Advances in techniques have been producing increasingly complex neural recordings, posing significant challenges for data analysis. This thesis discusses novel statistical methods for analyzing high-dimensional neural data. Part one discusses two extensions of state space models tailored to neural data analysis. First, we propose using a flexible count data distribution family in the observation model to faithfully capture over-dispersion and under-dispersion of the neural observations. Second, we incorporate nonlinear observation models into state space models to improve the flexibility of the model and get a more concise representation of the data. For both extensions, novel variational inference techniques are developed for model fitting, and simulated and real experiments show the advantages of our extensions. Part two discusses a fast region of interest (ROI) detection method for large-scale calcium imaging data based on structured matrix factorization. Part three discusses a method for sampling from a maximum entropy distribution with complicated constraints, which is useful for hypothesis testing for neural data analysis and many other applications related to maximum entropy formulation. We conclude the thesis with discussions and future works.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/D8D2240V |
| Date | January 2017 |
| Creators | Gao, Yuanjun |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0019 seconds