A Multiscale Dynamical Modeling and Identification Framework for Spike-Field Activity

Multiscale low-dimensional motor cortical state dynamics predict naturalistic reach-and-grasp behavior

“…2a and Methods). This algorithm modeled the spikes as point processes with a 10 ms time-scale and LFP activity as linear Gaussian processes with a 50 ms time-scale 42 and learned low-dimensional latent states to describe the combined spike-LFP dynamics. We also learned dynamical models for the spiking network activity alone and the LFP network activity alone using EM algorithms for point process 46 and linear Gaussian models 47 , respectively ( Fig.…”

“…For each principal mode, in addition to the decay-frequency pair, we quantify how well the mode predicts behavior and neural activity (Supplementary Note 1 and Methods). Similar to prior work, we model the behavior, defined as joint angle trajectories or 3D end-point hand kinematics (positions and velocities in 3D physical space), as linear projections of the latent state 19,22,42,49 .…”

“…This allowed our latent state to describe the dynamics and modes in neural population activity itself. We built these models not only for single-scale spiking or LFP activity, but also for combined spike-LFP activity using an unsupervised multiscale EM learning algorithm 42 , allowing us to characterize and compare low-dimensional dynamics and modes across scales and animals. An important consequence of our work is that behavioral predictions were learned by performing a regression of the already-estimated latent state onto behavior measurements and did not involve fitting the latent state itself to the behavior.…”

“…LFP activity is more durable, can augment the usable lifespan of recordings and enhance future BMIs for neural decoding and stimulation 24,41,[60][61][62][63] . Here, we learn the multiscale dynamics in spiking and LFP activity with a data-driven state-space model 6,18,19,21,22,42,46,47,49,53,60,[62][63][64][65][66][67][68] but by leveraging an unsupervised multiscale EM algorithm 42 that can jointly model multiscale spiking and LFP activities together. Various supervised and unsupervised machine learning algorithms have been developed for state-space models but are instead designed for single-scale activity 6,18,46,47,53,60,[62][63][64][65]68 .…”

“…Further, it is not known how low-dimensional dynamics across the different scales similarly or differentially predict reachand-grasp behavior and encode movement parameters. In addition to requiring a naturalistic reach-and-grasp task, investigating these questions requires developing a new machine learning algorithm that can extract the low-dimensional latent dynamics from combined spike-field activity-i.e., extract them across multiple scales simultaneously 42 (Methods). Answering these questions will help understand how naturalistic reach-and-grasp movements are controlled by the integrated action of large-scale brain networks as well as smaller-scale neural processes.…”