Authors: Chethan Pandarinath, Daniel J. OShea, Jasmine Collins, Rafal Jozefowicz, Sergey D. Stavisky, Jonathan C. Kao, et al. (Stanford University, Emory, Google Brain) · arXiv: 1608.06315
Models single-trial cortical spike trains as Poisson emissions driven by a nonlinear recurrent dynamical generator, inferring unmeasured external inputs and initial states via variational autoencoding.
Key Quantitative & Theoretical Takeaway: Decomposing single-trial spike variability into deterministic low-dimensional dynamical trajectories and inferred controller inputs predicts behavioral kinematics on millisecond timescales without trial averaging.
In neuroscience, a central goal is to understand how populations of neurons coordinate to generate behavior. We introduce Latent Factor Analysis via Dynamical Systems (LFADS), a deep learning method to infer latent dynamics from single-trial neural spiking data.
A bidirectional encoder infers the initial state g_0 of a recurrent generator RNN, while a separate controller network infers time-varying inputs u_t penalized by an autoregressive KL prior, isolating intrinsic dynamics from unexpected perturbations.
Replace arxiv.org with nujan.app on any arXiv URL (for example, https://nujan.app/abs/2006.11239) to open the PDF and AI research partner side-by-side.