Attentional Modulation of Neuronal Activity Depends on Neuronal Feature Selectivity

Abstract

Attention exerts a powerful influence on visual perception. The impact of attention on neuronal activity manifests at early visual information processing stages and progressively increases throughout the visual cortical hierarchy. However, the neuronal mechanisms of attention are unresolved. In particular, the rules governing attentional modulation of individual neurons, whether they are facilitated by or suppressed by attention, are not known. To obtain a more granular or neuron- and circuit-level understanding of the mechanisms of attention and to directly test the feature similarity gain model in V1, we compared attentional modulation with neuronal feature selectivity across a large population of V1 neurons in alert and behaving macaque monkeys trained on an attention-demanding contrast-change detection task. We utilized emerging multi-electrode array technology to record simultaneously from V1 neurons spanning all six cortical layers so that we could characterize the laminar position and physiological response properties of diverse V1 neuronal populations. We found significant relationships between attentional modulation and neuronal position within the cortical hierarchy, neuronal physiology, and neuronal feature selectivity. Our results support the feature similarity gain model and further suggest that attentional modulation depends critically upon the match between neuronal feature selectivity and the features required for the task.

Keywords: V1; feature attention; macaque monkey; spatial attention.

Figure 1. Attention task and behavioral results

(A) Schematic screen shots of the contrast-change detection task. A red central fixation dot (left) cued the monkey to attend to the drifting sinusoidal grating in the lower hemifield while a blue fixation dot (right) cued the monkey to attend to the grating in the upper hemifield. Dashed circle (not displayed in the actual task) indicates receptive field location of recorded neurons. The timeline for a trial is indicated far right. (B) Accuracy, as average percent correct, for each of the 3 monkeys on validly cued (attend-toward and attend-away conditions in red and blue, respectively) and invalidly cued (grey) trials. Asterisks indicate significant reductions in accuracy on invalidly cued trials (Monkey B: n=95 sessions, p=1.9×10⁻¹⁰; Attend-toward=82±1%, Attend-away=82±1%, Invalid=63±3%; Monkey O: n=23 sessions, p=0.007, Attend-toward=65±4%, Attend-away=68±5%, Invalid=25±13%; Monkey E: n=40 sessions, p=0.001, Attend-toward=70±3%, Attend-away=71±3%, Invalid=45±6%).

Figure 2. Example single-unit recordings and attentional modulation of neuronal firing rate

(A) Schematic representation of the linear array positioned to record from V1 neurons spanning the cortical layers. Average local field potential (LFP) responses to a flashed stimulus in the receptive field (flash onset indicated by vertical black line). Horizontal dashed line highlights the polarity reversal at LFP response onset, indicative of the layer 4C/5 border. See Figure S1 for receptive field locations of recorded neurons. (B) Current source density spectrum generated from LFPs traces in A. Vertical black line again indicates flash onset at time zero. Horizontal dashed lines indicate borders of laminar compartments: supragranular (SG), granular (G), and infragranular (IG). (C) Schematic representation of linear array with contacts assigned to SG (green), G (orange), and IG (purple) laminar compartments. Upper right images shows two single-unit waveforms (a simple [cyan] and a complex [magenta] neuron) isolated from one SG contact and the principal components analysis clustering of 1^st and 2^nd principal components of the same waveforms (both waveforms are significantly different from each other and from the noise [grey cluster] at p=5.7×10⁻³¹; waveform SNRs were 9.2 and 4.8). Lower right image shows the inter-spike-interval (ISI) distributions for the two waveforms and noise. Red dashed lines illustrate short-ISI cut-off and red bar above noise ISI distribution illustrates a high percentage of short ISI violations (3.6%) compared to 0% short ISI violation for the cyan and magenta waveforms. (D) Peri-stimulus time histograms (PSTHs) for 11 simultaneously recorded neurons spanning the SG (green), G (orange), and IG (purple) laminar compartments (contact number color-coded by laminar compartment at right) from a single session. Average spike count (in 1msec bins) prior to and following grating onset at time zero is illustrated separately for attend-toward (red) and attend-away (blue) trials. (E) Additional example PSTHs for 4 neurons, 3 facilitated (top three examples) and 1 suppressed (bottom example) by attention. PSTHs illustrate spike counts prior to and following grating onset at time zero on attend-toward (red) and attend-away (blue) trials. AI values for each neuron are indicated along with laminar compartment location and simple/complex type. Black curves underneath PSTHs illustrate differential modulation by attention (attend-toward minus attend-away), with attentional modulation during the grating display period illustrated by black fills, using the same scaling as above PSTHs. (F) Average firing rate for simple neurons (top; n=74) and complex neurons (bottom; n=94) just prior to and during grating presentation (grating onset at time zero) on attend-toward (red) and attend-away (blue) trials. Shaded regions represent SEMs.

Figure 3. Attentional modulation of firing rate across neuronal types and laminar locations

(A) Distributions of attention index (AI) values for simple neurons (white/light grey; n=74) and complex neurons (black/dark grey; n=94). Grey tones indicate neurons that were significantly modulated by attention (n=6 complex neurons; n=11 simple neurons). Average simple neuron AI=−0.004±0.006 (dashed line) and average complex neuron AI=0.03±0.005 (solid line). (B) AI values were significantly greater for complex neurons compared to simple neurons (asterisk, p=0.00017). Error bars represent SEMs. (C) Average AI values for simple (open bars) and complex (solid bars) neurons separated by laminar compartment location (supragranular, SG in green; granular, G, in orange; infragranular, IG, in purple). Asterisk indicates statistically significant differences in attentional modulation of firing rate across neuronal types and laminar locations (p=0.0046). Specifically, complex neurons in the G and IG laminar compartments were significantly more modulated by attention compared to simple neurons in the SG and IG laminar compartments. SG simple: n=26, average AI=−0.008±0.009; SG complex: n=37, average AI=0.018±0.007; G simple: n=28, average AI=0.005±0.01; G complex: n=25, average AI=0.029±0.008; IG simple: n=20, average AI=−0.012±0.013; IG complex: n=32, average AI=0.037±0.012.

Figure 4. Tuning curves for 6 representative V1 neurons

Each row illustrates orientation, contrast, spatial frequency, temporal frequency, and size tuning curves for a single V1 neuron with laminar compartment location, simple/complex type, and AI value listed above the middle (spatial frequency) plot. Data are dots and error bars represent SEMs. Lines are curve fits (Gaussian fits for orientation and size tuning data; power fits for contrast tuning data; and smoothing spline fits for spatial and temporal frequency data).

Figure 5. Relationships between neuronal feature selectivity and attentional modulation of firing rate

See Table 1 for full regression statistics corresponding to each plot. (A) Negative relationship, indicated by dashed line illustrating regression fit, between AI and c50, or contrast to evoke a half-maximal response, for V1 neurons (simple - open dots, complex - closed dots; R²=0.03, p=0.04, n=135). (B) Positive relationship, indicated by dashed line illustrating regression fit, between AI and direction selectivity index (DSI) for V1 neurons (R²=0.02, p=0.04, n=168). (C) Significant interaction between AI, DSI, and c50 illustrated by polynomial fit and 3-way regression model (R²=0.04, p=0.016, n=143) for V1 neurons (black dots). (D–F) Significant relationships (solid lines illustrate regression fits) were observed between AI and orientation HWHH for SG complex neurons (D, R²=0.11, p=0.04, n=37), between AI and DSI for G complex neurons (E, R²=0.32, p=0.006, n=22), and between AI and SSI for G complex neurons (F, R²=0.22, p=0.04, n=18). (G–I) Relationships (solid lines illustrate regression fits, R²<=0.1, p>=0.2 for all) between AI and orientation HWHH for G and IG complex neurons (G), DSI for SG and IG complex neurons (H), and SSI for SG and IG complex neurons (I). See Figure S2 and Table S1 for all other comparisons.

Figure 6. Attentional modulation of variance and contrast sensitivity among V1 neurons

(A) Average Fano Factor across attention conditions (red – attend-toward, blue – attend-away) for simple (open bars) and complex (closed bars) neurons located in each laminar compartment (x-axis labels). No differences in Fano Factor were observed across neuronal types, laminar compartments, or attention conditions (p=0.6; SG simple: n=26, attend-toward FF=3.1±0.6, attend-away FF=3.9±0.8; G simple: n=28, attend-toward FF=2.8±0.6, attend-away FF=2.9±0.6; IG simple: n=20, attend-toward FF=2.8±0.4, attend-away FF=2.7±0.4; SG complex: n=37, attend-toward FF=3.5±0.5, attend-away FF=3.3±0.5; G complex: n=25, attend-toward FF=3.8±0.7, attend-away FF=3.8±0.7; IG complex: n=32, attend-toward FF=2.3±0.3, attend-away FF=2.1±0.3). However, Fano Factors were lowest among IG complex neurons. (B) Average spike count correlations for 1239 total pairs of simultaneously recorded V1 neurons across attention conditions (p=0.19; average spike count correlations attend-toward=0.05±0.008; average spike count correlations attend-away=0.06±0.008). (C) Average c50 values for simple (open bars) and complex (closed bars) neurons across laminar compartments. There were no differences in contrast sensitivity across neuronal types or laminar compartments (p=0.5; SG simple: n=26, average c50=23±5%; SG complex: n=35, average c50=15±3%; G simple: n=12, average c50=11±2%; G complex: n=20, average c50=12±3%; IG simple: n=17, average c50=10±1%; IG complex: n=31, average c50=18±4%). (D) Significant positive relationship (black line illustrates regression fit; R²=0.2, p=0.007) between neuronal contrast discriminability (ROC AUC) and AI for 33 simple (open dots) and complex (closed dots) neurons.