose algorithms that can generate uniformly distributed samples of directed acyclic graphs. We introduce methods for the uniform generation of multi-connected and singly-connected networks for a given number of nodes; constraints on node degree and number of arcs can be easily imposed. After a directed acyclic graphis uniformly generated, the conditional distributions are produced by sampling Dirichlet distributions.

Approximate algorithms for credal networks with binary variables

This paper presents a family of algorithms for approximate inference in credal networks (that is, models based on directed acyclic graphs and set-valued probabilities) that contain only binary variables. Such networks can represent incomplete or vague beliefs, lack of data, and disagreements among experts; they can also encode models based on belief functions and possibilistic measures. All algorithms for approximate inference in this paper rely on exact inferences in credal networks based on polytrees with binary variables, as these inferences have polynomial complexity. We are inspired by approximate algorithms for Bayesian networks; thus the Loopy 2U algorithm resembles Loopy Belief Propagation, while the Iterated Partial Evaluation and Structured Variational 2U algorithms are, respectively, based on Localized Partial Evaluation and variational techniques.

Robust Brain Registration Using Adaptive Probabilistic Atlas

Elastic image registration is widely used to adapt brain images to a common template space, and, in complementary fashion, to adapt an anatomical template to a subjects anatomy. Although HAMMER is a very accurate image-registration algorithm, it requires a 3-class segmentation step prior to registration, and its performance is affected by segmentation quality. We here propose a new framework to improve this algorithms robustness to poor initial segmentation. Our new framework is based on Adaptive Generalized Expectation Maximization (AGEM) for unified segmentation and registration, in which we use an adaptive strategy to incorporate spatial information from a probabilistic atlas to improve segmentation and registration simultaneously. Our experiments using real MR brain images indicate that our integrated approach improves registration accuracy; we have also found that our iterative approach renders HAMMER robust to low tissue contrast, which hinders 3-class segmentation.