Spatial attention improves the quality of population codes in human visual cortex

Abstract

Selective attention enables sensory input from behaviorally relevant stimuli to be processed in greater detail, so that these stimuli can more accurately influence thoughts, actions, and future goals. Attention has been shown to modulate the spiking activity of single feature-selective neurons that encode basic stimulus properties (color, orientation, etc.). However, the combined output from many such neurons is required to form stable representations of relevant objects and little empirical work has formally investigated the relationship between attentional modulations on population responses and improvements in encoding precision. Here, we used functional MRI and voxel-based feature tuning functions to show that spatial attention induces a multiplicative scaling in orientation-selective population response profiles in early visual cortex. In turn, this multiplicative scaling correlates with an improvement in encoding precision, as evidenced by a concurrent increase in the mutual information between population responses and the orientation of attended stimuli. These data therefore demonstrate how multiplicative scaling of neural responses provides at least one mechanism by which spatial attention may improve the encoding precision of population codes. Increased encoding precision in early visual areas may then enhance the speed and accuracy of perceptual decisions computed by higher-order neural mechanisms.

Fig. 1.

Schematic depiction of the possible types of attention-induced scaling in the neuronal tuning function and, consequently, in population response profiles. A: multiplicative scaling or feature-dependent increase in response amplitude, where the increase depends on the proximity of the attended stimulus feature to the preferred feature of the neuron. B: additive scaling or feature-nonspecific increase in response amplitudes. C: bandwidth scaling or change in SD of response profile.

Fig. 2.

A: schematic of the task performed by subjects during functional magnetic resonance imaging (fMRI) scanning. Subjects attended to a flickering (2 Hz) sinusoidal grating that was rendered in one of 8 possible orientations [0, 22.5, 45, … , 157.5°]. Subjects were required to continuously fixate on the spot at the center of the display and to attend to either the left or the right stimulus based on a small central cue (B). Subjects pressed a button whenever they detected a slight dimming of the attended stimulus and ignored an equally probable dimming of the unattended stimulus. Target contrast decrements were titrated to maintain detection accuracy at about 75–80%. Stimuli in both hemifields were rendered in the same orientation to negate the influence of global feature-based attentional modulations (see methods). C: independent localizer scans were used to identify the most spatially selective voxels in V1–hV4. The stimulus was similar to that used in the attention task (above) except that it was visible in only one hemifield, to which the subject had to direct his/her attention.

Fig. 3.

Voxel tuning functions (VTFs) with attention (solid curve) and without attention (dashed curve) based on responses in V1, V2v, V3v, and hV4. These mean tuning functions were produced by centering all VTFs (for a visual area) at their preferred orientation and then averaging across subjects (lines represent best-fitting circular Gaussian; see methods). Error bars reflect ±1SE across subjects. The y-axis refers to the magnitude of the fit coefficients (beta weights) estimated using the general linear model (GLM; see methods).

Fig. 4.

Additive scaling parameter of best-fitting Gaussian function in areas (V1–hV4). Error bars reflect ±1SE across subjects.

Fig. 5.

A: tuning functions for top 25% (red) and bottom 25% (blue) of voxels in V1 ranked by MI score, with attention (solid curves) and without attention (dotted curves). Tuning functions of high-MI voxels (red lines; also see B) show significant multiplicative and additive scaling, whereas low-MI voxels show primarily additive scaling (blue lines; also see C and Figs. 6 and 7). Note that B and C have been derived from A, to better highlight the difference in shape between the VTFs from high-MI and low-MI voxels (with y-axis of both B and C covering equal range to permit a direct comparison). Error bars reflect ±1SE across subjects.

Fig. 6.

Additive scaling with attention for each MI quartile in V1, V2v, V3v, and hV4. Left panels show additive scaling parameter derived separately from attended data and unattended data. Right panels show the net change (attended minus unattended) in the parameter with attention. Top row shows data averaged across all visual areas; remaining rows show data from each visual area. Error bars reflect ±1SE across subjects.

Fig. 7.

Multiplicative scaling with attention for each MI quartile in V1, V2v, V3v, and hV4. Left panels show multiplicative scaling parameter derived separately from attended data and unattended data. Right panels show the net change (attended minus unattended) in the parameter with attention. Top row shows data averaged across all visual areas; remaining rows show data from each visual area. Error bars reflect ±1SE across subjects.

Fig. 8.

A: additive scaling. B: multiplicative scaling. C: bandwidth scaling (attended minus unattended) for voxels in V1 ranked by their MI score. Each point on the x-axis depicts an aggregation of top x% of voxels ranked by their MI score. Corresponding y-axis depicts the mean additive scaling, multiplicative scaling, and bandwidth scaling with attention for mean voxel tuning function of that group. Red color indicates data points that reached significance by repeated measures t-test (P < 0.05). Data points for which mean root mean square error (RMSE) of fit (across subjects) deviated >2SDs from overall RMSE mean (across all subjects and data points) were excluded; this was only an issue for the smallest aggregation of high-MI voxels (< top 3%).

Fig. 9.

Ratio of normalized MI (see methods) for voxels in V1, V2v, V3v, and hV4 between attended and unattended stimuli. Voxels were sorted by normalized MI based on unattended responses only. The abscissa marks different groups of top voxels (by percentage of the total voxels ranked by their MI score). /* refers to normalized MI (see methods) derived from unattended responses and attended responses, respectively. Ordinate refers to their ratio. Shaded gray patch around data points highlights the SD of values between subjects. High-MI voxels showed higher increase in overall information content with attention.