Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Explore a 12-minute conference talk introducing StateSpaceDynamics.jl, a new Julia package designed for probabilistic state-space modeling in computational neuroscience. Learn how this package addresses critical limitations in existing Julia libraries by supporting non-Gaussian observation models essential for analyzing neural spike count data that follows Poisson or discrete distributions. Discover the implementation of direct maximization of complete-data log-likelihood for Linear Dynamical Systems (LDS) models, which leverages block-tridiagonal Hessian matrix structure to compute the Kalman smoother in O(T) time. Understand how the package extends beyond Gaussian models to accommodate Poisson and Bernoulli observation noise models, making it suitable for modern neuroscience applications. Examine the variational expectation maximization method implementation that enables learning of hierarchical state-space models, including switching linear dynamical systems (sLDS) used in studying neural dynamics related to movement preparation and execution. Review the package's discrete state space model capabilities, including Hidden Markov Models (HMMs) fitted using expectation maximization algorithms, which support the development of more complex hierarchical models. Gain insights into how this package fills a crucial gap in the Julia ecosystem by providing computational neuroscientists with advanced tools for modeling high-dimensional neural recordings through dimensionality reduction combined with temporal dynamics modeling.
