Stable ensemble performance with single-neuron variability during reaching movements in primates

Abstract

Significant variability in firing properties of individual neurons was observed while two monkeys, chronically implanted with multielectrode arrays in frontal and parietal cortical areas, performed a continuous arm movement task. Although the degree of correlation between the firing of single neurons and movement parameters was nonstationary, stable predictions of arm movements could be obtained from the activity of neuronal ensembles. This result adds support to the idea that movement parameters are redundantly encoded in the motor cortex, such that brain networks can achieve the same behavioral goals through different patterns and relative contribution of individual neuron activity. This has important implications for neural prosthetics, suggesting that accurate operation of a brain-machine interface requires recording from large neuronal ensembles to minimize the effect of variability and ensuring stable performance over long periods of time.

Figure 1.

Experimental design and examples of stable and variable neurons. A, Schematics of the behavioral task and neuronal encoding analysis. The monkey controls a handheld pole while seated in front of a computer screen (i). The goal of the task is to move the cursor (small disk) to a visual target (large disk) that appears at random locations on the screen for each trial. Multidimensional linear regression is applied off-line for linearly mapping neuronal ensemble activity with motor output (ii). Independent models are generated for hand position and hand velocity coordinates. For each sample (bin size, 100 ms), 10 lags (i.e., 1 s total) before movement are assigned as free parameters to the linear model. B, Examples of stable and variable neurons during a 22 min recording session of continuous movement (first 10 min used for fitting the first model are not shown). Single-unit waveforms from PMd (top) and M1 (bottom) neurons recorded during the same session (i). Numbers in vertical axes depict the mean V_pp (in microvolts) of the action potentials. Interspike intervals (ISIs) of the depicted neurons are shown (ii). Evolution of the IRFs obtained by regressing a single unit with each motor parameter are shown (iii). Models were generated with data points from a 10 min sliding window in 30 s increments. Each model consisted of 10 free parameters from the corresponding time lags. The color bar denotes IRF values for each model. Note the different IRF patterns across time between a stable (top) and variable (bottom) neuron. Evolution of the single-unit ranking, single-unit R score, and ensemble R score across the recording session are shown (iv). For plotting purposes, neuron ranking is normalized from 0 (worst) to 1 (best) from the 94-neuron optimized ensemble. Notice the abrupt changes in both ranking and R score of the variable (bottom) neuron.

Figure 2.

Stable ensemble performance with single-neuron variability: spatiotemporal patterns, quantification, and implications of redundancy in neuronal ensembles. A, Spatiotemporal patterns of single-neuron variability based on ranking changes during a recording session. A color map shows an ensemble of neurons arranged by cortical area (i). The color bar denotes ranking (1, best; 94, worst) and the range of individual correlation coefficient scores for the whole ensemble. Arrows depict the top four neurons. A color map shows the same ensemble of neurons as in i but arranged by their initial ranking at the first epoch (ii). The mono-dimensional plot on top shows stable ensemble performance. A color map (red color) depicts statistically significant variability (p < 0.01) of a particular neuron at a particular epoch (iii). The ensemble is arranged as in ii. Mono-dimensional plots show the amount of significantly variable neurons at a given epoch (top) and the amount of epochs for a given neuron that were classified as variable within the recording session (right). B, Overall quantification of neuronal variability across all motor parameters and for four consecutive recording sessions. Variability was quantified as statistically significant (p < 0.01) changes in neuron ranking and firing rate. The numbers in parentheses indicate the size of the neuronal population in a particular cortical area. C, Cumulative (solid) and residual (dotted) neuron-ranked curves for position (X, Y) and velocity (VX, VY). Notice the smooth decay of the residual curves showing the high contribution of most of the neurons of the ensemble. D, Performance comparison between ensembles of the best 10 neurons in a given epoch and ensembles of 10 randomly selected neurons, of a total population of n = 94 units. For each epoch, 15 ensembles of randomly selected neurons were estimated (mean R ± STD).