Multimodal matrix and tensor factorization methods exploiting strong structural and side information
Abstract
We propose a new multi-scale imaging approach and new analytical methods to enable a new era of voltage imagingwith single-cell res"olution in large neuronal populations. The idea is to simultaneously perform, in the same neuronal population: (1) calcium imaging a"t low temporal resolution but high spatial resolution and (2) voltage imaging at hightemporal resolution but low spatial resolution". Video stream (1) provides information about where (and roughly when)spiking events occurred, and stream (2) provides more highly" detailed constraints about when the events occurred. Wewill develop statistical methods to fuse these two sources of data to obtai"n the best of both worlds, i.e., highlyspatiotemporally resolved estimates of neural activity, with imaging and computational metho"ds that are scalable tolarge neuronal populations. To accomplish our analytical goals we will develop novel matrix and tensor facto"rizationmethods exploiting the strong structural and side information inherent in this experimental data. In addition, we will also"pursue generalizations of these methods to applications outside neuroscience where strong but non-standard matrix ortensor structure can be exploited.Aim 1. Joint factorization for multimodal data fusion. We will develop algorithms and theory for jointly facto"rizingmultiple data matrices with different scales and signal types (e.g., voltage and calcium measurements) into estimatesof the" underlying shared signal structure with high spatial and temporal resolution.Aim 2. Amortized inference methods for incorporating stronger temporal prior information. We will exploit priorknowledge about the underlying structure of these temporal signals to develop improved fast approximate Bayesianmethods for recovering these signals.Aim 3. Exploiting multi-trial tensor structure for super-resolution and improved recovery. We will further exploithierarchical repeatable structure in these data matrices to estimate t"he spatial and temporal signal components withhigher resolution than available in the original data.Aim 4. Bayesian decoding, nonl""inear dimension reduction, network estimation, and optimal control given multimodal observations. We will extend previous models and"" methods for statistical analysis and control of point process data to exploit multimodal observations (e.g., incorporating not just" neuronal spike train observations but also subthreshold voltage data) to obtain improved estimates and enable new closed-loop experiments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 01, 2017
- Source ID
- N000141712843
Entities
People
- Liam Paninski
Organizations
- Office of Naval Research
- Trustees of Columbia University in the City of New York
- United States Navy