Prefrontal working memory signal controls phase-coded information within extrastriate cortex

Abstract

In order to understand how prefrontal cortex provides the benefits of working memory (WM) for visual processing we examined the influence of WM on the representation of visual signals in V4 neurons in two macaque monkeys. We found that WM induces strong $\beta$ oscillations in V4 and that the timing of action potentials relative to this oscillation reflects sensory information- i.e., a phase coding of visual information. Pharmacologically inactivating the Frontal Eye Field part of prefrontal cortex, we confirmed the necessity of prefrontal signals for the WM-driven boost in phase coding of visual information. Indeed, changes in the average firing rate of V4 neurons were correlated with WM-induced oscillatory changes. We present a network model to describe how WM signals can recruit sensory areas by inducing oscillations within these areas and discuss the implications of these findings for a sensory recruitment theory of WM through coherence.

Keywords: neural oscillations; neural phase code; prefrontal cortex; top-down control; working memory.

Figure 1.. WM alters β oscillatory power but not firing rates in V4.

A) Memory-guided saccade with background (MGS-background) task. The monkey fixated and a peripheral visual cue appeared (Cue). The monkey maintained fixation while remembering the cue location (~1s; Delay), and after the fixation point disappeared, executed a saccadic eye movement to the remembered location (Response) to receive a reward. Throughout the task, there was a task-irrelevant, full-field oriented bar background; the background contrast ranged from 0–64%, in one of 4 orientations. The memory location was either inside the extrastriate RF (IN condition, shown) or 180 degrees away (OUT condition). Neurophysiological recordings of spiking and LFP activity were made from extrastriate visual area V4, with linear array or single electrodes. B-C) Mean firing rate of a sample neuron over time for three different contrasts (B) or orientations (C) of the background stimulus, for the IN (left) and OUT (right) conditions. Shaded areas in all panels show standard error of mean (SEM). D-E) Mean firing rate of the population of 145 neurons over time for three different contrasts (low, medium, and high contrast, D) and four different orientations (preferred, nonpreferred, and middle 1 & 2 orientations, E) for the IN condition (left) and OUT (right) conditions. F) Time course of mean F-statistic values across 145 neurons, based on a one-way ANOVA for discrimination between 12 stimulus conditions for the IN (red) and OUT (black) conditions. G) Scatter plot of F-statistic averaged in the last 700ms of the delay period for each session, for the IN vs. OUT conditions. Histogram in the upper right shows the distribution of change in F-statistic (OUT-IN) across sessions. H) Mean power spectrum of the LFP recorded from the same channel as the sample V4 neuron in (B), during the delay period for the IN (red) vs. OUT (black) conditions. Inset panel shows 8–25 Hz. Asterisk indicates a significant difference (p<0.05) in the range shown. I) Mean power spectrum of population of the V4 LFPs (88 sessions) during the delay period for the IN (red) vs. OUT (black) conditions. Inset shows the power spectrum between 14–22 Hz. J) Scatter plot of power spectrum averaged in the $\mathrm{\beta }$ range for each session, for the IN vs. OUT conditions. Histogram in the upper right shows the distribution of change in power (OUT-IN) across sessions (***, p <0.001).

Figure 2.. WM alters the sensory representation in extrastriate cortex.

A) The distribution of spikes generated by a sample V4 neuron across various phases of $\mathrm{\beta }$ oscillations during the delay period. Arrows show the average of phase distributions for the IN (red) and OUT (black) conditions. B) Scatter plot of SPL in the $\mathrm{\beta }$ range for each neuron, for the IN vs. OUT conditions. Histogram in the upper right shows the distribution of change in SPL (OUT-IN) across neurons. Red cross indicates population mean. Green dot shows the selected sample neuron in A. C) Spike-triggered average of the normalized LFP of a sample neuron during the delay period, for the high contrast (red) and low contrast (blue) background stimuli, for the IN (left) and OUT (right) conditions. Shaded bars indicate the slopes in the falling phase. D) Histogram of the distribution of STA slopes $\left(\mathrm{a}\mathrm{b}\mathrm{s}\left({\mathrm{V}}_{\text{peak}}-{\mathrm{V}}_{\text{trough}}\right)/\left({\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}}_{\text{peak}}-{\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}}_{\text{trough}}\right)\right)$ across neurons, for high contrast (red) and low contrast (blue) stimuli, for the IN (left) and OUT (right) conditions. E) Population phase (left) and rate (right) coding over time, for the IN (red) and OUT (black) conditions, based on mutual information (MI) between 12 stimulus conditions. MI was measured in 100ms windows with steps of 100ms. Shaded areas show standard error of mean (SEM). F) Scatter plot of MI using a phase code in the $\mathrm{\beta }$ range (left) and rate code (right) for each neuron, for the IN vs. OUT conditions. Red crosses indicate population mean. Histograms in the upper right show the distribution of differences in MI (IN-OUT) across neurons. G) Phase coding, measured by MI (colorbar), as a function of frequency and time, for memory IN (left) and OUT (right). Black rectangle shows the time and frequency range selected for phase code analysis. (*, p <0.05; **, p <0.01; ***, p <0.001; ns, p>0.05)

Figure 3.. FEF inactivation alters WM behavioral performance and phase coding in visual areas.

A) V4 recordings were made before and after infusion of muscimol into FEF. Muscimol injections into FEF were made with a custom microinjectrode, at sites with stimulation-evoked saccade endpoints overlapping with simultaneous V4 recording site RFs. B) Eye traces for 8 MGS target locations, before (left) and after (right) FEF inactivation, for an example session where 0.5 microliter of muscimol was injected into the FEF; performance deficits were localized to the infusion hemifield. C) Average behavioral performance across sessions, at different locations over time following FEF inactivation (red pre-inactivation; green, blue, and black, 1, 2, and 3 hours after inactivation, respectively). Data from each session is aligned so that 0 degrees corresponds to the FEF RF. D-E) Normalized saccade error (D) and reaction times (E) for the memory IN (red) and OUT (black) conditions, over time relative to the FEF inactivation. Black bar indicates times with a significant difference between IN and OUT. Shaded areas show SEM across sessions. F) Heatmap shows phase coding (MI, colorbar) over time and frequency for 66 V4 neurons, for the IN condition, before (left) and after (right) FEF inactivation. Black rectangle indicates time and frequency range considered in (G-H): 14–22Hz, 200–800ms after start of delay period. G) Strength of $\mathrm{\beta }$ phase coding over time, for memory IN, before (red) and after inactivation (blue). The phase-code MI is averaged in the $\mathrm{\beta }$ range. Shading shows SEM across neurons. Gray area indicates time window plotted in (H). H) Scatter plot of $\mathrm{\beta }$ phase MI during the delay period (shaded area in G) of the IN condition for each V4 neuron, before vs. after FEF inactivation. Red square shows population mean. The histogram in the upper right shows the distribution of difference in MI (Pre-Post) across neurons. (*, p <0.05; **, p <0.01; ***, p <0.001; ns, p>0.05)

Figure 4.. Experimental and computational dependence of firing rate, phase-coded information, and rate-coded information on changes in peak oscillation frequency.

A) Schematic of dynamical neural field network architecture. Excitatory and inhibitory units are interconnected and organized to respond to different input stimuli (theta). Units also receive a global WM input (not shown); see Methods for description of connectivity weights. B) Example activity of excitatory units in the model over time, in response to an input at $\mathrm{\pi }/2$. Excitatory units are plotted along the y-axis according to their input tuning, which ranges from 0 to $\mathrm{\pi }$. Activity reflects both a beta-frequency oscillation across the entire population, and an earlier and stronger response of units whose preference matches the input feature (i.e., phase and rate coding). C) Information about the input stimulus feature coded by phase and rate (left and right y-axes; see Methods) in the neural field model, as a function of WM input strength. D) Phase information (shades of orange) and rate information (shades of blue) as a function of contrast levels and oscillation frequency, for the neural field model. Data for each contrast is divided into deciles based on oscillation frequency, variability in which comes from noise in the WM input strength. Note that rate code values are several orders of magnitude smaller than phase code values (left vs. right y-axis). Error bars show standard deviation. E) Correlation between oscillation frequency and firing rate in the neural field model under conditions of noisy WM input strength. F) Relationship between frequency of peak LFP power and firing rate during the delay period for the memory IN (red) and OUT (black) conditions, for two example neurons with increased (left) or decreased (right) firing rate during the IN condition. Each dot shows the average frequency of max power and normalized firing rate for a subsample consisting of 50% of trials (n=100 subsamples per neuron). G) Average normalized response as a function of peak frequency, pooled across subsamples of trials from each of 145 V4 neurons (100 subsamples/neuron), during the IN (red) and OUT (blue) conditions, during the delay period (left) and the cue period (right). Plot shows mean±SE for all subsamples with the peak frequency indicated on the x-axis.