@phdthesis{oai:oist.repo.nii.ac.jp:00000891,
 author = {Schulze, Jessica Verena},
 month = {2019-05-21, 2019-05-21},
 note = {Previous studies of effective connectivity inference from neural activity data benefited from simple regularization approaches such as L1 regularization, which promotes sparseness of the connection matrix. In this thesis we investigate the incorporation of two novel physiologically plausible priors based on spatial and modular organization of the neural circuit in the framework of Bayesian inference. First we formulate a spatial prior which incorporates distance-dependent connectivity in the linear non-linear Poisson (LNP) model. We consider distance-dependent L1 and L2 regularization of connection weights as well as a hierarchical prior with distance-dependent connection probability. We derive maximum a posteriori (MAP) estimation algorithms by gradient descent, Newton method, and Metropolis-Hastings sampling. We test the effectiveness of these algorithms using synthetic data based on physiologically realistic distance-dependent connection weights and clarify the effects of the regularization parameter and data size, as well as the problems with highly synchronous firing and self-connections. The methods are also tested with calcium imaging data from the mouse posterior parietal cortex (PPC). Next we formulate a modularity prior which assumes multiple modules in a network and different within-module and between-module weight distributions. We formulate a MAP inference by combining Gibbs sampling for module membership and Newton’s method for connection weights. The method is validated by synthetic data with various modular structures, including spatially localized modules with distance-dependent connections.},
 school = {Okinawa Institute of Science and Technology Graduate University},
 title = {神経活動データからの回路結合推定における位置情報とモジュール性に基づく制約},
 year = {}
}