Phase shifts in high-beta- and low-gamma-band local field potentials predict the focus of visual spatial attention

Abstract

The local field potential (LFP) contains rich information about activity in local neuronal populations. However, it has been challenging to establish direct links between LFP modulations and task-relevant behavior or cognitive processes, such as attention. We sought to determine whether LFP amplitude or phase modulations are predictive of the allocation of visual spatial attention. LFPs were recorded simultaneously in multiple early visual brain structures of alert macaque monkeys performing attention-demanding detection and discrimination tasks. Attention directed toward the receptive field of recorded neurons generated systematically larger phase shifts in high-beta- and low-gamma-frequency LFPs compared with LFP phase shifts on trials in which attention was directed away from the receptive field. This attention-mediated temporal advance corresponded to ~10 ms. LFP phase shifts also correlated with reaction times when monkeys were engaged in the tasks. Importantly, attentional modulation of LFP phase was consistent across monkeys, tasks, visual brain structures, and cortical layers. In contrast, attentional modulation of LFP amplitude varied across frequency bands, visual structures/layers, and tasks. Because LFP phase shifts were robust, consistent, and predictive of spatial attention, they could serve as a reliable marker for attention signals in the brain.

NEW & NOTEWORTHY Local field potentials (LFPs) reflect the activity of spatially localized populations of neurons. Whether alterations in LFP activity are indicative of cognitive processes, such as attention, is unclear. We found that shifts in the phase of LFPs measured in multiple visual brain areas reliably predicted the focus of spatial attention. LFP phase shifts could therefore serve as a marker for behaviorally relevant attention signals in the brain.

INTRODUCTION

The local field potential (LFP) provides a useful perspective on neuronal network activity. As low-pass-filtered voltage fluctuations, LFPs reflect subthreshold currents and lower frequency components of action potentials and are related to noninvasive signals measured with functional MRI (fMRI) and EEG (Buzsáki et al. 2012; Logothetis 2002). One major advantage of LFPs over single unit spikes is whereas single-unit spike rates are often low, LFPs are nearly continuously sampled, providing sufficient temporal resolution to evaluate modulations within single trials. Recent studies shed light on how specific neuronal types and circuits generate LFP oscillations (Bartos et al. 2007; Bharmauria et al. 2015; Cardin 2016; Otte et al. 2010; Sohal 2016; Veit et al. 2017), although uncertainties remain about the functional role of LFPs in perception, cognition, or behavior. Our objective in this study was to determine whether LFP modulations are predictive of the allocation of visual spatial attention. Attention impacts spike rate, variability, and correlation as well as LFP amplitude throughout the visual cortex; however, none of these neurophysiological readouts of attention are consistent across visual areas (see for example: Luck et al. 1997; Ruff and Cohen 2014; Siegel et al. 2008). A neurophysiological signal that undergoes a consistent modulation by attention across visual brain areas could serve as a “marker” for attention and could provide a foundation for causal tests of the neuronal mechanisms of attention. Furthermore, if that neurophysiological marker had sufficient temporal resolution, it would be possible to track the progress of attention signals across brain areas with single-trial resolution. Obtaining the necessary spatial coverage and temporal resolution to survey attention signals with single-trial resolution would be difficult based on single-unit spiking activity but might be possible with LFPs (Bastos et al. 2015).

There is a wealth of evidence supporting LFP modulation by incoming visual information as well as top-down cognitive processes such as visual attention. Visual cortical LFPs are selective for a variety of visual stimulus features including orientation, contrast, size, color, natural scenes, plaids, and stimulus repetition (Bartolo et al. 2011; Brunet et al. 2014, 2015; Gieselmann and Thiele 2008; Gray and Singer 1989; Jensen et al. 2015; Jia et al. 2013b; Katzner et al. 2009; Lashgari et al. 2012; Lima et al. 2011; Ray and Maunsell 2010; Roberts et al. 2013). Visual cortical LFPs are also modulated by attention, although attentional modulation of LFP power varies across visual areas and across layers within cortical areas (Bosman et al. 2012; Buffalo et al. 2011; Chalk et al. 2010; Fries et al. 2001; Gregoriou et al. 2009; Liang et al. 2005; Siegel et al. 2008; Spaak et al. 2012; Taylor et al. 2005). Thus LFP amplitude is modulated in a manner similar to spiking activity among visual cortical neurons, in that both are mainly modulated by incoming visual stimulus information with weaker or more variable modulation by visual attention.

Although most studies focus on changes in LFP amplitude in response to visual stimulation or allocation of spatial attention, parallel lines of inquiry suggest that LFP phase modulations convey information about the temporal dynamics of visual perception and attention. In visual cortical areas, neuronal spikes are often phase-locked to LFP oscillations in the γ (~25–80 Hz) frequency band (Cui et al. 2016; Denker et al. 2011; Gray and Singer 1989; Jia et al. 2013a; Martin and Schröder 2016; Womelsdorf and Fries 2007). Furthermore, spike-LFP coherence in the γ-band increases with attention in extrastriate cortex (Fries et al. 2001, 2008; Ni et al. 2016; Vinck et al. 2013; Womelsdorf et al. 2006), although not in primary visual cortex (Chalk et al. 2010). Whether LFP amplitude or phase modulations exert causal influences on visual perception or attention remains unresolved.

Also unresolved is whether there exists a causal link between LFP modulations and task-relevant behavior. Relationships between field potentials and task-relevant behaviors in visual detection tasks have been examined mostly with the use of noninvasive methods. A number of studies have demonstrated that phase shifts in EEG signals correlate with subjects’ reaction times and threshold detection probabilities (Busch et al. 2009; Callaway and Yeager 1960; Diederich et al. 2014; Gonzalez Andino et al. 2005; Linkenkaer-Hansen et al. 2004; Naruse et al. 2015; Samaha et al. 2015; Thut et al. 2006; Voloh et al. 2015). Among these studies are findings that EEG phase shifts, or phase resets, are greater on trials in which subjects have slower reaction times (Diederich et al. 2014; Naruse et al. 2015; Voloh et al. 2015). These findings have led to the proposal that phase shifts represent a pause or a reset when there is a mismatch between a prediction and incoming perceptual information. In the few studies directly comparing reaction times and LFP modulations in the visual cortex during visual detection tasks, correlations between faster reaction times and increased visual cortical LFP power or phase modulations in the γ-band were reported (Ni et al. 2016; Smith et al. 2015). However, γ-band LFPs and reaction times are both influenced by visual stimulus salience, making it difficult to tease apart the contributions of LFP modulations to behavior.

It is clear that LFPs in the visual cortex carry information about visual stimuli and attention state. However, significant open questions remain about the functional role of LFPs in visual perception, attention, and behavior. In this study, we addressed two key questions: 1) Can LFP phase predict reaction time in visual detection tasks? and 2) Does visual attention modulate LFP amplitude or phase in a robust and consistent manner? To answer these questions, we examined LFP amplitude and phase in the visual cortex and visual thalamus of alert monkeys performing discrimination tasks requiring covert shifts in visual spatial attention. We observed small but significant correlations between LFP phase shifts and reaction times when monkeys were engaged in the attention tasks. Moreover, attention modulated LFP phase in a manner that was robust and consistent across monkeys, visual areas, and tasks. Together, these findings suggest that LFP phase shifts could serve as a marker for behaviorally relevant visual attention.

MATERIALS AND METHODS

Three adult female macaque monkeys (Macacca mulatta) were used for this study. All of the procedures performed as a part of this study conformed to the guidelines set forth by the NIH and were approved by the Institutional Animal Care and Use Committees at the Geisel School of Medicine at Dartmouth and the University of California, Davis. Results of separate analyses of overlapping data sets have been published (Hembrook-Short et al. 2017; Mock et al. 2018).

Surgical preparation and cylinder maintenance.

Under full surgical anesthesia and with the use of sterile procedures, recovery surgeries were performed to secure headpost attachments, implant chronic electrodes, and/or make small craniotomies for recording access to visual structures. Individual headposts (Crist Instruments, Hagerstown, MD) were secured to the skull using bone screws (Synthes Vet, West Chester, PA) and bone cement (Biomet, Warsaw, IN), or three headpost platforms with pins (Thomas Recording, Giessen, Germany) were directly attached to the skull using self-tapping bone screws. Monkeys recovered from headpost placement surgery for 1–6 wk before being secured to a head-holding device while in the neurophysiological recording rig. To gain recording access to primary visual cortex (V1), a small craniotomy (~1 cm in diameter) was made over the parafoveal opercular surface of V1 and a recording cylinder was placed encircling the V1 craniotomy and secured with bone screws and bone cement. To record in the lateral geniculate nucleus of the thalamus (LGN) in one monkey, a small burr hole was made above the LGN and a seven-channel multielectrode array (Microprobes, Gaithersburg, MD) was advanced until neurophysiological responses to light stimulation through the eyes were recorded. Stereotaxic coordinates obtained from a macaque brain atlas were used to guide the chronically implanted array to the lower hemifield parafoveal representation in the LGN, and placement within this region of the LGN was verified with neurophysiological recordings conducted in surgery. Once positioned, the array was chronically cemented in place and secured with bone screws and bone cement. In two monkeys, access to the LGN was obtained by making an ~1-cm craniotomy above the LGN and encircling the craniotomy with a recording cylinder that was secured to the skull with bone screws and bone cement. Following recovery from surgery, cylinders were flushed with Betadine or chlorhexidine solution in sterile saline at least three times per week. Weekly 5-fluorouracil treatments and occasional dura scrapes were performed to maintain thin and healthy dura for ease of electrode penetration.

Visual stimulation and eye tracking.

Visual stimuli were generated using a VSG/5 or Visage system (Cambridge Research Systems, Rochester, UK). Stimuli were presented on a gamma-calibrated cathode ray tube monitor placed 56 cm in front of the monkeys’ eyes. The refresh rate of the monitor was 100 Hz, and the average luminance was 38 cd/m². The monitor was the sole source of illumination in the room with the monkey. All stimuli were presented under binocular viewing conditions. Monkeys’ eye positions were monitored by an infrared video eye tracker (Applied Science Laboratories, Bedford, MA or Thomas Recording, Giessen, Germany) with a refresh rate of 240 or 800 Hz. Monkeys’ eye position was calibrated at the start of each session. The calibration routine involved displaying four or nine dots (0.5° diameter), one at a time for up to 2 s each, on a symmetrical grid centered on the monitor. Dot fixation and calibration were determined manually or automatically.

All of the behavioral tasks involved presentation of a central fixation dot (0.3–0.5° diameter) along with one or more visual stimuli. Most visual stimuli were drifting sinusoidal or phase-reversing gratings with parameters optimized to best activate the majority of recorded neurons. To assess the location of recorded neuronal receptive fields, while monkeys maintained central fixation, a grating under computer mouse control was positioned at various locations on the monitor and varied in size, orientation, spatial frequency, and temporal frequency until the optimal position and parameters were determined. To generate robust LFP responses to determine laminar assignment of electrode contacts, while monkeys maintained central fixation, a 4-Hz flashing black/white disk (2–4° diameter) was centered over the receptive field. Neuronal tuning properties were measured during fixation-only trials in which monkeys fixated on a central dot while drifting sinusoidal gratings were presented for 1–2 s in the receptive field and varied in contrast (1–100%), orientation (0–324°), size (0.2–10°), spatial frequency (0.2–4 cycles/°), or temporal frequency (1–16 Hz) in steps of 10 (2–5 repeats per recording session). Monkeys were rewarded for maintaining fixation during fixation-only trials.

Gratings displayed in the attention task were optimized to the preferred orientation and spatial frequency of the majority of recorded neurons through online tuning analyses, were 2–4° in diameter, and had a fixed temporal frequency of 4 Hz (phase-reversing gratings had 2 phases per cycle appearing as 8 Hz). Starting grating contrast was 20–70%. The two gratings displayed in attention tasks were identical and equidistant from the central fixation dot within the same hemifield. For the change detection task, the contrast of one of two displayed gratings increased by 10%. For the contrast discrimination task, the contrast of one of the two displayed gratings either increased or decreased by 1, 5, 10, 15, or 20% and the monkey had to make an eye movement to one of two answer dots displayed above and below the fixation dot (6° of visual angle away from the central fixation dot). The gratings and central fixation dot were extinguished as the two answer dots were displayed. For the orientation discrimination task, the orientation of one of the two displayed gratings rotated either clockwise (CW) or counterclockwise (CCW) by 1, 2.5, 4, 6, or 8° and the monkey had to make an eye movement to one of two answer dots. The percentages of trials per block at each difficulty level were 4, 16, 34, 30, and 16% (4% of trials with 1% contrast or 1° orientation changes, 16% with 5% contrast or 2.5° orientation changes, and so forth). Visual stimulation of recorded neurons was identical across attention conditions.

Behavioral tasks and analyses.

Monkeys were trained to perform fixation-only tasks, contrast change detection (CD) tasks, and contrast change and orientation change discrimination tasks for juice rewards using standard operant conditioning. If the monkeys’ eye position deviated by more than 0.35° at any point during a trial in any task, the trial was aborted. For all tasks, trials were interleaved with a 1- to 2-s intertrial period during which the monitor was average luminance (gray) and monkeys were allowed to freely move their eyes.

For the CD task, monkeys were trained to detect a 10% contrast increment in one of two displayed gratings and to indicate detection by pushing a button or releasing a joystick. Monkeys had to maintain central fixation (fixation dot diameter: 0.3–0.5°) through the entire duration of each trial, including during the answer window. Monkeys learned to switch covert attention to the lower grating that always overlapped recorded receptive fields (attend-toward condition) or the upper grating (attend-away condition) according to the fixation dot color cue (see Fig. 1B), following a standard Posner cuing paradigm (Posner et al. 1980). Trials were presented in a block format with 5–20 correct trials per block, with equivalent numbers of correctly completed trials per attention condition per session.

Fig. 1.

Attention tasks, behavioral results, local field potential (LFP) recording locations, and average power. A: timeline for a single attention task trial. B, top: schematic screen shots of the change detection (CD) task. Red fixation dot cued monkeys to attend to the lower grating, overlapping recorded neuronal receptive fields (dashed circle, not displayed to monkeys; attend-toward condition); blue fixation dot cued monkeys to attend to the upper grating (attend-away condition; dots and gratings enlarged for display purposes). Attended grating contrast increased on validly cued trials (larger contrast increment shown for display purposes); nonattended grating increased in contrast on invalidly cued trials. Bottom: accuracy data for 3 monkeys on the CD task. Black curves illustrate individual session data. Only sessions with ≥5% invalidly cued trials were included in this analysis. Error bars are SE. Monkey B: *P = 0.0003; monkey O: *P = 0.004; monkey E: *P = 5.2 × 10⁻¹³, Kruskal-Wallis tests (Table 1). C, top: schematic screen shots for the contrast and orientation change discrimination tasks for attend-toward conditions (conventions as in B; CW, clockwise; CCW, counterclockwise; incr, increasing; decr, decreasing). Bottom: accuracy data for monkey E on contrast and orientation change discrimination tasks (conventions as in B). Contrast: *P = 0.001; orientation: *P = 0.0002, Kruskal-Wallis tests (Table 1). D: receptive field (RF) positions of LFPs recorded simultaneously (color coded) in lateral geniculate nucleus of the thalamus (LGN; open circles) and primary visual cortex (V1; closed circles); crosses represent RFs of additional V1 LFP recordings not paired with LGN recordings (monkey E). E: average z-scored V1 LFP power across frequencies for attend-toward (red) and attend-away (blue) trials in each monkey.

Trials progressed as follows (see Fig. 1A). After the intertrial period, the central fixation dot was displayed or monkeys initiated the fixation dot display by moving a joystick to one side. Monkeys were required to maintain the joystick in the side position throughout the duration of the trial until the answer period. Monkeys then acquired fixation and had to maintain fixation throughout the remainder of the trial. There was a cue period of 0.3 s during which no grating stimuli were displayed and the color of the fixation dot cued monkeys to attend to a particular location. The cue period was followed by a visual stimulus display period lasting 1–3 s. The duration of the visual stimulus display period varied per trial and was determined on a trial-by-trial basis according to a hazard function with an average of 1.7 s. Two drifting sinusoidal gratings were presented during the visual stimulus display period. Following the variable-duration visual stimulus display period, one of the gratings increased in contrast by 10%. Thus monkeys could not predict the timing of the contrast change on any trial. Both gratings remained on the monitor during the 1-s answer period in which monkeys signaled detection of the contrast change by moving the joystick to the original central position or pressing a button. Premature joystick movements or button presses before the contrast change resulted in aborted trials. Only correct detection of the contrast change, indicated by a correct joystick movement or button press within the answer period, while also maintaining central gaze fixation throughout the answer period was rewarded with juice. Miss trials were recorded as trials in which no answer was provided within the answer period.

Across blocks of trials, 95% of trials were validly cued, wherein the contrast change occurred at the attended location cued by the color of the fixation dot. In the remaining 5% of trials, the fixation dot color cue was invalid and the contrast change occurred at the unattended location. Accuracy was measured as the percent correct for each trial type (attend-toward validly cued, attend-away validly cued, invalidly cued), discounting aborted trials. We compared the proportion of aborted trials across attention conditions and monkeys using a nonparametric analysis of variance (Kruskal-Wallis test) and found no significant differences across attention conditions or monkeys (P > 0.5). For each session, the effect of cue on accuracy was quantified as an increase in accuracy on validly vs. invalidly cued trials greater than 2 SD of the mean accuracy for the session. Monkey B demonstrated cue effects on accuracy on 10 of 13 sessions; monkey O demonstrated cue effects on 5 of 6 sessions; and monkey E demonstrated cue effects on 38 of 39 sessions. In addition, Kruskal-Wallis tests were used to assess statistical differences in accuracy across trial types for each individual monkey (see Table 1). Reaction times were also measured as the time between the contrast change and button press or joystick movement for correctly completed trials.

Table 1.

Accuracy on CD and discrimination tasks

	Attend-Toward	Attend-Away	Invalidly Cued
CD task
Monkey B
Average ± SE	86 ± 3%	89 ± 2%	54 ± 7%
P value (Kruskal-Wallis test)		0.0003
No. of sessions		13
Monkey O
Average ± SE	63 ± 3%	67 ± 5%	13 ± 8%
P value (Kruskal-Wallis test)		0.004
No. of sessions		6
Monkey E
Average ± SE	77 ± 3%	77 ± 2%	48 ± 3%
p value (Kruskal-Wallis test)		5.2 × 10−13
Number sessions		39
Contrast discrimination task
Monkey E
Average ± SE	80 ± 2%	82 ± 2%	31 ± 10%
P value (Kruskal-Wallis test)		0.001
No. of sessions		7
Orientation discrimination task
Monkey E
Average ± SE	77 ± 2%	74 ± 2%	35 ± 7%
P value (Kruskal-Wallis test)		0.0002
No. of sessions		9

CD, change detection.

One monkey, monkey E, also learned two discrimination tasks: contrast change and orientation change discrimination tasks (see Fig. 1C). The discrimination tasks followed the same timeline and Posner cuing paradigm as the CD task; the fixation dot and fixation error window were the same, and the grating stimuli were similar. In the CD task, monkeys indicated detection of a grating change, whereas in the discrimination tasks, the monkey had to select an appropriate answer target given a grating change (akin to a two-alternative forced choice). The contrast change discrimination task differed from the CD task in that following the same visual stimulus display period, the contrast of the grating could either increase or decrease by a variable amount (see above), and the monkey had to indicate the direction of contrast change (increment or decrement) by making an eye movement to the correct answer target (high or low, respectively). Similarly, for the orientation change discrimination task, the grating rotated either CW or CCW by a variable amount (see above), and the monkey had to indicate the direction of rotation by making an eye movement to the correct answer target. Following the grating change, both gratings remained on the monitor for 0.5 s, after which both gratings and the central fixation dot disappeared and the two answer dots appeared and stayed on for 1 s during the answer period. When the grating change was very small (1% contrast change or 1° orientation change), the monkey was rewarded for a saccade to either answer target. Trials in which saccades were made to the wrong answer target were recorded as incorrect, but completed trials and reaction times were measured. Trials in which the monkey made no saccade were recorded as miss trials. Both discrimination tasks were run in longer blocks of 20–40 correct trials (with equivalent numbers of correctly completed trials per attention condition per session) to accommodate trials of varying difficulty levels. As for the CD task, 5% of discrimination trials were invalidly cued (with discriminations of mid-level difficulty). Accuracy was measured as described above. Monkey E demonstrated cue effects on accuracy on 15 of 16 discrimination task sessions. Reaction times in discrimination tasks were calculated relative to the time the fixation dot was extinguished.

A subset of recording sessions (28 sessions across 3 monkeys) included sufficient incorrect completed trials and/or miss trials to compare reaction times and/or LFP phase modulations on correct, incorrect, or miss trials. A total of 3,252 correct trials (1,318 trials from monkey B, 662 trials from monkey O, 1,272 trials from monkey E), 1,352 incorrect trials (from monkey E), and 1,387 miss trials (436 trials from monkey B, 448 trials from monkey O, 503 trials from monkey E) were analyzed. Reaction times and LFP phase modulations were recorded for incorrect completed trials. Only LFP modulations were measured for miss trials because these trials did not have valid reaction times. Trials in which monkeys’ reaction times were less than 90 ms were eliminated from further analyses.

Electrophysiological recordings.

LFPs were recorded in V1 and the LGN using single- or multielectrode arrays. All LFPs were referenced to the guide tube containing the electrodes, and no re-referencing was performed offline. LGN electrodes were either single electrodes (FHC, Bowdoin, ME) or seven-wire multielectrode arrays (MicroProbes). V1 electrodes were either single electrodes (Alpha Omega, Jerusalem, Israel) or 24-contact linear arrays (Plexon, Dallas, TX). All recorded LFPs had parafoveal receptive fields, and most V1 LFP recordings were paired with retinotopically aligned LGN recordings (see Fig. 1D). Continuous voltage data were amplified and digitized (at 10,000 Hz) using an Omniplex data acquisition system (Plexon) or a personal computer equipped with a Power 1401 system and Spike2 software package (Cambridge Electronic Design, Cambridge, UK). Continuous voltage recordings from each electrode/contact were low-pass filtered at 200 Hz and downsampled to 1,000 Hz. Laminar locations of V1 electrodes were determined for each recording session by measuring proximity to orthodromically activated geniculocortical recipient neurons in layer 4C (Briggs et al. 2013) or by determining the layer 4C/layer 5 border on the basis of the first polarity reversal in average LFP responses measured across all 24 contacts of the linear array (Maier et al. 2010, 2011) to flashed stimuli placed within recorded neuronal receptive fields. Granular (G) contacts included contacts up to 700 μm above the layer 4C/layer 5 border, and supragranular (SG) contacts included contacts >700 μm above the layer 4C/layer 5 border. Infragranular (IG) contacts were beneath the layer 4C/layer 5 border. Flat average LFP responses for top (e.g., above layer 1) or bottom (e.g., in the white matter) contacts on the linear array were excluded from further analysis.

LFP analyses and statistical comparisons.

Sessions in which monkeys completed at least 40 correct trials in each attention condition (minimum of 80 correct trials total) were analyzed. For all completed attention task trials, LFPs were extracted during the intertrial period (0.5-s window) to provide a baseline measurement of LFP amplitude, during the cue period (0.3 s window), and during the last four complete grating cycles of the visual stimulus display period before the grating change (1-s window). LFPs extracted during the visual stimulus display period on each trial were indexed to the same moment of the visual stimulus cycle to enable direct comparisons across trials, sessions, and attention conditions. As a control, LFPs during the last 1 s of the visual stimulus display period and randomized with respect to the visual stimulus grating cycle were also extracted and analyzed in parallel. Extracted LFPs were detrended and z scored. For multiarray recordings, LFPs were averaged across all contacts in each structure (in LGN or in V1) or averaged across contacts located in the SG, G, or IG laminar compartments. For all correct completed trials, LFPs were then separated by attention condition (attend-toward vs. attend-away trials) and separated into trials in which monkeys had fast or slow reaction times (less than or greater than average reaction time) or into trials with easy or hard discriminations (feature change greater or less than median feature change). The same criteria for defining fast or slow reaction times and easy or hard discriminations were applied to both attention conditions.

For all subsequent LFP analyses, data were analyzed first for all trials and sessions within individual monkeys [CD task: monkey B = 1,422 total attention trials from 13 sessions, data for V1 and LGN; monkey O = 714 total attention trials from 6 sessions, V1 and LGN; monkey E = 6,458 total attention trials from 34 V1 sessions and 1,292 total attention trials from 6 LGN sessions; discrimination tasks: monkey E = 6,244 total attention trials from 33 sessions, V1 and LGN (16 orientation and 17 contrast sessions, data combined because results were similar for contrast and orientation change discrimination tasks)]. For the analysis of easy vs. hard trials of the discrimination tasks, an additional 13 sessions (3,693 attention trials) from monkey E were analyzed. We observed no significant differences in power in θ+α (2–15 Hz), β (15–22 Hz), and low-γ (23–65 Hz) frequency bands across monkeys for either attention condition (P > 0.23 for all, Kruskal-Wallis tests; see Fig. 1E) and differential phase/frequency slopes were similar across monkeys (see Fig. 4). Ten CD task sessions performed by monkey E utilized phase-reversing rather than drifting gratings, and no differences were observed that were attributable to grating type. Therefore, LFP data from the CD task were combined as grand averages across three monkeys and included 53 sessions with V1 LFPs and 25 sessions with LGN LFPs. Grand-average LFP phase/frequency curves and differential phase/frequency curves (described below) were calculated for each attention condition and for each structure and laminar compartment (LGN, all V1, SG, G, IG laminar compartments). Grand-average LFPs, sampled at 1,000 Hz, were smoothed by 10 Hz for display purposes only, and LFPs recorded during the cue period were concatenated with LFPs recorded during the visual stimulus display period for display purposes (see Fig. 8), even though LFPs recorded during these periods were separated in real time.

Fig. 4.

Differential phase/frequency curves across individual monkeys. A: average differential phase/frequency curves for primary visual cortex (V1) local field potentials (LFPs) measured during the cue period, the visual stimulus display period for all trials (vis. stim. all), and the visual stimulus display period separated into fast reaction time trials and slow reaction time trials displayed for each monkey [monkey B (magenta), 711 trial pairs; monkey O (green), 357 trial pairs; monkey E (cyan), 3,229 trial pairs of change detection (CD) task, 1,650 trial pairs of discrimination task] with data from the CD task (closed circles and solid line) and the discrimination task (open cyan circles and dashed cyan line). Lines are linear regression fits; error bars are SE. R²>0.81, *P < 0.016 for all. B: average differential phase/frequency curves for lateral geniculate nucleus of the thalamus (LGN) LFPs measured during the same trial periods and types with conventions as in A. Numbers of trial pairs per monkey are the same as in A, except for monkey E (646 trial pairs for the CD task). R² > 0.72, *P < 0.03.

Fig. 8.

Average local field potentials (LFPs) and LFP power in primary visual cortex (V1) and lateral geniculate nucleus of the thalamus (LGN). A: average LFPs measured in V1 (53 sessions) and LGN (25 sessions) during attend-toward (red) and attend-away (blue) trials in the change detection (CD) task. Vertical line at time = 0/t ms indicates break in real time between activity recorded during cue (−300 to 0 ms) and the last 4 grating cycles of the visual stimulus display period (t to t+1,000 ms) of each trial. B: average LFP power (z-scored) in V1 and LGN measured during the cue (left) and visual stimulus display (Vis. stim.; right) periods of attend-toward (red) and attend-away (blue) trials in the CD task. Shading represents SE. Black dots under curves represent frequencies at which there were significant differences in power across attention conditions (Wilcoxon tests Bonferroni corrected at P < 0.05). C and D: average LFPs measured in V1 and LGN (C) and power (D; z-scored) measured in V1 and LGN during discrimination tasks (33 sessions). Conventions as in A and B; a.u., arbitrary units.

For attention task trials, LFP phase angles at frequencies between 2 and 70 Hz were determined per trial from Fourier-transformed LFPs by using Hamming windows with length equal to the analysis window duration. Phase data were extracted for the cue period (300 ms), the visual stimulus display period (1,000 ms), and for four 250-ms windows of the visual stimulus display period. Per trial, phase data for frequencies between 2 and 24 Hz were discretized every 4 Hz, and phase data for frequencies between 20 and 70 Hz were discretized every 10 Hz. Discretized phase data were plotted against increasing frequency, linear regression fits were applied to each curve (separately for data at 2–24 Hz and 20–70 Hz), and slopes of the phase/frequency linear regression fit were calculated per trial. With the same procedures, LFP phase/frequency slopes were measured for completed trials of fixation-only tasks (3,998 total fixation-only trials: 768 trials from monkey B, 430 trials from monkey O, 2,800 trials from monkey E). A nonparametric two-sample comparisons test (Wilcoxon rank sum test) was used to compare phase/frequency slopes on fixation-only trials with those on attention trials.

LFP phase/frequency slopes measured during correct and incorrect trials were plotted against reaction times. Linear regression analyses were performed on each plot to establish significant relationships and obtain R² and P values. Distributions of phase/frequency slopes for correct, incorrect, and miss trials were tested for significant deviations from zero using Kolmogorov-Smirnov tests. Wilcoxon tests were used to compare phase/frequency slopes across correct and incorrect trials or correct and miss trials using data from the same 28 sessions (described above). These comparisons were evaluated for each monkey individually and for data pooled across monkeys.

As a secondary test of the correlation between reaction time and LFP phase/frequency slope for correct trials of the attention task, we used a simple classifier to determine whether fast or slow reaction times (less than or greater than average reaction time) were predicted by phase/frequency slope. We performed a receiver operating characteristic (ROC) analysis with a bootstrapping procedure to obtain 95% confidence intervals (1,000 repeats per measure). We compared the true positive rates from classification of actual data with the true positive rates from classification of data in which phase/frequency slopes were randomized using a Wilcoxon test for significant differences.

All subsequent LFP phase analyses were performed on correct trial data from the attention tasks. Within each session, differential phase/frequency curves were calculated as attend-toward average phase/frequency minus attend-away average phase/frequency using discretized phase data described above. Grand averages of differential phase/frequency curves were made for the following trial periods and types of trials: cue period, visual stimulus display period, four 250-ms windows spanning the visual stimulus display period, visual stimulus display period for trials in which monkeys had fast vs. slow reaction times, and visual stimulus display period for trials in which the monkey successfully completed easy vs. hard discriminations. For each differential phase/frequency curve, we performed regression analyses, applied linear regression fits, and calculated goodness of fits. Slopes of differential phase/frequency linear regression fits were determined as one method to estimate the fixed temporal shift of LFP oscillations at increasing frequencies (Baldauf and Desimone 2014; Schoffelen et al. 2005). Because attend-toward and attend-away LFPs were both indexed to the same moment of the visual stimulus grating cycle, the slope of the differential phase/frequency linear regression fit estimates the time difference, as a lead or lag depending on the sign of the slope (Schoffelen et al. 2005), between attend-toward and attend-away LFPs. Differential phase/frequency slopes were compared across trial periods and types for LFPs measured in LGN, V1, and the laminar compartments (SG, G, and IG) within V1 using Kruskal-Wallis tests. As an illustration of progressive LFP phase shifts, we filtered single-session average LFPs at increasing frequencies (see Figs. 2A and 3A).

Fig. 2.

Local field potential (LFP) phase/frequency curves and reaction times. A, top: single-session average attend-toward trial LFPs (red) recorded during the visual stimulus display period (black line = 1,000 ms; gray curve represents drifting grating) show progressive advance and decrease in phase angle for bandpass-filtered LFP components between 20 and 60 Hz (filter ranges at right; expanded timescale, black line = 250 ms). Arrows highlight progressive advance of phase angle. Bottom: average phase/frequency curve for primary visual cortex (V1) LFPs measured during the visual stimulus display period of the change detection (CD) task (8,594 correctly completed trials from 53 sessions across all 3 monkeys). Line is linear regression fit to data (dots; R² = 0.97, *P = 0.0003). Phase/frequency curve slope (in ms), calculated from the fit, is indicated. Error bars are SE. Inset: average phase/frequency curve for V1 LFPs for low frequencies, conventions as in main panel (R² = 0.3, P = 0.3). B, left: phase/frequency slopes for correctly completed trials of the CD task (3,252 trials from 28 sessions across all 3 monkeys) plotted against reaction times (orange line illustrates regression fit: slope = −19, R² = 0.003, *P = 0.004). Dashed black line represents zero. Inset: receiver operating characteristic curve (orange) with 95% confidence intervals (gray shading). Middle: distribution of phase/frequency slopes across correct trials (*P = 1.3 × 10⁻³⁹, Kolmogorov-Smirnov test; orange line illustrates average phase/frequency slope = −21.3 ± 1.3 ms). Right: distributions of phase/frequency slopes across correct trials for each monkey individually (*P < 0.0001 for all, Kolmogorov-Smirnov tests). Colored lines illustrate average phase/frequency slopes: monkey B (magenta) = −33.6 ± 2.9 ms, monkey O (green) = −21.8 ± 3.8 ms, monkey E (cyan) = −8.9 ± 2.8 ms. C, left: phase/frequency slopes for incorrect but completed trials of the discrimination task (monkey E: 1,352 trials) plotted against reaction times (orange line illustrates regression fit: R² = 6 × 10⁻⁴, P = 0.2). Middle: distribution of phase/frequency slopes for incorrect trials of the discrimination task (orange line illustrates average phase/frequency slope = −2 ± 1 ms). Right: distributions of phase/frequency slopes for miss trials for each monkey individually. Colored lines (under dashed black line) illustrate average phase/frequency slopes: monkey B = −4.3 ± 3.3 ms, monkey O = −7.0 ± 3.4 ms, monkey E = −4.8 ± 1.7 ms.

Fig. 3.

Differential phase/frequency curves across tasks. A, left: single-session example of progressive advance and decrease in phase angle for filtered local field potentials (LFPs; as in Fig. 2A) for attend-toward (red) and attend-away (blue) trials. Right: average phase/frequency curves for primary visual cortex (V1) LFPs measured during the visual stimulus display period on attend-toward (red; 2,142 trials) and attend-away (blue; 2,142 trials) correct trials in which monkeys had slow reaction times in the change detection (CD) task. Lines are linear regression fits to data (dots). Error bars are SE. B: average differential phase/frequency curves for V1 (left; 2,142 trial pairs) and lateral geniculate nucleus of the thalamus (LGN) LFPs (right, flatter curve; 1,714 trial pairs) measured during the visual stimulus display period on slow reaction time trials, with linear regression fits (V1: R² = 0.94, *P = 0.001; LGN: R² = 0.71, *P = 0.04) and for LGN LFPs measured during the cue period (right, steeper curve; R² = 0.83, *P = 0.008). Error bars are SE (Table 2). Insets: average differential phase/frequency slopes for V1 and LGN LFPs calculated from differential phase/frequency linear regression fits (left, 53 V1 sessions; right, 25 LGN sessions) from the cue, visual stimulus display period (all trials), and visual stimulus period of trials in which monkeys had fast or slow reaction times. Error bars are SE. *P ≤ 0.02, Kruskal-Wallis tests (Table 2). C: average differential phase/frequency curves for V1 and LGN LFPs measured during the discrimination tasks with linear regression fits (Table 2; conventions as in B). Left: V1 slow reaction time trials (1,605 trial pairs); fit to 20–50 Hz data: R² = 0.94, *P = 0.03 (inset: data from 33 sessions, *P = 0.02, Kruskal-Wallis test). Right: LGN slow reaction time trials (1,605 trial pairs); fit: R² = 0.98, *P = 0.0001 (inset: data from 33 sessions). Error bars are SE. D: average differential phase/frequency curves for V1 (left; 1,441 trial pairs) and LGN (right; 740 trial pairs) LFPs measured during the visual stimulus display period of trials in which monkey E successfully completed hard discriminations (V1 fit: R² = 0.92, *P = 0.002; LGN fit: R² = 0.97, *P = 0.0003). Insets (46 sessions for V1, 33 sessions for LGN) illustrate slopes for cue, visual stimulus display period (all trials), and trials in which monkey E successfully completed easy or hard discriminations (*P = 0.04, Kruskal-Wallis test; Table 2). Error bars are SE. Conventions as in B and C.

As an independent measure of temporal shifts between attend-toward and attend-away LFPs, we performed an LFP amplitude cross-correlation. A negatively shifted peak in the cross-correlation of attend-toward with attend-away LFPs recorded in the same session would indicate that the attend-toward LFP led the attend-away LFP (Adhikari et al. 2010; Catanese et al. 2016). As with the calculation of differential phase/frequency described above, direct comparisons of attend-toward and attend-away LFPs were possible because both measurements were indexed to the same moment of the visual stimulus grating cycle. To perform the amplitude cross-correlation, session average LFPs in each attention condition, separated into trials in which monkeys had fast or slow reaction times, were first filtered using a 20- to 70-Hz bandpass filter. Next, the amplitude modulation of the filtered LFP was calculated using the absolute value of the Hilbert transform of the filtered LFP. The amplitude modulations for each attention condition were cross-correlated using the “xcov” MATLAB function to remove any direct current offsets in the data. These steps were performed on slow and fast reaction time trials separately (for each session in all tasks). The peak in the amplitude cross-correlation was determined for each session in which at least one cross-correlogram had a significant peak, defined as a peak >2 SD above the cross-correlation generated after shuffling of the Fourier components (Catanese et al. 2016). Twenty-five V1 sessions (of 86; 7 from monkey B, 2 from monkey O, 16 from monkey E) had significant amplitude cross-correlation peaks for either slow or fast reaction time trials. Distributions of these amplitude cross-correlation peak values were made for slow and fast reaction time trials, and a Wilcoxon test was used to determine whether the distributions differed significantly. Kruskal-Wallis tests were used to compare cross-correlation peak values for fast or slow reaction time trials across monkeys.

To test the notion that attention-mediated LFP phase shifts could give rise to latency changes for phase-aligned neuronal spikes, V1 single-unit spiking data obtained from the same three monkeys were analyzed. Single-unit identification and spike sorting were performed as described previously (Hembrook-Short et al. 2017). Spike trains for 130 V1 single units (95 single units from monkey E, 23 single units from monkey B, 12 single units from monkey O) were aligned with simultaneously recorded LFPs averaged across V1 contacts and bandpass filtered from 20 to 70 Hz. For each trial, spikes occurring within ±5 ms of filtered LFP peaks or troughs were separated and collapsed into single 11-ms windows centered on the peak/trough. Collapsed spike data were then summed across trials in each attention condition. Spike density functions were calculated from the collapsed/summed spike data using a Gaussian filter with a standard deviation of 10 ms. We determined the latency of each spike density function as the time of the first peak. Separate spike density functions were created from spikes occurring on attend-toward and attend-away trials, and latencies were determined for each. Single units without clear peaks in the spike density functions were discarded from further analyses (20 units excluded from peak analysis, 33 units excluded from trough analysis). Latencies were then compared across attention conditions for all remaining single units using Wilcoxon tests. As an additional comparison, spike onset latencies, computed as the time to the session-average peak in the spiking peristimulus time histogram (PSTH), were also calculated for the same single units, and these latencies were compared across attention conditions using a Wilcoxon test. For a subset of the same V1 single units (60; 44 from monkey E, 11 from monkey B, 5 from monkey O), spike onset latencies were also calculated for responses to high (>50%) and low (<30%) contrast grating stimuli.

Per-trial power spectra were computed from LFPs by using Welch’s method with a Hamming window of 1,000 ms and a sampling frequency of 1,000 Hz. LFPs were not windowed for amplitude measurements. Per-trial power spectra were averaged across trials for each electrode/contact, and the evoked power spectrum was calculated for each electrode/contact as the cue period or visual stimulus display period average per-trial power minus the average intertrial period (baseline) power divided by the standard deviation of the intertrial power, to produce a z-scored power measurement (Gieselmann and Thiele 2008). Evoked power spectra for each electrode/contact were averaged within structures or laminar compartments and separated by attention condition. Because power in the frequency bands of interest was not different across monkeys, power data were combined across sessions and monkeys and SEs were calculated across all trials (total trials described above). Nonparametric two-sample comparisons tests (Wilcoxon test) with Bonferroni correction for multiple comparisons at P = 0.05 were utilized to determine statistically significant differences in power at each frequency across attention conditions. As an additional quantification of LFP amplitude, average power within three frequency bands, θ+α (2–15 Hz), β (15–22 Hz), and low γ (23–65 Hz), was quantified per structure/laminar compartment (LGN and SG, G, and IG laminar compartments) across sessions for each task. Kruskal-Wallis tests were used to determine statistical differences across structure/laminar compartment and attention conditions for each frequency band analyzed. Kruskal-Wallis tests were also used to compare average power in the same three frequency bands across monkeys.

RESULTS

Direct links between LFP modulations and task-relevant behaviors have been difficult to establish, and it is unclear whether LFP modulations by themselves could serve as a consistent proxy for attentional modulation of neuronal activity (Ray and Maunsell 2015). In this study, we explored whether LFP modulations are predictive of the focus of visual spatial attention.

Attention tasks and behavioral results.

We recorded LFPs in primary visual cortex (V1) and the visual thalamus (LGN) of three alert monkeys performing grating change discrimination tasks requiring covert shifts in visual spatial attention. Attention task trials followed the same timeline (Fig. 1A). Following an intertrial period, monkeys initiated a trial by acquiring and maintaining fixation on a central dot throughout the cue period lasting 0.3 s. No visual stimuli were present in neuronal receptive fields during the cue period; however, the fixation dot color cued monkeys to attend to a particular location on the monitor. The visual stimulus display period lasted 1–3 s, during which two identical drifting or phase-reversing sinusoidal gratings were presented. One grating was placed inside and the other outside the receptive fields of recorded neurons, and both stimuli were equidistant from the central fixation dot. After an unpredictable duration, the contrast or orientation of one grating changed, prompting monkeys to make an appropriate answer during the answer period. In the contrast-increment detection task, termed the CD task (Fig. 1B, top), monkeys detected a 10% increase in grating contrast by pressing a button or releasing a lever. One monkey also performed two additional discrimination tasks in which the monkey had to indicate the direction of grating contrast change (increment or decrement) or orientation change (CW or CCW) by making a saccade to the appropriate answer target (Fig. 1C, top). Discrimination difficulty varied across trials within each block of the discrimination tasks. In both the CD and discrimination tasks, 5% of grating changes were invalidly cued (Posner et al. 1980), and monkeys were rewarded for correct detections of grating changes at invalidly cued locations. Monkeys performed between ~65% and 90% correct on validly cued trials, and all were less accurate on trials in which the grating change was invalidly cued (Fig. 1, B and C; P ≤ 0.004 for all, Kruskal-Wallis tests; see Table 1 for statistics), indicating that all three monkeys learned to correctly shift the spatial locus of attention according to the cue.

Reaction times on correctly completed trials varied across monkeys and answer methods. Saccading to an answer target resulted in the shortest reaction times (average reaction time for monkey E = 212 ± 6 ms for validly cued trials), followed by joystick release (average reaction times: monkey B = 313 ± 7 ms, monkey O = 349 ± 12 ms for validly cued trials), followed by button press (average reaction time for monkey E = 447 ± 11 ms for validly cued trials). Across monkeys and tasks, reaction times were longer on correctly completed, invalidly cued trials [19% longer for the CD task (see Hembrook-Short et al. 2017), 48% longer for the discrimination tasks]. All three monkeys displayed behaviors suggestive of speed-accuracy trade-off strategies for the CD and discrimination tasks. In the CD task, all three monkeys had relatively high rates of aborting trials by premature answering, or “jumping the gun”: ~40% of total trials in which fixation was maintained for >1 s were aborted by premature lever release or button press. Furthermore, reaction times on attend-toward and attend-away trials were significantly shorter on incorrect but completed trials compared with correct trials of the discrimination tasks (P = 1.8 × 10⁻³⁵, Wilcoxon rank sum test; monkey E: average correct trial reaction time = 233 ± 3 ms, average incorrect trial reaction time = 218 ± 2 ms). These behavioral data suggested that monkeys’ attention states varied, providing unique opportunities to examine LFP modulations across varying states of attention, described below.

LFP recordings.

With the use of single electrodes or multielectrode arrays, LFPs were recorded at retinotopically aligned regions in V1 and the LGN corresponding to parafoveal eccentricities (Fig. 1D). Additional V1 recordings not paired with LGN recordings were also made in one monkey. Voltage signals from all electrodes were digitized and low-pass filtered at 200 Hz to generate LFPs recorded from all electrodes. Average V1 LFP power measured during visual stimulus presentation in attention tasks was largely consistent across monkeys (Fig. 1E). Although session-average power in θ+α (2–15 Hz), β (15–22 Hz), and low-γ (23–65 Hz) frequency bands was not significantly different across monkeys for either attention condition (P > 0.23 for all, Kruskal-Wallis tests), attentional modulation of LFP power varied across frequency bands and tasks, discussed further below. Importantly, coherence between the LGN and V1 in high-β- and low-γ-frequency bands was significantly elevated relative to baseline during the cue and visual stimulus display periods of the attention task (Mock et al. 2018). We therefore focused primarily on attentional modulation of LFPs in the high-β- and low-γ-frequency bands.

LFP phase shifts.

When we examined the phase of LFPs between 2 and 70 Hz, we noted a trend whereby phase was progressively more negative at higher frequencies. Figure 2A, top, illustrates a representative single session in which V1 LFPs bandpass filtered at increasing frequencies demonstrated a progressive advance and decrease in phase angle. Average phase/frequency curves were flat for frequencies <20 Hz (Fig. 2A, bottom inset) but were highly linear with negative slopes for frequencies from 20 to 70 Hz (Fig. 2A, bottom). A negative linear relationship between phase and frequency suggested that LFP oscillations in the high-β-to-γ-frequency bands were shifted forward by a constant time, with the slope defining the temporal shift (Baldauf and Desimone 2014; Schoffelen et al. 2005). The slope of the grand-average V1 LFP phase/frequency curve for all correct trials of the CD task corresponded to a temporal lead of −12.1 ms. In other words, high-β-to-γ-frequency LFP components underwent a progressive advance and decrease in phase angle, corresponding to a constant temporal shift of −12.1 ms. Going forward, we will refer to this progressive advance and decrease in phase angle corresponding to a constant temporal lead as “phase shift” for simplicity.

For most analyses, LFPs were extracted from the last four complete visual stimulus grating cycles of each trial; in other words, LFPs across trials were indexed to the same moment in the visual stimulus grating cycle. This enabled direct comparison of LFP phase shifts on attend-toward and attend-away trials, discussed below. Importantly, LFP phase shifts were not dependent on a fixed relationship to the visual stimulus grating cycle. LFPs extracted from the last second of each trial and randomized with respect to the visual stimulus grating cycle also displayed linear phase/frequency curves with significant negative slopes corresponding to a temporal lead of −8 ms, on average. Additionally, because LFPs were extracted from the last second of the visual stimulus display period for each trial and the visual stimulus display period had a minimum duration of 1.2 s, the observed LFP phase shifts were not due to visual stimulus onset transient responses, which typically occur in V1 within 50 ms of stimulus onset (Maunsell and Gibson 1992). Similarly, LFP phase shifts were relatively consistent across the 1-s analysis window. When phase shifts were measured for four 250-ms LFP segments per trial, phase/frequency slopes were slightly more negative during the first and last 250 ms per trial, consistent with the notion that attention varies dynamically over the course of trials (Mock et al. 2018), but were not significantly different across segments (P = 0.99, Kruskal-Wallis test).

LFP phase shifts and behavioral responses.

Motivated by prior results that phase shifts in EEG signals correlate with reaction times in visual detection tasks (Callaway and Yeager 1960; Naruse et al. 2015; Samaha et al. 2015), we compared LFP phase with reaction time in the CD and discrimination tasks. We observed a small but significant negative correlation between V1 LFP phase/frequency slope and monkeys’ reaction times across correct trials of the CD task (Fig. 2B), consistent with similar negative correlations between EEG phase shifts and reaction times observed previously (Diederich et al. 2014; Naruse et al. 2015; Voloh et al. 2015). More negative phase/frequency slopes (corresponding to greater constant temporal leads) occurred on trials in which monkeys had slower reaction times (Fig. 2B, left; regression fit slope = −19, R² = 0.003, P = 0.004; 3,250 total correct trials from 28 sessions across all 3 monkeys). Because the correlation between LFP phase shifts and reaction times was small, we performed a secondary test of the relationship between LFP phase/frequency slope and reaction time by classifying fast or slow reaction times on the basis of phase/frequency slopes. A simple classifier performed significantly better than chance (area under ROC curve = 0.54; Fig. 2B, inset), and the true positive rate for classification of the original phase/frequency slope and reaction time data was significantly greater than the true positive rate generated from classification of shuffled data (P = 0.04, Wilcoxon test). This link between larger phase shifts and slower reaction times, combined with behavioral evidence that monkeys employed speed-accuracy trade-off strategies, hinted that LFP phase shifts were predictive of focused attention.

Across correct trials in the CD task, phase/frequency slopes were significantly negative (Fig. 2B, middle; P = 1.3 × 10⁻³⁹, Kolmogorov-Smirnov test; average phase/frequency slope, or temporal lead = −21.3 ± 1.3 ms). Similar negative distributions of phase/frequency slopes were observed for correct trials in the CD task from each monkey individually (Fig. 2B, right; P < 0.0001 for all, Kolmogorov-Smirnov tests). Interestingly, the distribution of phase/frequency slopes for incorrect trials was not shifted from zero (Fig. 2C, middle; P = 0.2, Kolmogorov-Smirnov test; average phase/frequency slope =−2 ± 1 ms), suggesting an absence of LFP phase shifts on trials in which monkeys were guessing, not paying attention, or made incorrect responses. Furthermore, LFP phase shifts were not correlated with reaction times on incorrect trials, because there was no relationship between V1 LFP phase/frequency slope and reaction time across incorrect trials (Fig. 2C, left; regression fit R² = 6 × 10⁻⁴, P = 0.2; 1,352 incorrect trials from discrimination tasks performed by monkey E). Additionally, phase/frequency slopes for correct and incorrect trials in the discrimination tasks were significantly different from one another (P = 5.2 × 10⁻⁸, Wilcoxon test). Examination of attention trials in which monkeys missed the stimulus change altogether revealed significant differences in LFP phase/frequency slope between miss trials and correct trials (P = 2.4 × 10⁻¹⁰, Wilcoxon test; average phase/frequency slope for miss trials = −4.8 ± 1.7 ms). Phase/frequency slopes were significantly different across miss and correct trials in two of three monkeys individually (Fig. 2C, right; P < 0.02, Wilcoxon tests). Taken together, these results support a link between LFP phase shifts and task-relevant behavior. Namely, LFP phase shifts in the high-β-to-γ-frequency bands were predictive of slower reaction times on trials in which monkeys paid attention, optimizing accuracy over speed, and successfully completed the task. In contrast, on trials in which monkeys were less attentive, or were optimizing speed over accuracy, LFP phase shifts were random and near zero, on average.

Attentional modulation of LFP phase.

We next sought to explore the impact of attention on LFP phase shifts. We noted that phase/frequency slopes were 10% more negative, on average, for attend-toward trials (P = 0.04, Wilcoxon test) even though phase/frequency slopes were significantly shifted from zero for both attention conditions (attend-toward average phase/frequency slope = −21.6 ± 1.8 ms, significantly negative P = 2.2 × 10⁻³³, Kolmogorov-Smirnov test; attend-toward phase/frequency slope vs. reaction time: regression slope = −21, R² = 0.004, P = 0.012; attend-away average phase/frequency slope = −19.6 ± 1.8 ms, significantly negative P = 1.2 × 10⁻²⁶, Kolmogorov-Smirnov test; attend-away phase/frequency slope vs. reaction time: regression slope = −21, R² = 0.004, P = 0.013). The small increase in LFP phase-shift magnitude with attention was reminiscent of attentional modulations of neuronal spiking activity, such as small increases, on average, in neuronal firing rate or small decreases, on average, in neuronal spike count correlations (Cohen and Maunsell 2009; Luck et al. 1997; Mitchell et al. 2009; Motter 1993). Because pooling LFP phase-shift data across all recording sessions could mask within-session attention effects, we focused on comparisons of attentional modulation of LFP phase across sequential blocks of trials within individual sessions. We hypothesized that attention directed toward the visual stimulus in the receptive field would cause consistently and significantly larger phase shifts for LFPs in the high-β-to-γ-frequency bands relative to LFP phase shifts on attend-away trials when measured in the same session. Direct comparison of LFPs recorded on attend-toward and attend-away trials during a single session revealed a larger progressive advance and decrease in phase angle for attend-toward compared with attend-away V1 LFPs bandpass filtered at increasing frequencies (Fig. 3A, left; same session illustrated in Fig. 2A with addition of attend-away trial data). Grand-average V1 LFP phase/frequency curves calculated separately for attend-toward and attend-away trials in which monkeys had slower reaction times in the CD task revealed a negative linear relationship between phase and frequency for high-β-to-γ-frequency LFPs on attend-toward trials only; the relationship between phase and frequency was flat for attend-away trials (Fig. 3A, right; this and all subsequent analyses were performed on correct trials only). Qualitatively, these grand averages supported our hypothesis that attention directed toward the stimulus in the receptive field caused systematically larger LFP phase shifts among V1 LFPs.

To quantify attentional modulation of LFP phase, we took advantage of the fact that LFPs extracted on all trials were indexed to the same moment of the visual stimulus cycle. We directly compared the relative phase of attend-toward and attend-away trial LFPs within each recording session by calculating the differential phase at each frequency between 20 and 70 Hz (attend-toward phase/frequency curve minus attend-away phase/frequency curve per session, analogous to red minus blue curves depicted in Fig. 3A, right). If per-session differential phase/frequency curves were linear with negative slopes, this would indicate that LFP phase shifts were consistently larger when attention was directed toward the stimulus within the receptive field. Figure 3B, left, illustrates the grand-average differential phase/frequency curve measured from V1 LFPs recorded during the visual stimulus display period on slow reaction time trials in the CD task. There was a significant negative linear relationship between differential phase and frequency for high-β-to-γ-frequency LFPs (regression fit: R² = 0.94, P = 0.001), indicating a significant effect of attention. The slope of the differential phase/frequency curve corresponded to a constant temporal lead of −14.3 ms. Importantly, this temporal lead also corresponded to the V1 LFP phase shift attributable to attention. Not surprisingly, attentional modulation of LFP phase was limited to the 20- to 70-Hz frequency range. Because session-average phase/frequency curves, regardless of attention condition, were flat for frequencies <20 Hz (Fig. 2A, bottom inset), differential phase/frequency curves were also flat for frequencies below 20 Hz; and this effect was consistent across monkeys and tasks (average regression slope = −0.01, average R² = 0.81, phase/frequency slopes not different from zero, P > 0.5, Kolmogorov-Smirnov tests).

We performed two quantifications of attentional modulation of LFP phase with separate statistical tests for each: 1) we determined whether attention significantly increased the magnitude of LFP phase shifts (as described above for V1 LFPs measured during slow reaction time trials of the CD task and illustrated in Fig. 3B, left); and 2) we compared attentional modulation of LFP phase across different trial outcomes (e.g., trials in which monkeys had faster or slower reaction times) and across different trial periods (e.g., cue and visual stimulus display periods). In the second quantification of V1 LFP phase from the CD task, we found that differential phase/frequency slopes measured during the visual stimulus display period were significantly more negative on slow reaction time trials compared with fast reaction time trials and were also significantly different from slopes measured during the cue period (Fig. 3B, left inset; see Table 2 for statistics). Thus qualitative and two quantitative measures of V1 LFP phase supported our hypothesis that attention significantly impacted LFP phase, with the largest attention-mediated LFP phase shifts occurring on trials in which monkeys had slower reaction times.

Table 2.

Differential phase/frequency slopes for V1 and LGN LFPs

	Cue Period	Vis. Stim. Period (All Trials)	Fast Reaction Time Trials	Slow Reaction Time Trials
CD task (Kruskal-Wallis tests used for all)
V1
Average ± SE	+1.0 ± 1.4 ms	−2.0 ± 1.6 ms	−0.5 ± 1.7 ms	−14.3 ± 3.5 ms
No. of sessions	53
P value	0.001
Differences	Slow reaction time trials < cue and fast reaction time trials
LGN
Average ± SE	−8.2 ± 2.3 ms	+0.3 ± 2.4 ms	+1.3 ± 3.1 ms	−2.1 ± 1.1 ms
No. of sessions	25
P value	0.02
Differences	Cue < visual stimulus period

	Cue Period	Vis. Stim. Period (All Trials)	Fast Reaction Time Trials	Slow Reaction Time Trials	Easy Discrimination Trials	Hard Discrimination Trials
Discrimination task (Kruskal-Wallis tests used for all)
V1
Average ± SE	+3.2 ± 2 ms	+3.0 ± 2.9 ms	+4.4 ± 3.5 ms	−17.2 ± 9.9 ms	−8.5 ± 3.7 ms	−12.6 ± 4.6 ms
No. of sessions	33
P value	0.02				0.04
Differences	Slow reaction time trials < cue and fast reaction time trials				Hard trials < all other trials
LGN
Average ± SE	−2.4 ± 1.8 ms	−0.8 ± 1.8 ms	+0.9 ± 2 ms	−11.9 ± 6 ms	−0.9 ± 2.1 ms	−3.4 ± 2.3 ms
No. of sessions	33
P value	0.2				0.6
Differences

CD, change detection; LFPs, local field potentials; LGN, lateral geniculate nucleus of the thalamus; V1, primary visual cortex; Vis. stim., visual stimulation.

In order for LFP phase shifts to serve as a reliable marker or proxy for attention, they must be present in multiple visual brain structures. We therefore asked whether attention facilitated LFP phase shifts in another visual brain structure, the LGN. Attention significantly increased the magnitude of LFP phase shifts in the LGN during slow reaction time trials and also during the cue period of the CD task, because grand-average differential phase/frequency curves for both had significant negative regression slopes (P < = 0.04; Fig. 3B, right; Table 2). The average differential phase/frequency slopes for LGN LFPs measured during the cue period and for slow reaction time trials corresponded to attention-mediated temporal leads of −8.2 ± 2.3 and −2.1 ± 1.1 ms, respectively. Additionally, attention-mediated phase shifts in the LGN were larger during the cue period compared with all visual stimulus period trials (Fig. 3B, right and inset). Thus attention caused a significant shift in LFP phase in both structures that was pronounced on trials in which monkeys had slower reaction times.

Further comparisons across tasks and visual brain structures revealed remarkable consistency in attentional modulation of LFP phase. Attention increased the magnitude of LFP phase shifts for V1 and LGN LFPs measured during the visual stimulus display period of the discrimination tasks: differential phase/frequency slopes were significantly negative for slow reaction time trials (P ≤ 0.03, Kruskal-Wallis tests; Fig. 3C; Table 2). Additionally, in V1, attentional modulation of LFP phase was larger for slow reaction time trials compared with fast reaction time trials and cue period data (P = 0.02, Kruskal-Wallis tests; Fig. 3C, left inset; Table 2). The average differential phase/frequency slopes for V1 and LGN LFPs measured during the discrimination tasks corresponded to attention-mediated temporal leads of −17.2 ± 9.9 and −11.9 ± 6 ms, respectively (Fig. 3C; Table 2).

With the discrimination task dataset, we had the opportunity to further explore the relationship between attentional modulation of LFP phase and behavior by examining LFP phase across trials varying in difficulty. Evidence suggests that attentional modulation of neuronal activity scales with increasing task difficulty (Chen et al. 2008) and that EEG phase correlates with reaction times on difficult detection tasks (Busch et al. 2009; Diederich et al. 2014; Linkenkaer-Hansen et al. 2004). Our findings up to this point indicated that LFP phase shifts were greatest on trials in which monkeys focused spatial attention toward the stimulus in the receptive field and had slower reaction times. We therefore hypothesized that attention-mediated LFP phase shifts would also be larger on trials with more difficult discriminations. To examine the relationship between attentional modulation of LFP phase and task difficulty, we separated LFP data into trials in which the monkey successfully completed easy or hard grating change discriminations. We observed consistent attention-mediated increases in the magnitude of LFP phase shifts during the visual stimulus display period of hard trials in both V1 and LGN (Fig. 3D). Differential phase/frequency slopes corresponded to attention-mediated temporal leads of −12.6 ± 4.6 and −3.4 ± 2.3 ms in V1 and the LGN, respectively. In confirmation of our hypothesis, attention had a significantly larger effect on LFP phase during hard discrimination trials: differential phase/frequency slopes for V1 LFPs were significantly more negative on hard discrimination trials compared with all other trials (P = 0.04; Fig. 3D, insets; Table 2).

In addition to consistent attentional modulation of LFP phase across tasks and visual brain structures, attention effects on LFP phase were broadly consistent across monkeys. Figure 4 illustrates differential phase/frequency curves for each monkey individually as well as curves for different trial periods and types. Significant negative slopes were observed for V1 (Fig. 4A, right) and LGN (Fig. 4B, right) LFPs measured during slow reaction time trials in two of three monkeys. Even when differential phase/frequency slopes for slow reaction time trials did not reach statistical significance, curves were similarly shaped across monkeys (Fig. 4, A and B, right) with slopes similar to those observed in grand averages (Fig. 3). Significant effects of attention on phase were also observed for cue period LFPs in all three monkeys and were most pronounced in the LGN (Fig. 4, A and B, left). Although not always significant, there was a trend whereby attention increased the magnitude of LGN LFP phase shifts during the cue period in the grand-average data for both types of tasks (Fig. 3, B–D, top right insets). This observation was noteworthy in light of results of a recent parallel study in which we demonstrated that attention facilitates feedforward and feedback communication between the LGN and V1 during the cue period (Mock et al. 2018). It is therefore possible that the attention-mediated phase shifts observed in the LGN during the cue reflect the relay of attention signals in early visual circuits.

To further test the notion that LFP phase shifts could serve as a consistent proxy for attention, we examined attentional modulation of LFP phase across the different layers in V1. For this analysis, we separated V1 LFPs according to their laminar compartment location. For each recording session, V1 electrodes were assigned to SG, G, or IG laminar compartments by determining their position relative to orthodromically stimulated geniculocortical recipient neurons in layer 4C (Briggs et al. 2013) or the border between layer 4C and layer 5 according to the first polarity reversal among average LFP responses to flashed stimuli (Maier et al. 2010, 2011) (Fig. 5A). LFPs recorded on V1 electrodes within each of the SG, G, or IG laminar compartments were then averaged together. Figure 5, B and C, illustrates differential phase/frequency curves for LFPs measured in the SG, G, and IG laminar compartments during the visual stimulus display period of slow reaction time trials in the CD and discrimination tasks, respectively. Differential phase/frequency slopes measured from LFPs in all three laminar compartments were significantly negative for the CD task (Fig. 5B; Table 3), and attention had a larger effect on slow reaction time trial LFPs measured in the SG and IG laminar compartments compared with other trial periods and types (Figs. 5B, inset; Table 3). Whereas attentional modulation of LFP phase was only significant for LFPs recorded in the IG layers in the discrimination tasks, LFPs measured from the other laminar compartments on slow reaction time trials were similarly sloped (Fig. 5C; Table 3). Overall, differential phase/frequency slopes calculated from LFPs recorded in each laminar compartment (approximately −6 to −13 ms) were within the same range as those observed for LGN LFPs and LFPs averaged across all V1 contacts. Together, these findings suggested that attention-mediated increases in the magnitude of LFP phase shifts were remarkably consistent throughout the structures and layers of the early visual system.

Fig. 5.

Differential phase/frequency curves across primary visual cortex (V1) laminar compartments. A: schematic of the V1 linear electrode array next to average local field potentials (LFPs) recorded in response to a flashed stimulus (vertical black line = flash onset; dashed line = layer 4C/5 border; purple and cyan traces around border thickened for visibility). B: average differential phase/frequency curves with linear regression fits (lines) for LFPs measured from the supragranular (SG, light gray; R² = 0.92, *P = 0.002), granular (G, gray; R² = 0.89, *P = 0.003), and infragranular (IG, black; R² = 0.91, *P = 0.002) laminar compartments of V1 during the visual stimulus display period for slow reaction time trials in the contrast change detection (CD) task (Table 3). Error bars are SE. Insets: average differential phase/frequency slopes for SG, G, and IG laminar compartments (38–43 sessions) from the cue, visual stimulus display period (all trials), and visual stimulus display period in which monkeys had fast or slow reaction times. Error bars are SE. *P < 0.011, Kruskal-Wallis tests (Table 3). C: average differential phase/frequency curves with linear regression fits (lines) for LFPs measured from the SG (R² = 0.81, P = 0.07), G (R² = 0.83, P = 0.06), and IG (R² = 0.93, *P = 0.02) laminar compartments of V1 in the discrimination tasks (Table 3), with conventions as in B. Error bars are SE. Insets show data from 33 sessions each.

Table 3.

Differential phase/frequency slopes for V1 LFPs measured in each laminar compartment

	Laminar Compartment SG				Laminar Compartment G				Laminar Compartment IG
	Cue period	Vis. Stim. Period (All trials)	Fast Reaction Time Trials	Slow Reaction Time Trials	Cue Period	Vis. Stim. Period (All trials)	Fast Reaction Time Trials	Slow Reaction Time Trials	Cue Period	Vis. Stim. Period (All trials)	Fast Reaction Time Trials	Slow Reaction Time Trials
CD task (Kruskal-Wallis tests used for all)
Average ± SE	−0.8 ± 1.7 ms	−2.8 ± 2.0 ms	−0.01 ± 2 ms	−10.6 ± 3.4 ms	−0.03 ± 1.7 ms	−1.4 ± 1.9 ms	+0.4 ± 2.1 ms	−8.3 ± 3.4 ms	−1.0 ± 1.8 ms	−1.3 ± 1.7 ms	−0.4 ± 1.7 ms	−11.4 ± 3.3 ms
No. of sessions	38				40				43
P value	0.02				0.1				0.008
Differences	Slow reaction time trials < cue and fast reaction time trials								Slow reaction time trials < fast reaction time trials
Discrimination task (Kruskal-Wallis tests used for all)
Average ± SE	+3.3 ± 2. ms	+3.0 ± 2.5 ms	+2.7 ± 2.9 ms	−5.8 ± 6.2 ms	+2.4 ± 1.9 ms	+3.4 ± 2.6 ms	+4.7 ± 3.0 ms	−6.8 ± 6.6 ms	+3.6 ± 2.4 ms	+2.7 ± 2.1 ms	+2.3 ± 2.7 ms	−12.7 ± 6.6 ms
No. of sessions	33				33				33
P value	0.5				0.11				0.09
Differences

CD, change detection; G, granular laminar compartment; IG, infragranular laminar compartment; LFPs, local field potentials; SG, supragranular laminar compartment; V1, primary visual cortex; Vis. stim., visual stimulation.

To test whether LFP phase shifts were unique to attention trials, we compared LFP phase shifts measured during attention trials with those measured during fixation-only trials. In fixation-only trials, monkeys were required to maintain fixation on a central fixation dot while drifting sinusoidal gratings were placed within the receptive fields of recorded neurons. We hypothesized that because monkeys were not instructed to attend to any particular location, LFP phase advances would not be observed on fixation-only trials. Consistent with our hypothesis, we observed a small phase lag, on average, for LFPs recorded during the visual stimulus display period of fixation-only trials (average LFP phase/frequency slope for fixation-only trials = +7.8 ± 1.1 ms). Furthermore, phase/frequency slopes for fixation-only and attention trials were significantly different from one another (P = 3.9 × 10⁻¹⁸, Wilcoxon test). These findings suggested that LFP phase advances were unique to trials in which monkeys allocated spatial attention.

Cross-correlation of LFP amplitude modulation.

Because attentional modulation of LFP phase was robust and consistent, we hypothesized that the attention effect could also be detected by cross-correlating LFP amplitude modulations measured for each attention condition. Accordingly, the attention-mediated temporal leads computed from the LFP phase analyses (approximately −10 ms) should be equivalent to the latency of the amplitude cross-correlation peak. We performed an independent assessment of attentional modulation of the relative timing of LFPs by cross-correlating the amplitude modulations of LFPs measured on attend-toward and attend-away trials (Adhikari et al. 2010; Catanese et al. 2016). Figure 6A illustrates LFPs measured during three example sessions separated by attention condition and fast or slow reaction time. Figure 6B illustrates the same LFPs bandpass filtered between 20 and 70 Hz. Qualitative observations of the filtered LFPs revealed a number of cycles in which attend-toward LFPs led attend-away LFPs by ~10 ms (dots in Fig. 6B); these forward shifts occurred more often during slow reaction time trials. Cross-correlating the amplitude modulations of filtered LFPs (attend-toward cross-correlated with attend-away) from fast and slow reaction time trials in the same example sessions revealed negative cross-correlation peaks for slow reaction time trials (Fig. 6C). The distributions of amplitude cross-correlation peak times for fast and slow reaction time trials were significantly different (P = 0.001, Wilcoxon text; Fig. 6D), with slow reaction time trials displaying an average amplitude cross-correlation peak time of −9.6 ± 2.7 ms. Cross-correlation peaks for fast and slow reaction time trials were similar across sessions from each monkey (P = 0.95, Kruskal-Wallis test). Thus, not only did the independent amplitude cross-correlation analysis reveal peaks with negative latencies, consistent with the observed attentional modulation of LFP phase, but cross-correlation peak times were equivalent to attention-mediated temporal lead values calculated from LFP phase analyses. Together, these independent analyses indicated that attention directed toward the stimulus in the receptive field caused a forward shift in high-β-to-γ-frequency LFPs corresponding to about −10 ms.

Fig. 6.

Local field potential (LFP) amplitude cross-correlation. A: 3 example sessions (left from monkey E, middle from monkey B, right from monkey O) showing average LFPs measured during the visual stimulus display period for attend-toward (red) and attend-away (blue) conditions on fast (top row) and slow (bottom row) reaction time trials. B: filtered LFPs (20–70 Hz) for the same example sessions, with conventions as in A. Red dots highlight cycles with small temporal leads for attend-toward filtered LFPs relative to attend-away filtered LFPs. C: cross-correlation of the filtered (20–70 Hz) LFP amplitude modulations of attend-toward and attend-away trials computed separately for slow (orange) and fast (purple) reaction time trials for the same example sessions. Dashed black lines illustrate amplitude cross-correlation peak times for the slow reaction time trials in each example session. D: distribution of amplitude cross-correlation peak times for fast (purple) and slow (orange) reaction time trials (25 sessions that met criteria, see materials and methods). Orange and purple vertical lines illustrate average amplitude cross-correlation peak times for slow (−9.6 ± 2.7 ms) and fast (+2.6 ± 1.4 ms) reaction time trials (*P = 0.001, Wilcoxon test).

Modulation of spike latency by attention and stimulus contrast.

We next explored the impact of shifting high-β-to-γ-frequency LFP oscillations forward by about −10 ms on spiking activity in V1. Some have suggested that spikes aligned with the phase of particular LFP oscillations contain more information and are more likely to be relayed from one cortical area to another (Fries et al. 2007; Vinck et al. 2010a; Womelsdorf et al. 2012). Alternatively, LFP phase advances could synchronize activity in downstream target areas in a latency-independent manner (Palmigiano et al. 2017). To test whether attention-mediated phase shifts for high-β-to-γ-frequency LFPs caused a change in the latencies of phase-aligned spikes, we calculated spike density functions for well-isolated, visually responsive V1 single-unit spike trains recorded simultaneously with V1 LFPs. Figure 7, top, illustrates spike density functions for a representative single unit generated from spikes aligned to either peaks (left) or troughs (right) of the LFP filtered at 20–70 Hz. As illustrated by both the example single unit and the sample population, attention generated a small but significant reduction in the latency of phase-aligned spikes (Fig. 7, bottom; peaks: P = 7.7 × 10⁻⁶; troughs: P = 0.02, Wilcoxon tests). Whereas phase-aligned spike latencies were reduced by ~0.5–1 ms with attention, spike onset latencies for all spikes, without phase alignment, were not altered by attention (P > 0.05, Wilcoxon test; for example PSTHs, see Fig. 2 of Hembrook-Short et al. 2017). Thus the small attention-mediated latency reduction for phase-aligned spikes, but not for unaligned spikes, provided support for the notion that attention prioritizes selective spikes through LFP phase alignment, perhaps to enhance the transmission of relevant visual information.

Fig. 7.

Latency shift with attention among phase-aligned spikes. Top: spike density functions for an example primary visual cortex (V1) single unit illustrating latencies for spikes aligned with peaks (left) or troughs (right) of the simultaneously recorded local field potential (LFP) bandpass filtered from 20 to 70 Hz. Separate spike density functions are illustrated for spikes recorded on attend-toward (red) and attend-away (blue) trials. Arrowheads mark latencies. Bottom: comparison of latencies on attend-toward and attend-away trials for spikes aligned with peaks (left; 110 single units) or troughs (right; 97 single units) of simultaneously recorded LFPs. Green stars illustrate average latencies across recorded single units (spikes aligned with peaks: average latency attend-toward = 2.31 ± 0.1 ms; average latency attend-away = 3.1 ± 0.14 ms; significant reduction in latency with attention: P = 7.7 × 10⁻⁶, Wilcoxon test; spikes aligned with troughs: average latency attend-toward = 2.4 ± 0.12 ms; average latency attend-away = 2.9 ± 0.16 ms; significant reduction in latency with attention: P = 0.02, Wilcoxon test).

The attention-mediated spike latency reduction we observed was quite small (~1 ms) but statistically significant, as reported previously (Sundberg et al. 2012). In comparison, we examined spike onset latencies in response to gratings varying in contrast for 60 V1 neurons recorded in the same monkeys. V1 neuronal response latencies were reduced by 14 ms, on average, in response to high-contrast compared with low-contrast stimuli, in line with previous reports (Albrecht 1995; Carandini and Heeger 1994; Dean and Tolhurst 1986; Kremers et al. 1997; Reid et al. 1992; Saul and Humphrey 1990; Sclar 1987). However, the contrast-dependent latency reduction we observed was not statistically significant (P = 0.1, Wilcoxon test). It is interesting to consider similarities between contrast-dependent and attentional modulation of visual cortical neurons (Reynolds and Heeger 2009) even though latency changes as a result of each are an order of magnitude different. In this study, we found qualitative similarities between attention-mediated temporal leads corresponding to LFP phase shifts and contrast-dependent latency reductions.

Attentional modulation of LFP amplitude.

In stark contrast to robust and consistent attentional modulation of LFP phase across early visual structures, attentional modulation of LFP amplitude was highly variable. Grand-average voltage traces for LFPs measured across V1 and in the LGN during the CD and discrimination tasks revealed little obvious attentional modulation of LFP amplitude (Fig. 8, A and C). Attentional modulation of LFP power varied across frequencies, tasks, structures, and trial periods (Fig. 8, B and D). Attention had little impact on V1 LFP evoked power across all trial periods of all tasks, although there were small but significant reductions in V1 LFP evoked power with attention in some frequency bands (Fig. 8, B and D, left). Attention directed toward the stimulus in the receptive field also suppressed LFP evoked power in the LGN across many frequency bands during the CD task (Fig. 8B, right), whereas few significant differences in LGN LFP evoked power were observed during the discrimination task (Fig. 8D, right). Overall, evoked power in LGN LFPs was small except during the cue period for the discrimination tasks, whereas V1 LFP evoked power was higher except during the visual stimulus display period of the discrimination tasks (Fig. 8, B and D).

Because LFP power is known to vary across the layers of V1 (Buffalo et al. 2011; Maier et al. 2010; van Kerkoerle et al. 2014; Xing et al. 2012), it is possible that averaging LFPs across all V1 contacts obscured attentional modulation of LFP power in V1. We therefore compared evoked power between LFPs recorded in the LGN and the SG, G, and IG laminar compartments in V1 to identify effects of attention and structure/layer on power. LFP evoked power in the LGN was significantly reduced with attention during both the cue and visual stimulus display periods of the CD task (Fig. 9A; Table 4), but LGN LFP evoked power was enhanced with attention in the θ+α-band during the cue period of the discrimination task (Fig. 9B, top left; Table 4). Evoked power in the θ+α-band was also reduced with attention in all laminar compartments in V1 during the CD task (Fig. 9A, top row; Table 4). Small reductions in evoked power with attention were also observed in the β-band for LFPs measured in the IG compartment during the visual stimulus display period of the CD task (Fig. 9A, middle right; Table 4) and in the γ-band for LFPs measured in the SG compartment during the cue period of the discrimination task (Fig. 9B, bottom left; Table 4). A small increase in evoked power with attention was observed in the γ-band for LFPs measured in the G compartment during the cue period of the discrimination task (Fig. 9B, bottom left; Table 4). Across both tasks, γ-band activity was suppressed during the visual stimulus display period and there was little attentional modulation of γ-band LFP evoked power (Fig. 9, A and B, bottom rows; Table 4). Additionally, LGN LFP evoked power was often significantly lower than V1 LFP evoked power in the CD task (Fig. 9A; Table 4), consistent with prior observations of reduced power at higher frequencies in the LGN (Bastos et al. 2014; Gray and Singer 1989). However, this was not the case for the discrimination task, where LGN LFP evoked power was sometimes significantly greater than evoked power in V1 laminar compartment LFPs (Fig. 9B; Table 4). Although we observed some laminar differences in evoked power, these were not consistent across frequency bands or task periods (Table 4). Overall, attentional modulation of LFP amplitude was largely inconsistent across frequency bands, structures/laminar compartments, and tasks.

Fig. 9.

Quantification of average power per structure/laminar compartment across attention conditions. A: average power (z-scored), quantified for θ+α (top row), β (middle row), and γ (bottom row) frequency bands, measured in the lateral geniculate nucleus of the thalamus (LGN; 25 sessions) and the supragranular (SG; 38 sessions), granular (G; 40 sessions), and infragranular (IG; 43 sessions) laminar compartments of primary visual cortex (V1) on attend-toward (red) and attend-away (blue) trials during the cue period (left) and the last 4 grating cycles of the visual stimulus display (Vis. stim.) period (right) in the change detection (CD) task. Error bars are SE. *P < 0.0001 (Kruskal-Wallis tests) across attention conditions (Table 4). Note differences in scale. B: average power (z-scored), quantified across frequency bands and structure/laminar compartment (33 sessions for all) on attend-toward and attend-away trials during the cue period and the visual stimulus display period in the discrimination tasks (Table 4). Conventions as in A; a.u., arbitrary units. *P < 0.0001, Kruskal-Wallis tests.

Table 4.

LFP power per structure/laminar compartment

	LGN		SG		G		IG
	Attend-toward	Attend-away	Attend-toward	Attend-away	Attend-toward	Attend-away	Attend-toward	Attend-away
CD task cue period (Kruskal-Wallis tests used for all)
θ + α
Average ± SE	−0.7 ± 0.2	2.8 ± 0.3	2.4 ± 0.3	2.9 ± 0.3	2.3 ± 0.3	2.9 ± 0.3	2.0 ± 0.2	2.6 ± 0.3
No. of sessions	25		38				40
P value	1.6 × 10–38
Differences (attention)	Attend-toward < Attend-away				Attend-toward < Attend-away		Attend-toward < Attend-away
Differences (structure/laminar)	No structure/laminar differences
β
Average ± SE	−0.5 ± 0.2	2.1 ± 0.2	2.8 ± 0.3	2.9 ± 0.3	2.8 ± 0.3	2.9 ± 0.3	2.4 ± 0.3	2.7 ± 0.3
No. of sessions	25		38		40		43
P value	1.9 × 10–35
Differences (attention)	Attend-toward < Attend-away
Differences (structure/laminar)	LGN < SG and G
γ
Average ± SE	2.1 ± 0.5	10.5 ± 0.7	24 ± 1.4	24.3 ± 1.3	23.2 ± 1.3	24.2 ± 1.4	22.3 ± 1.1	25.3 ± 1.3
No. of sessions	25		38		40		43
P value	9.7 × 10–115
Differences (attention)	Attend-toward < Attend-away
Differences (structure/laminar)	LGN < all V1; IG > all
CD task visual stimulus display period (Kruskal-Wallis tests used for all)
θ + α
Average ± SE	−1.8 ± 0.1	1.3 ± 0.2	4.0 ± 0.2	4.6 ± 0.2	4.0 ± 0.2	4.7 ± 0.2	3.7 ± 0.2	4.4 ± 0.2
No. of sessions	25		38		40		43
P value	1.0 × 10–501
Differences (attention)	Attend-toward < Attend-away		Attend-toward < Attend-away		Attend-toward < Attend-away		Attend-toward < Attend-away
Differences (structure/laminar)	LGN < all V1 laminar compartments
β
Average ± SE	−1.7 ± 0.1	1.2 ± 0.3	9.1 ± 0.3	10.0 ± 0.3	9.2 ± 0.3	10.1 ± 0.3	6.0 ± 0.2	7.4 ± 0.3
No. of sessions	25		38		40		43
P value	1.0 × 10–55
Differences (attention)	Attend-toward < Attend-away						Attend-toward < Attend-away
Differences (structure/laminar)	LGN < all V1; IG < SG and G
γ
Average ± SE	−9.9 ± 0.3	2.9 ± 0.9	−3.2 ± 0.7	−3.5 ± 0.5	−1.8 ± 0.6	−1.2 ± 0.6	−0.4 ± 0.6	2.1 ± 0.8
No. of sessions	25		38		40		43
P value	3.8 × 10–67
Differences (attention)	Attend-toward < Attend-away
Differences (structure/laminar)	IG > all V1 and LGN
Discrimination task cue period (Kruskal-Wallis tests used for all)
θ + α
Average ± SE	3.3 ± 0.3	2.6 ± 0.3	2.4 ± 0.3	2.4 ± 0.2	1.9 ± 0.2	2.0 ± 0.2	1.6 ± 0.3	1.4 ± 0.2
No. of sessions	33		33		33		33
P value	1.4 × 10–23
Differences (attention)	Attend-toward < Attend-away
Differences (structure/laminar)	LGN > IG
β
Average ± SE	5.2 ± 0.3	4.0 ± 0.2	4.6 ± 0.3	4.0 ± 0.2	4.6 ± 0.3	3.4 ± 0.2	3.7 ± 0.2	3.0 ± 0.2
No. of sessions	33		33		33		33
P value	1.2 × 10–70
Differences (attention)	No attentional modulation
Differences (structure/laminar)	LGN > all V1 laminar compartments
γ
Average ± SE	21.2 ± 1.6	20.4 ± 0.6	19.5 ± 0.9	19.7 ± 0.8	17.4 ± 0.9	16.7 ± 0.8	15.5 ± 0.8	14.2 ± 0.7
No. of sessions	33		33		33		33
P value	2.9 × 10–146
Differences (attention)			Attend-toward < Attend-away		Attend-toward > Attend-away
Differences (structure/laminar)	LGN > all V1 laminar compartments; SG > G and IG
Discrimination task visual stimulus display period (Kruskal-Wallis tests used for all)
θ + α
Average ± SE	−0.4 ± 0.1	0.2 ± 0.1	−0.8 ± 0.1	0.1 ± 0.2	−0.5 ± 0.1	−0.7 ± 0.1	−0.1 ± 0.1	−0.2 ± 0.1
No. of sessions	33		33		33		33
P value	9.7 × 10–14
Differences (attention)	No attentional modulation
Differences (structure/laminar)	IG > SG
β
Average ± SE	−1.7 ± 0.1	−1.6 ± 0.1	−2.1 ± 0.1	−1.9 ± 0.1	−2.1 ± 0.1	−2.2 ± 0.1	−1.5 ± 0.1	−1.4 ± 0.1
No. of sessions	33		33		33		33
P value	1.7 × 10–69
Differences (attention)	No attentional modulation
Differences (structure/laminar)	LGN > SG and G; IG > SG and G
γ
Average ± SE	−13.6 ± 0.3	−13.6 ± 0.2	−13.3 ± 0.2	−11.8 ± 0.3	−15.0 ± 0.2	−14.0 ± 0.3	−13.4 ± 0.2	−12.5 ± 0.3
No. of sessions	33		33		33		33
P value	1.8 × 10–44
Differences (attention)	No attentional modulation
Differences (structure/laminar)	G < LGN and SG and IG

CD, change detection; G, granular laminar compartment; IG, infragranular laminar compartment; LFPs, local field potentials; LGN, lateral geniculate nucleus of the thalamus; SG, supragranular laminar compartment; V1, primary visual cortex.

DISCUSSION

Although LFPs in the visual cortex contain rich information about visual stimulus features and attention state, it has been challenging to establish causal links between LFP modulations and attention or behavior. Our overarching goal in this study was to explore whether attention modulates LFPs in a consistent manner across visual brain areas. Regardless of whether or not LFPs are causal for perception or cognition, establishing direct links between specific and consistent LFP modulations and attention would signify that those LFP modulations could serve as a reliable marker for attention signals in the brain.

In alert monkeys performing attention-demanding visual discrimination tasks, we recorded LFPs from retinotopically aligned regions of the LGN and V1 (Fig. 1). We observed a small but significant correlation between LFP phase shifts in the high-β-to-γ-frequency bands and monkeys’ reaction times such that larger phase shifts predicted slower reaction times (Fig. 2B). Furthermore, LFP phase shifts were near zero and not correlated with reaction times on incorrect and miss trials in which monkeys were less engaged in the tasks (Fig. 2C). On fixation-only trials, LFP phase shifts were lagged and significantly different from leading phase shifts observed during attention trials. LFP phase shifts were also larger when monkeys successfully completed the most difficult discriminations (Fig. 3D). The latter finding is especially relevant in light of prior evidence that attentional modulation in V1 is larger when subjects successfully complete difficult trials (Chen et al. 2008). Together, these findings suggest that advances in LFP phase were unique to trials in which monkeys were engaged in the task and correctly allocated spatial attention. Although these data support links between LFP phase and task-relevant behavior, the fact that larger LFP phase shifts correlated with slower reaction times appeared at first counterintuitive. Although the same type of correlation has been observed in human subjects (Diederich et al. 2014; Naruse et al. 2015; Voloh et al. 2015), one might expect faster reaction times on trials in which subjects allocate more attention. Further examination of monkeys’ behavior revealed that all monkeys adopted speed-accuracy trade-off strategies in the attention tasks. In this context, behavioral and neurophysiological data fit with the idea that LFP phase shifts occurred when monkeys were engaged and applied focused attention to optimize accuracy over speed.

Our most auspicious observation was that LFP phase shifts in the high-β-to-γ-frequency bands were significantly greater on trials in which monkeys directed attention toward the stimulus in the receptive field compared with trials in which attention was directed away (Fig. 3). Differential phase/frequency curves illustrated the shift in LFP phase attributed to attention, corresponding to a constant temporal lead of ~10 ms. Consistent with the link between LFP phase shifts and reaction times, attention generated larger LFP phase shifts on trials in which monkeys had slower reaction times (Fig. 3, B and C). In the LGN, attentional modulation of LFP phase often also occurred during the cue period, perhaps reflecting the relay of attention signals between early visual structures (Mock et al. 2018). Attentional modulation of LFP phase was remarkably consistent across monkeys, tasks, and early visual structures including the LGN and the SG, G, and IG laminar compartments in V1 (Figs. 3–5). An independent analysis of LFP amplitude cross-correlations revealed cross-correlation peaks with the same latency as the attention-mediated constant temporal leads computed from the LFP phase analyses (Fig. 6), indicating that attention caused a robust and consistent temporal shift in high-β-to-γ-band LFPs. In addition to shifting LFP phase, attention also led to a small but significant reduction in the latency of phase-aligned spikes for V1 neurons recorded simultaneously with LFPs (Fig. 7).

In contrast to robust and consistent attentional modulation of LFP phase, attentional modulation of LFP amplitude was inconsistent (Figs. 8 and 9). We observed variable attentional modulation of γ-band LFP power in the LGN and across the laminar compartments of V1, often manifesting as a net reduction in γ-band power with attention, similar to prior findings (Bastos et al. 2014; Buffalo et al. 2011; Chalk et al. 2010; Gray and Singer 1989; Hansen and Dragoi 2011; Maier et al. 2010; van Kerkoerle et al. 2014; Xing et al. 2012). It is noteworthy that differential attentional modulation of γ-band LFP power across visual cortical areas that we and others observed is reminiscent of variations in attentional modulation of neuronal firing rate and spike-spike coherence across visual cortical areas (Buffalo et al. 2010; Luck et al. 1997; McAdams and Maunsell 1999; Nandy et al. 2017). Because visual cortical LFP power is strongly modulated by visual stimulation and weakly and/or variably modulated by attention, it may be worth reevaluating LFP amplitude as a readout for visual stimulus selectivity rather than attention.

Our findings suggest that attention generates phase shifts in high-β-to-γ-band LFPs such that attended task-relevant signals are shifted forward in time by ~10 ms. What is the advantage of these LFP phase shifts from a neuronal coding perspective? One possibility is that phase shifts, or phase resets, represent a pause that signifies a comparison between an internal prediction about the stimulus and the incoming perceptual information about the stimulus. Neurophysiological evidence supporting this theory suggests that lower frequency oscillations represent the prediction (or the cue) whereas higher frequency oscillations represent incoming sensory information, and a mismatch between prediction and perception leads to a phase shift (Bauer et al. 2014; Lakatos et al. 2008; Lima et al. 2011; Naruse et al. 2015; Samaha et al. 2015; Voloh et al. 2015). Our findings are in some ways consistent with this theory. For example, V1 LFP phase shifts in the high-β-to-γ-frequency band could reflect a comparison between a top-down prediction/cue signal and incoming stimulus information, in line with the recent finding that γ-band LFPs in the visual cortex are strongly driven by the visual stimulus but still susceptible to phase resetting by top-down prediction signals (Richter et al. 2017). Alternatively, our finding that LFP phase shifts correlate with slower reaction times is consistent with the notion of phase resets. Additional neuronal processing time required to compare prediction and perceptual signals could introduce delays or lengthen the time needed to accumulate adequate information about an ambiguous visual stimulus (Roitman and Shadlen 2002).

Although our behavioral data suggest that LFP phase shifts in V1 ultimately led to more careful evaluation of the stimulus and slower responses, neurophysiological data within V1 suggest that attention sped up the processing of task-relevant signals in the early visual pathways. Supporting the notion that LFP phase influences the timing of neuronal spiking activity (Cui et al. 2016; Gray and Singer 1989; Vinck et al. 2010a) and spiking consistency across cortical networks (Vinck et al. 2010b), we observed a reduction in the latency of spikes phase-aligned to high-β-to-γ-band LFPs with attention. We can further infer that attentional modulation of LFP phase could prioritize the transmission of spikes encoding task-relevant visual information, given that spikes phase-aligned to LFPs carry greater information about visual stimuli and attention state (Fries 2005; Fries et al. 2007; Ni et al. 2016; Roberts et al. 2013). Although the spike latency reduction we observed with attention was small, as reported previously (Sundberg et al. 2012), it is important to note that small reductions in spike latencies that are synchronized across neuronal ensembles can exert appreciable impact on downstream recipient neurons (Alonso et al. 1996). Finally, although it is tempting to assume that an LFP phase advance indicates a temporal advance, it is important to consider that LFP phase leads need not evoke shorter neuronal response latencies to influence downstream neuronal processing. Theoretical simulations show that coordinated LFP phase leads in one or more visual cortical structures can synchronize downstream target neurons in a latency-independent manner to improve visual perception (Palmigiano et al. 2017). One can envision alternative, but not mutually exclusive, scenarios in which 1) LFP phase shifts reflect a prediction/perception comparison that increases neuronal processing time to transmit and accumulate accurate information, or 2) LFP phase shifts align and synchronize spiking activity to speed up and/or boost the efficacy of task-relevant stimulus information. Future work will be required to elucidate the neuronal mechanisms involved and whether attentional modulation of LFP phase speeds up or slows down neuronal processing of task-relevant information.

In summary, we present compelling evidence that LFP phase shifts are predictive of task-relevant behavior and the allocation of focused visual spatial attention. Unlike attentional modulations of LFP amplitude, attentional modulations of LFP phase were reliable, robust, and consistent across monkeys, tasks, and visual brain areas. As such, LFP phase shifts could serve as a proxy for attention signals throughout the visual system. In contrast, other neurophysiological signals, such as spike rate, variability, or LFP amplitude, are not consistently modulated by attention. Moreover, LFPs have sufficient temporal resolution to provide information about attentional modulation with single-trial resolution, opening up the possibility of tracking attention signals, through LFP phase shifts, across visual brain areas on a behaviorally relevant timescale. In the present study we observed attention-mediated LFP phase shifts in the LGN during the cue period and in V1 during the visual stimulus display period, providing preliminary evidence that LFP phase shifts reflect the relay of attention signals in early visual circuits. Additionally, experimental manipulations of LFP phase may reveal whether LFPs themselves are causal for behavior, perception, or cognition.

GRANTS

This work was funded by National Institutes of Health Grants EY018683, EY013588, EY025219 (to F. Briggs) and EY023165 (to J. R. Hembrook-Short), National Science Foundation Established Program to Stimulate Competitive Research Grant 1632738, the Whitehall Foundation, and the Hitchcock Foundation. V. L. Mock was supported by a Graduate Fellowship from the Albert J. Ryan Foundation.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

F.B. conceived and designed research; V.L.M. and J.R.H.-S. performed experiments; V.L.M., K.L.L., and F.B. analyzed data; V.L.M., K.L.L., and F.B. interpreted results of experiments; V.L.M., K.L.L., and F.B. prepared figures; V.L.M. and F.B. drafted manuscript; V.L.M., K.L.L., J.R.H.-S., and F.B. edited and revised manuscript; V.L.M., K.L.L., J.R.H.-S., and F.B. approved final version of manuscript.