The activity of neurons in primary motor cortex provides the signals that control our ability to execute movements. One of the crucial questions, still unresolved, is that of identifying the code used by this neural ensemble. Do neurons in primary motor cortex use a representation of the external physical space, or an actuator representation of the state of the muscles? We address this question through the analysis of data obtained for awake behaving monkeys executing stereotyped reaches. The data includes both simultaneous recordings of the activity of about one hundred neurons in primary motor cortex, and of the activity of about ten muscles in the relevant limb. The analysis of this data involves a variety of techniques, from linear regression models to nonlinear methods for dimensionality reduction. I will review the current level of achievement in this active area of research and discuss its implications, both for understanding aspects of neural information processing that relate to natural behaviors and for extracting from these neural signals the information needed to guide prosthetic limbs and other types of external devices.
The presentation will focus on two projects:
i) A natural representation for the ensemble activity is provided by a
high-dimensional space in which each axis represents the activity of a
single neuron as an independent degree of freedom. However, the observed
correlations among neurons whose activity is detectably modulated by a task
suggest that the population activity defines a low-dimensional space within
the high-dimensional space of independent firing activities. We have used
linear and nonlinear methods for dimensionality reduction to find the
low-dimensional structure that captures the underlying relationship between
population neural activity and behavioral task. The use of multidimensional
scaling in conjunction with an empirical measure of geodesic distances
yields a low-dimensional manifold whose intrinsic coordinates capture the
geometry of the task in the external physical space. This nonlinear mapping
provides a highly informative representation for the prediction of reach
directions.
ii) Ensemble activity is also predictive of electromyographic (EMG)
recordings of muscle activity during grasps. Appropriately regularized
linear models easily achieve 60% to 80% prediction of dynamic EMG during
palmar, lateral, and precision grasps. The ability to predict EMG signals
from neural recordings can then be used to generate useful patterns of
muscle activation to guide limb motion during nerve block. This paradigm
provides biophysically generated signals for the real-time functional
electrical stimulation (FES) of muscles that have lost nerve innervation.