Dynamics and Constancy in Cortical Spatiotemporal Patterns of Orientation Processing

See allHide authors and affiliations

Science  18 Jan 2002:
Vol. 295, Issue 5554, pp. 512-515
DOI: 10.1126/science.1065916


How does the high selectivity to stimulus orientation emerge in the visual cortex? Thalamic feedforward-dominated models of orientation selectivity predict constant selectivity during the visual response, whereas intracortical recurrent models predict dynamic improvement in selectivity. We imaged the cat visual cortex with voltage-sensitive dyes to measure orientation-tuning dynamics of a large neuronal population. Tuning-curve width did not narrow after response onset, whereas the difference between preferred and orthogonal responses (modulation depth) first increased, then declined. We identified a suppression of the evoked responses, referred to as the evoked deceleration-acceleration (DA) notch, which was larger for the orthogonal response. Furthermore, peak selectivity of the tuning curves was contemporaneous with the evoked DA notch. These findings suggest that in the cat brain, sustained visual cortical processing does not narrow orientation tuning; rather, intracortical interactions may amplify modulation depth and suppress the orthogonal response relatively more than the preferred. Thus, feedforward models and recurrent models of orientation selectivity must be combined.

Visual cortical neurons are highly selective for the orientation of stimuli presented within their receptive field (1), a property not shared by their thalamic inputs (2). How orientation selectivity arises in the cortex is still debated. Previous experiments (3–10) have suggested mechanisms that include feedforward (thalamically dominated) (1, 11) and recurrent (intracortically dominated) (12–15) models. The input impinging on orientation-selective neurons has constant selectivity in feedforward models, whereas recurrent models predict improvement in selectivity during the visual response. Although single-unit methodologies excel at determining the properties of individual neurons, these properties are highly variable, making it extraordinarily difficult to obtain large samples on which to base estimates of neuronal population behavior. In optical imaging with voltage-sensitive dyes (16), the recorded signal accurately represents membrane-potential changes at the neuronal population level (17, 18), emphasizing synaptic potentials in the dendritic tufts of cortical neurons from superficial and deep layers. Recently, this method has been improved substantially, enabling in vivo imaging of cortical neuronal population activity with millisecond temporal resolution and spatial resolution of 50 to 100 μm (19). We therefore used optical imaging to explore the dynamics of orientation selectivity (9,20–22).

We imaged the responses of area 18 in the cat visual cortex to high-contrast square-wave gratings of six different orientations (23). We examined the recording period starting 50 ms before stimulus onset and lasting 300 ms (thus avoiding late intrinsic signal artifacts). Figure 1A shows the evoked response (24) to 150° gratings at a sampling rate of 9.6 ms per frame: the time series of single-condition responses (25), normalized to activity recorded during presentation of a blank screen. At 36 ms after stimulus onset, a brightening began (the dye fluorescence increased in response to depolarization) across the entire imaged area rather than only in the preferred orientation patches. Thus, this brightening signal had both an orientation-nonselective component and an orientation-selective component (Fig. 1, A and B).

Figure 1

Voltage-sensitive dye-imaging of the evoked-response dynamics. (A) Time series of single-condition maps: raw (unfiltered) response to a 150° grating stimulus, normalized to activity during the presentation of a blank isoluminant screen. Time (in milliseconds after stimulus onset) is indicated at the top right, with the orientation of the grating stimulus. The sampling rate is 9.6 ms per frame. Gray-scale values are 0 to 1.8‰ (parts in a thousand). The scale bar here and in all other figures is 1 mm. (B) Pixels within the patches delineated in (A) (rightmost frame) brighten more in response to a 150° grating (blue) than to a 60° grating (red). Thus, the 150° stimulus is preferred by neurons in these patches. The yaxis is the fractional response (ΔF/F, ‰) (the brightness of pixels relative to the resting fluoresence level). The two dots highlight the evoked DA notch (see text). (C) Orientation tuning curve, for the pixels delineated in (A) (rightmost frame), at the time marked by the two dots in (B), with the best-fitting Gaussian. HWHH, MD, and NSC are marked by arrows. The NSC here is considerably larger than that normally reported in simple cells. However, the optical signal is dominated by the superficial layers, in which the more common neurons are complex cells. In these cells, the intracellular NSC is considerably larger than in simple cells—in a few cells it was up to 67%, although in others it was almost absent (39). Furthermore, the optical signal emphasizes activity in dendritic tufts that spread laterally, beyond the borders of orientation domains of neuronal somata.

We examined the orientation-selective response by fitting a Gaussian to the tuning curve (26) at every point in time (27), as exemplified in Fig. 1C for the data frame at 65 ms, for pixels within the delineated regions of Fig. 1A (rightmost frame). Preferred orientation is obtained from the peak angle of the Gaussian. Selectivity is determined by three additional independent attributes of the tuning curve, obtained from the Gaussian fit (27): half width at half height (HWHH), modulation depth (MD, preferred – orthogonal), and nonselective component (NSC). We combined these attributes into a single selectivity index (28).

We analyzed preferred orientation, obtained from pixel-by-pixel Gaussian fitting (26, 27) from the same hemisphere shown in Fig. 1. Figure 2A shows a sequence of polar maps (29) in which preferred orientation and MD are represented by color and brightness, respectively. After the map emerged at 36 ms, its strength increased until 74 ms. Before one can draw any conclusions from these maps, it is important first to examine their reproducibility by comparing two independent subsets of stimulus presentations. Before response onset, the apparent “orientation map” was not reproducible, whereas it reached near-maximal reproducibility at 46 ms, ∼10 ms after response onset (18). All subsequent analysis was done exclusively for the pixels in the high-reproducibility region indicated by the dotted line (in Fig. 2C).

Figure 2

Dynamics of orientation maps. (A) Time course of the polar orientation map. Colors represent the preferred orientation of each pixel (ranging from 0° to 180°; bottom to top, respectively, of the color scale, on the right), and brightness represents the MD of each pixel's tuning curve (ranging from 0 to 0.5‰; left to right of the color scale). After peaking at 74 ms, map strength declines gradually to ∼65% of the maximal value, at 120 ms (not shown). (B) Temporal stability of orientation preference: correlation between each frame and the map averaged over a set of later, independent frames [shown in (C)], for pixels within the delineated region in (C). The gray line is the 99% confidence limit of prestimulus level, based on the 18 frames before stimulus onset (one-tailed t test). (C) Polar orientation map constructed by averaging over responses at 257 to 526 ms. The high-reproducibility region is delineated.

Qualitative visual inspection of Fig. 2A suggested that preferred orientation at each cortical location was stable over time, because the color at each pixel did not change appreciably between frames in the time sequence. This was quantified by comparing the map at each frame to a high signal-to-noise ratio map. Figure 2B demonstrates the correlation coefficient between preferred orientation in each frame and in this average map (Fig. 2C). As soon as the response was fully reproducible (46 ms, close to the time of the first cortical spikes) (18), the correlation coefficient stabilized at ∼0.8, and the change in preferred orientation between successive frames had a median of 4.4°, well within the effect of residual noise in the images. Having observed the same result in two additional hemispheres, we concluded that as soon as the signal was reproducible, preferred orientation was constant as a function of time.

We next calculated the evolution in time of orientation selectivity for the entire population by Gaussian fitting of the tuning curve averaged over all pixels delineated in Fig. 2C, on a frame-by-frame basis.Figure 3A shows the tuning curves and their fits for a few frames for the same hemisphere shown in Figs. 1and 2. To facilitate comparison, we removed the offset of the tuning curves (NSC) so that they all started at the same baseline value of 0. These tuning curves suggest that as the response develops, MD increases, but tuning width does not change appreciably. We used two different approaches to assess the reliability of the Gaussian fits and the parameters derived from them (18). The response for early data points was small and indistinguishable from background noise (marked in red and orange). The first frames that showed highly reliable fits are marked in green throughout Fig. 3. For the three experiments shown in Fig. 3, a reliable fit was obtained already in the first or second frame after onset of the evoked responses (e.g., Fig. 1B).

Figure 3

Dynamics of orientation selectivity. (A) Tuning curves (mean ± SE) were averaged over pixels in the reproducible region (Fig. 2C) for several frames, the times of which are indicated on the right. The hemisphere used was the same as that in Figs. 1 and 2. (Insets). Two additional experiments: the frames shown are the same as those in the main panel, with modifications of frame color notation shown in the small boxes. The blue- and magenta-colored insets in (A) to (D) show the same two experiments. Frame color notations are constant for each experiment; red and orange curves mark the last frames before reliable fits were obtained, and green curves mark the first frames with reliable fits. For the two experiments shown in the insets, the first frame of the response showed an intermediate significance level, marked in yellow (18). (B) Half-width at half-height (HWHH) obtained from a Gaussian fit. Ninety-nine percent confidence intervals (40) are shown in light and dark gray shading, for frames with reliable and unreliable fits, respectively; after response onset they were between 8° and 15° wide. (C) Amplitude (MD) of the Gaussian fit. Ninety-nine percent confidence intervals are shown in gray. (D) Selectivity index (28). We found that MD normalized by orthogonal, preferred, or [orthogonal + preferred] (33), as well as 1-[circular variance] (15), all exhibit an extremely similar time course to the selectivity index (not shown).

Figure 3B shows the time course of orientation tuning width. HWHH of the fitted Gaussian started at 35.5° already in the first frame with an acceptable fit, at 36 ms. It then widened gradually to ∼38° during about the next 60 ms. For the three experiments, the narrowest initial values of HWHH range from 33° to 36°, and late values range from 36° to 38°. This range is well within that found for the intracellular response averaged over hundreds of milliseconds (30–33). Therefore, the measurement of tuning dynamics averaged over several populations of thousands of neurons did not show an improvement in tuning width (narrowing) as cortical response continued to increase (Fig. 1B).

Modulation depth is presented in Fig. 3C as the amplitude of the fitted Gaussian. It increased until 74 ms after stimulus onset (during the first 50 ms of the response), after which it decreased to a relatively stable level of ∼65% of maximal response during the course of ∼40 ms. This behavior was consistent in all three hemispheres, with a late MD of 65 to 90% of the peak value. Moreover, because MD is the difference between preferred and orthogonal responses, it can be assessed in experiments in which only two orientations were presented, and these results were confirmed in the other 10 hemispheres in this study (18), in which only two stimulus orientations were used. Taken together, we concluded that MD peaked tens of milliseconds after cortical response began, rather than at its onset.

To evaluate overall selectivity, we determined the selectivity index (Fig. 3D) (28). After response onset, selectivity increased and peaked at 55 ms, decaying to about one-third of its peak value at ∼150 ms. Because HWHH is fairly constant throughout the response (Fig. 3B), this temporal profile was affected mainly by the ratio between the selective component (MD) and the nonselective one. This normalized MD (preferred – orthogonal/orthogonal) peaked at 55 to 100 ms over the additional 10 hemispheres in this study (18).

What mechanisms are involved in creating the peak in selectivity observed 55 to 100 ms after response onset? One possibility is suggested by the finding that the selectivity index (as well as normalized MD) peaked simultaneously with a small transient drop in the rate at which the evoked response increased. This deceleration and subsequent acceleration, which we term the evoked deceleration-acceleration (DA) notch, is marked by two dots in Fig. 1B and suggests a temporary suppression: The response first slowed down, then sped up. Peak selectivity (Fig. 3D) was attained at the same time, at 55 ms. The evoked DA notch was detected in 12 of the 13 hemispheres examined. Because the evoked DA notch is rather small, we investigated whether it was present in existing in vivo intracellular recordings in area 18 of the cat visual cortex (34). Not surprisingly, we detected the notch in the averaged intracellular recording of the visually evoked response.

Enlargements of two optically detected evoked DA notches are shown in the red and green insets of Fig. 4 [for the complete set of hemispheres, see (18)]. In these 12 hemispheres, peak normalized MD (or peak selectivity index, in the two hemispheres where six orientations were presented) occurred within 10 ms of the acceleration at the end of the evoked DA notch, with a correlation coefficient between the two events of 0.83 (P < 0.001, two-tailed t test). A comparison between the evoked DA notch at the preferred and orthogonal orientations (Fig. 4, insets) (18) showed that, although the magnitude of the evoked DA notch varied among experiments, it was always stronger for the orthogonal orientation than for the preferred. Comparison of the evoked DA notch index (35) at the orthogonal and preferred orientations is shown in Fig. 4 for the 12 hemispheres in which an evoked DA notch was detected. A larger notch index corresponds to a greater deflection of the response. Equal values lie on the dashed line, so in all cases the orthogonal stimulus induced a larger evoked DA notch than the preferred stimulus (P< 0.01, one-tailed paired t test). Therefore, analysis of the evoked DA notch suggests that suppression, which has a stronger relative influence when the orthogonal orientation is presented, coincides with maximal selectivity.

Figure 4

The evoked DA notch. The orthogonal notch index (35) is larger than the preferred notch index for all 12 hemispheres. The green cross is from the experiment shown in Figs. 1 through 3 (main panels); the blue cross is from the experiment shown in the blue insets in Fig. 3. Insets: The evoked response around notch occurrence. Green inset, as described above; red inset, an additional experiment (also shown in the main panel). The thick trace represents the response to preferred orientation, and the thin trace the response to orthogonal orientation. See also (18).

The data presented here imply that sustained cortical processing does not narrow tuning width and is not required to establish preferred orientation at a given cortical location (36). However, the dynamic time course of MD suggests that cortical interactions are involved in determining the amplification of the tuning curve; both intracortical excitation (12–14) and inhibition (7–11, 15, 37) may be involved. We identified an evoked DA notch in the evoked response at ∼50 to 80 ms, which we interpret as the peaking of a suppressive mechanism, simultaneous with peak normalized (MD). Questions remain regarding the orientation selectivity of the underlying synaptic mechanism and whether it results from increased inhibition or from withdrawal of excitation. Shunting inhibition has been shown to peak as early as 50 to 70 ms (38), and in light of evidence for the involvement of cross-orientation inhibition in orientation selectivity (7, 10), we favor the involvement of inhibition. However, the present results suggest that the inhibition observed in these previous studies should be involved in amplifying rather than narrowing the tuning curve—increasing its normalized MD by preventing the response to the orthogonal orientation from increasing as rapidly as the response to the preferred orientation.

  • * To whom correspondence should be addressed. E-mail: dahlia.sharon{at}


View Abstract

Stay Connected to Science

Navigate This Article