Thank you for visiting Nature.com. The browser version you are using has limited CSS support. For the best experience, we recommend that you use an updated browser (or disable Compatibility Mode in Internet Explorer). In the meantime, to ensure continued support, we will render the site without styles and JavaScript.
A carousel showing three slides at the same time. Use the Previous and Next buttons to move through three slides at a time, or use the slider buttons at the end to move through three slides at a time.
High-resolution vision requires fine retinal sampling and integration to reconstruct object properties. It is important to note that when mixing local samples from different objects, accuracy is lost. Therefore, segmentation, the grouping of areas of an image for separate processing, is critical to perception. In previous work, bistable lattice structures, which can be considered as one or more moving surfaces, were used to study this process. Here, we report the relationship between activity and segmentation judgments in intermediate regions of the primate visual pathway. In particular, we found that selectively orienting median temporal neurons were sensitive to texture cues used to distort the perception of bistable gratings and showed a significant correlation between trials and subjective perception of persistent stimuli. This correlation is greater in units that signal global movement in patterns with multiple local directions. Thus, we conclude that the intermediate time domain contains signals used to separate complex scenes into constituent objects and surfaces.
Vision relies not only on the precise discrimination of elementary image characteristics such as edge orientation and speed, but more importantly on the correct integration of these characteristics to calculate environmental properties such as object shape and trajectory1. However, problems arise when retinal images support several equally plausible feature groups 2, 3, 4 (Fig. 1a). For example, when two sets of speed signals are very close, this can reasonably be interpreted as one moving object or several moving objects (Fig. 1b). This illustrates the subjective nature of segmentation, i.e. it is not a fixed property of the image, but a process of interpretation. Despite its obvious importance for normal perception, our understanding of the neural basis of perceptual segmentation remains incomplete at best.
A cartoon illustration of a perceptual segmentation problem. An observer’s perception of depth in a Necker cube (left) alternates between two possible explanations (right). This is because there are no signals in the image that allow the brain to uniquely determine the figure’s 3D orientation (provided by the monocular occlusion signal on the right). b When multiple motion signals are presented in spatial proximity, the vision system must determine whether the local samples are from one or more objects. The inherent ambiguity of local motion signals, i.e. a sequence of object motions can produce the same local motion, resulting in multiple equally plausible interpretations of the visual input, i.e. vector fields here can arise from the coherent motion of a single surface or the transparent motion of overlapping surfaces. c (left) An example of our textured grid stimulus. Rectangular gratings drifting perpendicular to their direction (“component directions” – white arrows) overlap each other to form a grating pattern. The lattice can be perceived as a single, regular, connected movement of directions (red arrows) or a transparent movement of compound directions. The perception of the lattice is distorted by the addition of random point texture cues. (Middle) The area highlighted in yellow is expanded and displayed as a series of frames for coherent and transparent signals, respectively. The movement of the dot in each case is represented by green and red arrows. (Right) Graph of the position (x, y) of the selection point versus the number of frames. In the coherent case, all textures drift in the same direction. In the case of transparency, the texture moves in the direction of the component. d A cartoon illustration of our motion segmentation task. The monkeys began each trial by fixing a small dot. After a short delay, a certain type of grating pattern (coherence/transparency) and texture signal size (eg contrast) appeared at the location of the MT RF. During each test, the grating may drift in one of two possible directions of the pattern. After stimulus withdrawal, selection targets appeared above and below MT RF. The monkeys must indicate their perception of the grid in saccades to the appropriate selection target.
The processing of visual movements is well characterized and thus provides an excellent model for studying the neural circuits of perceptual segmentation. Several computational studies have noted the utility of two-stage motion processing models in which high-resolution initial estimation is followed by selective integration of local samples to remove noise and restore object velocity7,8. It is important to note that vision systems must take care to limit this ensemble to only those local samples from ordinary objects. Psychophysical studies have described physical factors that influence how local motion signals are segmented, but the shape of anatomical trajectories and neural codes remain open questions. Numerous reports suggest that orientation-selective cells in the temporal (MT) region of the primate cortex are candidates for neural substrates.
Importantly, in these previous experiments, changes in neural activity correlated with physical changes in visual stimuli. However, as mentioned above, segmentation is essentially a perceptual process. Therefore, the study of its neural substrate requires linking changes in neural activity with changes in the perception of fixed stimuli. Therefore, we trained two monkeys to report whether the perceived bistable grating pattern formed by superimposed drifting rectangular gratings was a single surface or two independent surfaces. To study the relationship between neural activity and segmentation judgments, we recorded a single activity in the MT when the monkeys performed this task.
We found a significant correlation between the study of MT activity and perception. This correlation existed whether or not the stimuli contained overt segmentation cues. In addition, the strength of this effect is related to sensitivity to segmentation signals, as well as to the pattern index. The latter quantifies the degree to which the unit radiates global rather than local movement in complex patterns. Although selectivity for fashion direction has long been recognized as a defining feature of MT, and fashion-selective cells show tuning to complex stimuli consistent with human perception of those stimuli, to the best of our knowledge, this is the first evidence for a correlation between patterns. index and perceptual segmentation.
We trained two monkeys to demonstrate their perception of drifting grid stimuli (coherent or transparent movements). Human observers usually perceive these stimuli as coherent or transparent movements of approximately the same frequency. In order to give the correct answer in this trial and set the basis for the operant reward, we created segmentation signals by texturing the raster of the component that forms the lattice (Fig. 1c, d). Under coherent conditions, all textures move along the direction of the pattern (Fig. 1c, “coherent”). In the transparent state, the texture moves perpendicular to the direction of the grating on which it is superimposed (Fig. 1c, “transparent”). We control the difficulty of the task by changing the contrast of this texture label. In the cued trials, monkeys were rewarded for responses corresponding to texture cues, and rewards were offered randomly (50/50 odds) in trials containing patterns without texture cues (zero texture contrast condition).
Behavioral data from two representative experiments are shown in Figure 2a, and responses are plotted as a proportion of judgments of coherence versus contrast of texture signals (transparency contrast is assumed to be negative by definition) for patterns that shift up or down, respectively. Overall, the monkeys’ perception of coherence/transparency was reliably affected by both the sign (transparent, coherent) and strength (contrast) of the texture cue (ANOVA; monkey N: direction – F = 0.58, p = 0.45, sign – F = 1248, p < 10−10, contrast – F = 22.63, p < 10;−10 monkey S: direction – F = 0.41, p = 0.52, sign – F = 2876.7, p < 10−10, contrast – F = 36.5, p < 10−10). Overall, the monkeys’ perception of coherence/transparency was reliably affected by both the sign (transparent, coherent) and strength (contrast) of the texture cue (ANOVA; monkey N: direction – F = 0.58, p = 0.45, sign – F = 1248, p < 10−10, contrast – F = 22.63, p < 10; −10 monkey S: direction – F = 0.41, p = 0.52, sign – F = 2876.7, p < 10−10, contrast – F = 36.5, p < 10−10). В целом на восприятие обезьянами когерентности/прозрачности достоверно влияли как знак (прозрачность, когерентность), так и сила (контрастность) текстурного признака (ANOVA; обезьяна N: направление — F = 0,58, p = 0,45, знак — F = 1248, p < 10−10, контраст – F = 22,63, p < 10; −10 обезьяна S: направление – F = 0,41, p = 0,52, признак – F = 2876,7, p < 10−10, контраст – F = 36,5, р < 10-10). In general, the perception of coherence/transparency by monkeys was significantly affected by both the sign (transparency, coherence) and the strength (contrast) of the textural feature (ANOVA; monkey N: direction — F = 0.58, p = 0.45, sign — F = 1248, p < 10−10, contrast – F = 22.63, p < 10; -10 monkey S: direction – F = 0.41, p = 0.52, sign – F = 2876.7, p < 10 −10, contrast – F = 36.5, p < 10-10).总体而言，猴子对连贯性/透明度的感知受到纹理提示（ANOVA）的符号（透明、连贯）和强度（对比度）的可靠影响；猴子N：方向- F = 0.58，p = 0.45，符号- F = 1248, p < 10−10, 对比度– F = 22.63, p < 10;−10 猴子S: 方向– F = 0.41, p = 0.52, 符号– F = 2876.7, p < 10−10, 对比度– F = 36.5，p < 10-10）。总体而言，猴子对连贯性/透明度的感知受到纹理提示（ANOVA）的符号（透明、连贯）和强度（对比度）的可靠影响；猴子N：方向- F = 0.58，p = 0.45，符号- F = 1248, p < 10−10, 对比度– F = 22.63, p < 10;−10 36.5，p < 10-10）。 In general, the perception of monkey coherence/transparency was significantly affected by the sign (transparency, coherence) and intensity (contrast) of texture signals (ANOVA); обезьяна N: ориентация – F = 0,58, p = 0,45, знак – F = 1248, p < 10−10, Контрастность — F = 22,63, p < 10; monkey N: orientation – F = 0.58, p = 0.45, sign – F = 1248, p < 10−10, Contrast – F = 22.63, p < 10; −10 Обезьяна S: Ориентация — F = 0,41, p = 0,52, Знак — F = 2876,7, p < 10−10, Контрастность — F = 36,5, p < 10-10). −10 Monkey S: Orientation – F = 0.41, p = 0.52, Sign – F = 2876.7, p < 10-10, Contrast – F = 36.5, p < 10-10). Gaussian cumulative functions were fitted to the data from each session to characterize the psychophysical characteristics of the monkeys. On fig. 2b shows the distribution of agreement for these models over all sessions for both monkeys. Overall, the monkeys completed the task accurately and consistently, and we rejected less than 13% of the two-monkey sessions due to poor fit to the cumulative Gaussian model.
a Behavioral examples of monkeys in representative sessions (n ​​≥ 20 trials per stimulus condition). In the left (right) panels, data from one N(S) monkey session are plotted as coherent selection scores (ordinate) versus sign contrast of texture signals (abscissa). Here it is assumed that transparent (coherent) textures have negative (positive) values. Responses were built separately according to the direction of movement of the pattern (up (90°) or down (270°)) in the test. For both animals, performance, whether the response is divided by 50/50 contrast (PSE – solid arrows) or the amount of textural contrast required to support a certain level of performance (threshold – open arrows), is in these drift directions. b Fitted histogram of R2 values ​​of the Gaussian cumulative function. Monkey S(N) data are shown on the left (right). c (Top) PSE measured for the grid shifted down (ordinate) compared to the PSE shifted up the grid (abscissa) plotted, with edges representing the PSE distribution for each condition and arrows indicating the mean for every condition. Data for all N(S) monkey sessions are given in the left (right) column. (Bottom) Same convention as for PSE data, but for fitted feature thresholds. There were no significant differences in PSE thresholds or fashion trends (see text). d PSE and slope (ordinate) are plotted depending on the normalized raster orientation of the angular separation component (“integral grating angle” – abscissa). The open circles are the means, the solid line is the best fitting regression model, and the dotted line is the 95% confidence interval for the regression model. There is a significant correlation between PSE and normalized integration angle, but not slope and normalized integration angle, suggesting that the psychometric function shifts as the angle separates the component lattices, but not sharpening or flattening. (Monkey N, n = 32 sessions; Monkey S, n = 43 sessions). In all panels, the error bars represent the standard error of the mean. Haha. Coherence, PSE subjective equality score, norm. standardization.
As noted above, both the contrast of texture cues and the direction of movement of the pattern varied across trials, with stimuli drifting up or down in a given trial. This is done to minimize psychophysical11 and neuronal28 adaptive effects. Pattern orientation versus bias (subjective equality point or PSE) (Wilcoxon rank sum test; monkey N: z = 0.25, p = 0.8; monkey S: z = 0.86, p = 0.39) or fitted function threshold (sum of Wilcoxon ranks; monkey N: z = 0.14, p = 0.89, monkey S: z = 0.49, p = 0.62) (Fig. 2c). In addition, there were no significant differences between monkeys in the degree of texture contrast required to maintain performance threshold levels (N monkey = 24.5% ± 3.9%, S monkey = 18.9% ± 1.9%; Wilcoxon rank sum , z = 1.01, p = 0.31).
In each session, we changed the interlattice angle separating the orientations of the component lattices. Psychophysical studies have shown that people are more likely to perceive cell 10 as connected when this angle is smaller. If the monkeys were reliably reporting their perception of coherence/transparency, then based on these findings, one would expect PSE, the texture contrast corresponding to a uniform split between coherence and transparency choices, to increase upon interaction. lattice angle. This was indeed the case (Fig. 2d; collapsing across pattern directions, Kruskal–Wallis; monkey N: χ2 = 23.06, p < 10−3; monkey S: χ2 = 22.22, p < 10−3; correlation between normalized inter-grating angle and PSE – monkey N: r = 0.67, p < 10−9; monkey S: r = 0.76, p < 10−13). This was indeed the case (Fig. 2d; collapsing across pattern directions, Kruskal–Wallis; monkey N: χ2 = 23.06, p < 10−3; monkey S: χ2 = 22.22, p < 10−3; correlation between normalized inter- grating angle and PSE – monkey N: r = 0.67, p < 10−9; monkey S: r = 0.76, p < 10−13). Это действительно имело место (рис. 2d; коллапс поперек направления паттерна, Крускал-Уоллис; обезьяна N: χ2 = 23,06, p < 10–3; обезьяна S: χ2 = 22,22, p < 10–3; корреляция между нормализованными угол решетки и PSE – обезьяна N: r = 0,67, p < 10-9, обезьяна S: r = 0,76, p < 10-13). This indeed occurred (Fig. 2d; collapse across the direction of the pattern, Kruskal-Wallis; monkey N: χ2 = 23.06, p < 10–3; monkey S: χ2 = 22.22, p < 10–3; correlation between normalized lattice angle and PSE – monkey N: r = 0.67, p < 10-9, monkey S: r = 0.76, p < 10-13).情况确实如此（图2d；跨模式方向折叠，Kruskal-Wallis；猴子N：χ2 = 23.06，p < 10-3；猴子S：χ2 = 22.22，p < 10-3；标准化间光栅角和PSE – 猴子N：r = 0.67，p < 10-9；猴子S：r = 0.76，p < 10-13）。情况 确实 如此 （图 图 2D ； 方向 折叠 ， kruskal-wallis ； n ： ：2 = 23.06 ， p <10-3 ； 猴子 ： ：2 = 22.22 ， p <10-3 ； 间 光栅角 和 pse-猴子 猴子 猴子 猴子 猴子 猴子 猴子N：r = 0.67，p < 10-9；猴子S：r = 0.76，p < 10-13）。 Это действительно имело место (рис. 2d; кратность по оси моды, Крускал-Уоллис; обезьяна N: χ2 = 23,06, p < 10-3; обезьяна S: χ2 = 22,22, p < 10-3; нормализованный межрешеточный угол). This was indeed the case (Fig. 2d; fold along the mode axis, Kruskal-Wallis; monkey N: χ2 = 23.06, p < 10-3; monkey S: χ2 = 22.22, p < 10-3; normalized interlattice corner). PSE-обезьяна N: r = 0,67, p < 10–9, обезьяна S: r = 0,76, p < 10–13). PSE monkey N: r = 0.67, p < 10–9, monkey S: r = 0.76, p < 10–13). In contrast, changing the interlattice angle had no significant effect on the slope of the psychometric function (Fig. 2d; cross-modal orientation fold, Kruskal-Wallis; monkey N: χ2 = 8.09, p = 0.23; monkey S χ2 = 3.18 , p = 0.67, correlation between normalized interlattice angle and slope – monkey N: r = -0.4, p = 0.2, monkey S: r = 0.03, p = 0.76). Thus, in accordance with the psychophysical data of a person, the average effect of changing the angle between the gratings is a shift in displacement points, and not an increase or decrease in sensitivity to segmentation signals.
Finally, rewards are randomly assigned with a probability of 0.5 in trials with zero texture contrast. If all monkeys were aware of this unique randomness and were able to distinguish between zero texture contrast and cue stimuli, they could develop different strategies for the two types of trials. Two observations strongly suggest that this is not the case. First, changing the grating angle had qualitatively similar effects on cue and zero texture contrast scores (Fig. 2d and Supplementary Fig. 1). Second, for both monkeys, the bistable trial selection is unlikely to be a repeat of the most recent (previous) reward selection (binomial test, N monkeys: 0.52, z = 0.74, p = 0.22; S monkeys: 0.51 , r = 0.9, p = 0.18).
In conclusion, the behavior of the monkeys in our segmentation task was under good stimulus control. The dependence of perceptual judgments on the sign and size of texture cues, as well as changes in PSE with grating angle, indicate that the monkeys reported their subjective perception of motor coherence/transparency. Finally, the responses of the monkeys in the zero texture contrast trials were not affected by the reward history of previous trials and were significantly affected by inter-raster angular changes. This suggests that the monkeys continue to report their subjective perception of the lattice surface configuration under this important condition.
As mentioned above, the transition of texture contrast from negative to positive is equivalent to the perceptual transition of stimuli from transparent to coherent. In general, for a given cell, the MT response tends to increase or decrease as the texture contrast changes from negative to positive, and the direction of this effect usually depends on the direction of movement of the pattern/component. For example, the directional tuning curves of two representative MT cells are shown in Figure 3 along with the responses of these cells to gratings containing low or high contrast coherent or transparent texture signals. We have tried in some way to better quantify these grid responses, which could be related to the psychophysiological performance of our monkeys.
Polar plot of the directional tuning curve of a representative monkey MT cell S in response to a single sinusoidal array. The angle indicates the direction of movement of the grating, the magnitude indicates the emissivity, and the preferred cell direction overlaps by about 90° (up) with the direction of one of the components in the direction of the grating pattern. b Weekly stimulus-time histogram (PSTH) of the response grid, shifted in the template direction by 90° (shown schematically on the left) for the cell shown in a. Responses are sorted by texture hint type (cohesive/transparent – middle/right panel respectively) and Michelson contrast (PSTH color hint). Only correct attempts are shown for each type of low-contrast and high-contrast texture signals. Cells responded better to upward drifting lattice patterns with transparent texture cues, and the response to these patterns increased with increasing texture contrast. c, d are the same conventions as in a and b, but for MT cells other than monkey S, their preferred orientation nearly overlaps that of the downward drifting grid. The unit prefers downward drifting gratings with coherent texture cues, and the response to these patterns increases with increasing texture contrast. In all panels, the shaded area represents the standard error of the mean. Spokes. Spikes, seconds. second.
To explore the relationship between lattice surface configuration (coherent or transparent) as indicated by our texture signals and MT activity, we first regressed the correlation between cells for coherent movement (positive slope) or transparent movement (negative slope) by regression. given to classify cells by sign response rate compared to contrast (separately for each mode direction). Examples of these lattice tuning curves from the same example cell in Figure 3 are shown in Figure 4a. After classification, we used receiver performance analysis (ROC) to quantify each cell’s sensitivity to modulation of texture signals (see Methods). The neurometric functions obtained in this way can be directly compared with the psychophysical characteristics of the monkeys in the same session in order to directly compare the psychophysical sensitivity of neurons to lattice textures. We performed two signal detection analyzes for all units in the sample, calculating separate neurometric features for each direction of the pattern (again, up or down). It is important to note that, for this analysis, we only included trials in which (i) the stimuli contained a texture cue and (ii) the monkeys responded in accordance with that cue (i.e., “correct” trials).
Fire rates are plotted against texture sign contrast, respectively, for gratings shifting up (left) or down (right), the solid line represents the best fit linear regression, and the data in the top (bottom) row are taken from those shown in Fig. . Rice. 3a Cell, b (Fig. 3c, d). Regression slope features were used to assign preferred texture cues (coherent/transparent) to each cell/lattice orientation combination (n ≥ 20 trials per stimulus condition). Error bars represent the standard deviation of the mean. The neurometric functions of the units shown in ba are described together with the psychometric functions collected during the same session. Now, for each feature, we plot the preferred tooltip choice (ordinate) (see text) as a percentage of the texture’s sign contrast (abscissa). Texture contrast has been changed so that preferred tooltips are positive and blank tooltips are negative. The data from the upward (downward) drifting grids are shown in the left (right) panels, in the upper (lower) rows – the data from the cells shown in Fig. 3a,b (Fig. 3c,d). The ratios of neurometric and psychometric threshold (N/P) are shown in each panel. Spokes. Spikes, seconds. sec, directory. direction, province preferred, psi. Psychometry, Neurology.
Lattice tuning curves and neurometric functions of two representative MT cells and their associated psychometric functions, aggregated together with these responses, are shown in the top and bottom panels of Figure 4a,b, respectively. These cells show a roughly monotonous increase or decrease as the hint of texture goes from transparent to coherent. In addition, the direction and strength of this bond depend on the direction of lattice motion. Finally, the neurometric functions calculated from the responses of these cells only approached (but still did not correspond to) the psychophysical properties of the unidirectional grid movement. Both neurometric and psychometric functions were summarized with thresholds, i.e. corresponding to about 84% of a correctly chosen contrast (corresponding to the mean + 1 sd of the fitted cumulative Gaussian function). Across the entire sample, the N/P ratio, the ratio of the neurometric threshold to the psychometric one, averaged 12.4 ± 1.2 in monkey N and 15.9 ± 1.8 in monkey S, and for the lattice to move in at least one direction, only At ~16% (18). %) units from monkey N (monkey S) (Fig. 5a). From the cell example shown in the figure. As seen in Figures 3 and 4, the sensitivity of neurons can be affected by the relationship between the preferred orientation of the cell and the direction of lattice movement used in the experiments. In particular, the orientation adjustment curves in Fig. 3a,c demonstrate the relationship between the neuron orientation setting of a single sinusoidal array and its sensitivity to transparent/coherent motion in our textured array. This was the case for both monkeys (ANOVA; relative preferred directions binned with 10° resolution; monkey N: F = 2.12, p < 0.01; monkey S: F = 2.01, p < 0.01). This was the case for both monkeys (ANOVA; relative preferred directions binned with 10° resolution; monkey N: F = 2.12, p < 0.01; monkey S: F = 2.01, p < 0.01). Это имело место для обеих обезьян (ANOVA; относительные предпочтительные направления объединены в группы с разрешением 10°; обезьяна N: F = 2,12, p <0,01; обезьяна S: F = 2,01, p <0,01). This was the case for both monkeys (ANOVA; relative preferred directions grouped at 10° resolution; monkey N: F=2.12, p<0.01; monkey S: F=2.01, p<0.01) .两只猴子都是这种情况（方差分析；以10° 分辨率合并的相对首选方向；猴子N：F = 2.12，p < 0.01；猴子S：F = 2.01，p < 0.01）。两 只 猴子 都 是 这 种 （方差 分析 以 以 10 ° 分辨率 合并 的 相对 方向 ； 猴子 n ： f = 2.12 ， p <0.01 ； ： ： f = 2.01 ， p <0.01。。。。。。。）））））））））））））））））））））））））） Это имело место для обеих обезьян (ANOVA; относительная предпочтительная ориентация объединена при разрешении 10°; обезьяна N: F = 2,12, p <0,01; обезьяна S: F = 2,01, p <0,01). This was the case for both monkeys (ANOVA; relative preferred orientation pooled at 10° resolution; monkey N: F=2.12, p<0.01; monkey S: F=2.01, p<0.01). Given the large degree of variability in neuron sensitivity (Fig. 5a), in order to visualize the dependence of neuron sensitivity on relative preferred orientations, we first normalized each cell’s preferred orientation to the “best” orientation for the movement of the grid pattern (i.e. direction). in which the grating forms the smallest angle between the preferred cell orientation and the orientation of the grating pattern). We found that the relative thresholds of neurons (threshold for “worst” lattice orientation/threshold for “best” lattice orientation) varied with this normalized preferred orientation, with peaks in this threshold ratio occurring around patterns or component orientations (Figure 5b). )). This effect could not be explained by a bias in the distribution of preferred directions in the units in each sample towards one of the plaid pattern or component directions (Fig. 5c; Rayleigh test; monkey N: z = 8.33, p < 10−3, circular mean = 190.13 deg ± 9.83 deg; monkey S: z = 0.79, p = 0.45) and was consistent across plaid inter-grating angles (Supplementary Fig. 2). This effect could not be explained by a bias in the distribution of preferred directions in the units in each sample towards one of the plaid pattern or component directions (Fig. 5c; Rayleigh test; monkey N: z = 8.33, p < 10−3 , circular mean = 190.13 deg ± 9.83 deg; monkey S: z = 0.79, p = 0.45) and was consistent across plaid inter-grating angles (Supplementary Fig. 2). Этот эффект нельзя было объяснить смещением распределения предпочтительных направлений в единицах в каждой выборке в сторону одного из клетчатых направлений или направлений компонентов (рис. 5в; критерий Рэлея; обезьяна N: z = 8,33, p < 10–3). This effect could not be explained by a shift in the distribution of preferred directions in units in each sample towards one of the checkered directions or component directions (Fig. 5c; Rayleigh test; monkey N: z = 8.33, p < 10–3). , circular mean = 190.13 degrees ± 9.83 degrees; monkey S: z = 0.79, p = 0.45) and was the same for all corners of the plaid grid (Supplementary Figure 2).这种效应不能通过每个样本中单元中的优选方向分布偏向格子图案或组件方向之一来解释（图5c；瑞利测试；猴子N：z = 8.33，p < 10-3 ，圆形平均值= 190.13 度± 9.83 度；猴子S：z = 0.79，p = 0.45）并且在格子间光栅角上是一致的（补充图2）。这 种 效应 不 能 通过 每 样本 中 单元 中 优选 方向 分布 偏向 偏向 图案 或 组件 方向 来 解释 （图 图 图 图 瑞利 测试 ； 猴子 n ： z = 8.33 ， p <10-3 ， 平均值 平均值 圆形 圆形 圆形 圆形 圆形 圆形 圆形 圆形 圆形 圆形 圆形z Этот эффект не может быть объяснен тем, что распределение предпочтительных ориентаций в клетках в каждом образце смещено либо в сторону структуры решетки, либо в сторону одной из ориентаций компонентов (рис. 5в; критерий Рэлея; обезьяна N: z = 8,33, p < 10–3). This effect cannot be explained by the fact that the distribution of preferred orientations in cells in each sample is biased either towards the lattice structure or towards one of the component orientations (Fig. 5c; Rayleigh’s test; monkey N: z = 8.33, p < 10–3). , circular average) = 190.13 degrees ± 9.83 degrees; monkey S: z = 0.79, p = 0.45) and were equal in lattice angles between grids (Supplementary Fig. 2). Thus, the sensitivity of neurons to textured grids depends, at least in part, on the fundamental properties of MT tuning.
The left panel shows the distribution of N/P ratios (neuron/psychophysiological threshold); each cell provides two data points, one for each direction in which the pattern moves. The right panel plots psychophysical thresholds (ordinate) versus neuronal thresholds (abscissa) for all units in the sample. The data in the top (bottom) row are from monkey N (S). b Normalized threshold ratios are plotted against the magnitude of the difference between the optimal lattice orientation and the preferred cell orientation. The “best” direction is defined as the direction of the grating structure (measured with a single sinusoidal grating) closest to the preferred cell direction. Data were first binned by normalized preferred orientation (10° bins), then threshold ratios were normalized to the maximum value and averaged within each bin. Cells with a preferred orientation slightly larger or smaller than the orientation of the lattice components had the largest difference in sensitivity to the orientation of the lattice pattern. c Pink histogram of the preferred orientation distribution of all MT units recorded in each monkey.
Finally, the response of the MT is modulated by the direction of the grating movement and the details of our segmentation signals (texture). Comparison of neuronal and psychophysical sensitivity showed that, in general, MT units were much less sensitive to contrasting texture signals than monkeys. However, the sensitivity of the neuron changed depending on the difference between the preferred orientation of the unit and the direction of the grid movement. The most sensitive cells tended to have orientational preferences that almost covered the lattice pattern or one of the constituent orientations, and a small subset of our samples were as sensitive or more sensitive than monkeys’ perception of contrast differences. To determine whether signals from these more sensitive units were more closely associated with perception in monkeys, we examined the correlation between perception and neuronal responses.
An important step in establishing a connection between neural activity and behavior is to establish correlations between neurons and behavioral responses to constant stimuli. In order to link neural responses to segmentation judgments, it is critical to create a stimulus that, despite being the same, is perceived differently in different trials. In the present study, this is explicitly represented by a zero texture contrast grating. Although we emphasize that, based on the psychometric functions of animals, gratings with minimal (less than ~20%) textural contrast are usually considered coherent or transparent.
To quantify the extent to which MT responses correlate with perceptual reports, we performed a selection probability (CP) analysis of our grid data (see 3). In short, CP is a non-parametric, non-standard measure that quantifies the relationship between spike responses and perceptual judgments30. Restricting the analysis to trials using grids with zero textural contrast and sessions in which monkeys made at least five choices for each type of these trials, we calculated SR separately for each direction of grid movement. Across monkeys, we observed a mean CP value significantly greater than we would expect by chance (Fig. 6a, d; monkey N: mean CP: 0.54, 95% CI: (0.53, 0.56), two-sided t-test against null of CP = 0.5, t = 6.7, p < 10−9; monkey S: mean CP: 0.55, 95% CI: (0.54, 0.57), two-sided t-test, t = 9.4, p < 10−13). Across monkeys, we observed a mean CP value significantly greater than we would expect by chance (Fig. 6a, d; monkey N: mean CP: 0.54, 95% CI: (0.53, 0.56), two-sided t-test against null of CP = 0.5, t = 6.7, p < 10−9; monkey S: mean CP: 0.55, 95% CI: (0.54, 0.57), two-sided t-test, t = 9.4, p < 10−13) . In monkeys, we observed a mean CP significantly higher than randomly expected (Fig. 6a, d; monkey N: mean CP: 0.54, 95% CI: (0.53, 0.56), two-tailed t-test vs. null values). CP = 0,5, t = 6,7, p < 10–9; CP = 0.5, t = 6.7, p < 10–9; обезьяна S: среднее CP: 0,55, 95% ДИ: (0,54, 0,57), двусторонний t-критерий, t = 9,4, p < 10–13) . monkey S: mean CP: 0.55, 95% CI: (0.54, 0.57), two-tailed t-test, t = 9.4, p < 10–13).在猴子中，我们观察到平均CP 值显着大于我们偶然预期的值（图6a，d；猴子N：平均CP：0.54，95% CI：（0.53，0.56），针对空值的双边t 检验CP = 0.5, t = 6.7, p < 10−9;猴子S: 平均CP: 0.55, 95% CI: (0.54, 0.57), 双边t 检验, t = 9.4, p < 10−13) .在 猴子 中 ， 我们 观察 平均 平均 值 显着 大于 我们 偶然 的 值 （图 图 图 6a ， d ； n ： 平均 ： 0.54，95% Ci ： 0.53，0.56） ， 空值 检验 CP = 0.5, t = 6.7, p < 10−9; 猴子S: 平均CP: 0.55, 95% CI: (0.54, 0.57), 双边t检验, t=9.4, p < 10−13) У обезьян мы наблюдали средние значения CP, значительно превышающие то, что мы могли бы ожидать случайно (рис. 6a, d; обезьяна N: среднее CP: 0,54, 95% ДИ: (0,53, 0,56), двусторонний t- тест CP против нуля = 0,5, t = 6,7, p < 10-9, обезьяна S: средний CP: 0,55, 95% ДИ: (0,54, 0,57), двусторонний t-критерий, t = 9,4, p < 10- 13) . In monkeys, we observed mean CP values ​​well above what we might expect by chance (Fig. 6a, d; monkey N: mean CP: 0.54, 95% CI: (0.53, 0.56), two-tailed t-test CP vs. zero = 0.5, t = 6.7, p < 10-9, monkey S: mean CP: 0.55, 95% CI: (0.54, 0.57), two-tailed t- criterion, t = 9.4, p < 10-13). Thus, MT neurons tend to fire more strongly even in the absence of any overt segmentation cues, when the animal’s perception of the lattice motion matches the cell’s preferences.
a Selection probability distribution for grids without texture signals for samples recorded from monkey N. Each cell can contribute up to two data points (one for each direction of grid movement). A mean CP value above random (white arrows) indicates that overall there is a significant relationship between MT activity and perception. b To examine the impact of any potential selection bias, we calculated CP separately for any stimulus for which the monkeys made at least one error. The selection probabilities are plotted as a function of the selection ratio (pref/null) for all stimuli (left) and the absolute values ​​of texture mark contrast (right, data from 120 individual cells). The solid line and shaded area in the left panel represent the mean ± sem of the 20-point moving average. The selection probabilities calculated for stimuli with unbalanced selection ratios, such as grids with high signal contrast, differed more and were clustered around possibilities. The gray-shaded area in the right panel emphasizes the contrast of the features included in the calculation of the high selection probability. c The probability of a large choice (ordinate) is plotted against the threshold of the neuron (abscissa). The probability of selection was significantly negatively correlated with threshold. The convention df is the same as ac but applies to 157 single data from monkey S unless otherwise noted. g Highest selection probability (ordinate) is plotted against the normalized preferred direction (abscissa) for each of the two monkeys. Each MT cell contributed two data points (one for each direction of the lattice structure). h Large box plots of selection probability for each inter-raster angle. The solid line marks the median, the lower and upper edges of the box represent the 25th and 75th percentiles respectively, the whiskers are extended to 1.5 times the interquartile range, and outliers beyond this limit are noted. The data in the left (right) panels are from 120 (157) individual N(S) monkey cells. i Highest probability of selection (ordinate) is plotted against the time of onset of the stimulus (abscissa). Large CP was calculated in sliding rectangles (width 100 ms, step 10 ms) throughout the test and then averaged over units.
Some previous studies have reported that CP depends on the relative number of trials in the basal rate distribution, meaning this measure is less reliable for stimuli that cause large differences in the proportion of each choice. To test this effect in our data, we calculated the CP separately for all stimuli, regardless of the sign texture contrast, and the monkeys performed at least one false trial. The CP is plotted against the selection ratio (pref/null) for each animal in Figure 6b and e (left panel), respectively. Looking at the moving averages, it’s clear that CP remains above probability over a wide range of selection odds, declining only when the odds fall (increase) below (above) 0.2 (0.8). Based on the psychometric characteristics of animals, we would expect selection coefficients of this magnitude to apply only to stimuli with high-contrast texture cues (coherent or transparent) (see examples of psychometric features in Fig. 2a, b). To determine whether this was the case and whether a significant PC persisted even for stimuli with clear segmentation signals, we examined the effect of absolute textural contrast values ​​on PC (Fig. 6b, e-right). As expected, CP was significantly higher than the probability for stimuli containing up to moderate (~20% contrast or lower) segmentation cues.
In orientation, speed, and mismatch recognition tasks, MT CP tends to be highest in the most sensitive neurons, presumably because these neurons carry the most informative signals30,32,33,34. Consistent with these findings we observed a modest but significant correlation between grand CP, calculated from z-scored firing rates across the texture cue contrasts highlighted in the rightmost panel of Fig. 6b, e, and neuronal threshold (Fig. 6c, f; geometric mean regression; monkey N: r = −0.12, p = 0.07 monkey S: r = −0.18, p < 10−3). 6b, e, and neuronal threshold (Fig. 6c, f; geometric mean regression; monkey N: r = −0.12, p = 0.07 monkey S: r = −0.18, p < 10−3). Consistent with these findings, we observed a modest but significant correlation between the large CP calculated from the excitation frequency z-score from the texture signal contrasts highlighted in the rightmost panel of Fig. 6b, e, and neuronal threshold (Fig. 6c, f; geometric). geometric mean regression; обезьяна N: r = -0,12, p = 0,07 обезьяна S: r = -0,18, p < 10-3). monkey N: r = -0.12, p = 0.07 monkey S: r = -0.18, p < 10-3).与这些发现一致，我们观察到大CP 之间存在适度但显着的相关性，这是根据图6b、e 和神经元阈值（图6c、f；几何平均回归；猴子N：r = -0.12，p = 0.07 猴子S：r = -0.18，p < 10-3）。与 这些 发现 一致 ， 我们 到 大 大 之间 存在 适度 但 显着 的 相关性 这 是 根据 图 6b 、 e 和 元 阈值 （图 图 6c 、 f ； 回归 ； 猴子 n ： r = -0.12 ， p = 0.07 猴子S：r = -0.18，p < 10-3）。 Consistent with these findings, we observed a modest but significant correlation between large CVs as shown in Fig. 6b,e and neuron thresholds (Figure 6c,f; geometric mean regression; monkey N: r = -0.12, p = 0.07). Обезьяна S: г = -0,18, р < 10-3). Monkey S: r = -0.18, p < 10-3). Therefore, cues from the most informative units tended to show greater covariance with subjective segmentation judgments in monkeys, which is important regardless of any textural cues added to the perceptual bias.
Given that we had previously established a relationship between sensitivity to grid texture signals and preferred neuronal orientation, we wondered if there was a similar relationship between CP and preferred orientation (Fig. 6g). This association was only slightly significant in monkey S (ANOVA; monkey N: 1.03, p=0.46; monkey S: F=1.73, p=0.04). We observed no difference in CP for lattice angles between lattices in any animal (Fig. 6h; ANOVA; monkey N: F = 1.8, p = 0.11; monkey S: F = 0.32, p = 0. 9).
Finally, previous work has shown that CP changes throughout the trial. Some studies have reported a sharp increase followed by a relatively smooth selection effect,30 while others have reported a steady increase in the selection signal over the course of the trial31. For each monkey, we calculated the CP of each unit in trials with zero textural contrast (respectively, according to pattern orientation) in 100 ms cells stepping every 20 ms from pre-stimulus start to post mean pre-stimulus offset. The average CP dynamics for two monkeys is shown in Fig. 6i. In both cases, the CP remained at a random level or very close to it until almost 500 ms after the onset of the stimulus, after which the CP increased sharply.
In addition to changing sensitivity, CP has been shown to be also affected by certain qualities of cell tuning characteristics. For example, Uka and DeAngelis34 found that CP in the binocular mismatch recognition task depends on the symmetry of the device’s binocular mismatch tuning curve. In this case, a related question is whether pattern direction selective (PDS) cells are more sensitive than component direction selective (CDS) cells. PDS cells encode the general orientation of patterns containing multiple local orientations, while CDS cells respond to movement of directional pattern components (Fig. 7a).
a Schematic representation of mode component tuning stimulus and hypothetical grating (left) and grating orientation tuning curves (right) (see Materials and Methods). In short, if a cell integrates across the grid components to signal pattern movement, one would expect the same tuning curves for individual grid and grid stimuli (last column, solid curve). Conversely, if the cell does not integrate the directions of the components into the motion of the signal pattern, one would expect a bipartite tuning curve with a peak in each direction of grating motion that translates one component into the cell’s preferred direction (last column, dashed curve). . b (left) curves for adjusting the orientation of the sinusoidal array for the cells shown in Figures 1 and 2. 3 and 4 (top row – cells in Figs. 3a,b and 4a,b (top); bottom panel – cells of Fig. 3c, d and 4a, b (bottom)). (Middle) Pattern and component predictions calculated from lattice tuning profiles. (Right) Adjusting the grid of these cells. The cells of the top (bottom) panel are classified as template (component) cells. Note that there is no one-to-one correspondence between the classifications of the pattern components and the preferences for coherent/transparent cell movement (see the texture lattice responses for these cells in Fig. 4a). c Coefficient of partial correlation of the z-score mode (ordinate) plotted against the partial correlation coefficient of the z-score component (abscissa) for all cells recorded in N (left) and S (right) monkeys. Thick lines indicate the significance criteria used to classify cells. d Plot of high selection probability (ordinate) versus mode index (Zp – Zc) (abscissa). Data in the left (right) panels refer to monkey N(S). Black circles indicate data in approximate units. In both animals, there was a significant correlation between high selection probability and pattern index, suggesting a greater perceptual correlation for cells with signal pattern orientation in stimuli with multiple component orientation.
Therefore, in a separate test set, we measured responses to sinusoidal grids and grids to classify the neurons in our samples as PDS or CDS (see Methods). Lattice tuning curves, template component predictions built from this tuning data, and lattice tuning curves for the cells shown in Figures 1 and 3. Figures 3 and 4 and Supplementary Figure 3 are shown in Figure 7b. The pattern distribution and component selectivity, as well as the preferred cell orientation in each category, are shown for each monkey in Fig. 7c and supplementary fig. 4 respectively.
To assess the dependence of the CP on the correction of the pattern components, we first calculated the pattern index 35 (PI), the larger (smaller) values ​​of which indicate a larger PDS (CDS) similar behavior. Given the above demonstration that: (i) neuronal sensitivity varies with the difference between the preferred cell orientation and the direction of stimulus movement, and (ii) there is a significant correlation between neuronal sensitivity and selection probability in our sample, we found a relationship between PI and the total CP was studied for the “best” direction of movement of each cell (see above). We found that CP was significantly correlated with PI (Fig. 7d; geometric mean regression; grand CP monkey N: r = 0.23, p < 0.01; bi-stable CP monkey N r = 0.21, p = 0.013; grand CP monkey S: r = 0.30, p < 10−4; bi-stable CP monkey S: r = 0.29, p < 10−3), indicating that cells classified as PDS exhibited greater choice-related activity than CDS and unclassified cells. We found that CP was significantly correlated with PI (Fig. 7d; geometric mean regression; grand CP monkey N: r = 0.23, p < 0.01; bi-stable CP monkey N r = 0.21, p = 0.013; grand CP monkey S: r = 0.30, p < 10−4; bi-stable CP monkey S: r = 0.29, p < 10−3), indicating that cells classified as PDS exhibited greater choice-related activity than CDS and unclassified cells. Мы обнаружили, что CP значительно коррелирует с PI (рис. 7d; регрессия среднего геометрического; большая обезьяна CP N: r = 0,23, p <0,01; бистабильная обезьяна CP N r = 0,21, p = 0,013; большая обезьяна CP S: r = 0,30, p < 10-4; бистабильный CP обезьяны S: r = 0,29, p < 10-3), что указывает на то, что клетки, классифицированные как PDS, проявляли большую активность, связанную с выбором, чем CDS и неклассифицированные клетки. We found that CP was significantly correlated with PI (Figure 7d; geometric mean regression; great monkey CP N: r = 0.23, p < 0.01; bistable monkey CP N r = 0.21, p = 0.013; great monkey CP S: r = 0.30, p < 10-4; monkey S bistable CP: r = 0.29, p < 10-3), indicating that cells classified as PDS showed more activity, associated with choice than CDS and unclassified cells.我们发现CP 与PI 显着相关（图7d；几何平均回归；大CP 猴N：r = 0.23，p < 0.01；双稳态CP 猴N r = 0.21，p = 0.013；大CP 猴S： r = 0.30，p < 10-4；双稳态CP 猴S：r = 0.29，p < 10-3），表明分类为PDS 的细胞比CDS 和未分类的细胞表现出更大的选择相关活性。 CP 与PI 显着相关（图7d；几何平均回归；大CP猴N：r = 0.23，p < 0.01；双稳态CP 猴N r = 0.21，r，p = 0.21，p3； 0.0 Мы обнаружили, что CP был значительно связан с PI (рис. 7d; регрессия среднего геометрического; большая обезьяна CP N: r = 0,23, p <0,01; бистабильная обезьяна CP N r = 0,21, p = 0,013; большая обезьяна CP S: r = 0,013) 0,30, p < 10-4; We found that CP was significantly associated with PI (Figure 7d; geometric mean regression; great monkey CP N: r = 0.23, p < 0.01; bistable monkey CP N r = 0.21, p = 0.013; great monkey CP S: r = 0.013) 0.30, p < 10-4; бистабильный CP обезьяны S: r = 0,29, p < 10-3), что указывает на то, что клетки, классифицированные как PDS, проявляли большую селекционную активность, чем клетки, классифицированные как CDS и неклассифицированные. monkey S bistable CP: r = 0.29, p < 10-3), indicating that cells classified as PDS exhibited greater selection activity than cells classified as CDS and unclassified. Because both PI and neuron sensitivity correlated with CP, we performed multiple regression analyzes (with PI and neuron sensitivity as independent variables and large CP as dependent variable) to rule out a correlation between the two measures leading to the possibility of an effect. . Both partial correlation coefficients were significant (monkey N: threshold vs. CP: r = −0.13, p = 0.04, PI vs. CP: r = 0.23, p < 0.01; monkey S: threshold vs. CP: r = −0.16, p = 0.03, PI vs CP: 0.29, p < 10−3), suggesting that CP increases with sensitivity and in an independent fashion increases with PI. Both partial correlation coefficients were significant (monkey N: threshold vs. CP: r = −0.13, p = 0.04, PI vs. CP: r = 0.23, p < 0.01; monkey S: threshold vs. CP: r = −0.16, p = 0.03, PI vs CP: 0.29, p < 10−3), suggesting that CP increases with sensitivity and in an independent fashion increases with PI. Оба частных коэффициента корреляции были значимыми (обезьяна N: порог против CP: r = -0,13, p = 0,04, PI против CP: r = 0,23, p <0,01; обезьяна S: порог против CP: r = -0,16, p = 0,03, PI vs CP: 0,29, p < 10-3), предполагая, что CP увеличивается с чувствительностью и независимым образом увеличивается с PI. Both partial correlation coefficients were significant (monkey N: threshold vs. CP: r=-0.13, p=0.04, PI vs. CP: r=0.23, p<0.01; monkey S: threshold vs. CP: r = -0.16, p = 0.03, PI vs CP: 0.29, p < 10-3), suggesting that CP increases with sensitivity and increases independently with PI.两个偏相关系数均显着（猴子N：阈值与CP：r = -0.13，p = 0.04，PI 与CP：r = 0.23，p < 0.01；猴子S：阈值与CP：r = -0.16， p = 0.03，PI vs CP：0.29，p < 10-3），表明CP 随灵敏度增加而增加，并且以独立方式随PI 增加。两个偏相关系数均显着（猴子N：阈值与CP：r = -0.13，p = 0.04，PI = 0.03，PI vs CP：0.29，p < 10-3），表明CP Оба частных коэффициента корреляции были значимыми (обезьяна N: порог против CP: r = -0,13, p = 0,04, PI против CP: r = 0,23, p <0,01; обезьяна S: порог против CP: r = -0,16, p = 0,03 , PI против CP: 0,29, p < 10-3), что указывает на то, что CP увеличивалась с чувствительностью и увеличивалась с PI независимым образом. Both partial correlation coefficients were significant (monkey N: threshold vs. CP: r=-0.13, p=0.04, PI vs. CP: r=0.23, p<0.01; monkey S: threshold vs. CP: r = -0.16, p = 0.03, PI vs CP: 0.29, p < 10-3), indicating that CP increased with sensitivity and increased with PI in an independent manner.
We recorded a single activity in the MT area, and the monkeys reported their perception of patterns that could appear as coherent or transparent movements. The sensitivity of neurons to segmentation cues added to biased perceptions varies widely and is determined, at least in part, by the relationship between the unit’s preferred orientation and the direction of stimulus movement. In the entire population, neuronal sensitivity was significantly lower than psychophysical sensitivity, although the most sensitive units matched or exceeded behavioral sensitivity to segmentation signals. In addition, there is a significant covariance between firing frequency and perception, suggesting that MT signaling plays a role in segmentation. Cells with preferred orientation optimized their sensitivity to differences in lattice segmentation signals and tended to signal global movement in stimuli with multiple local orientations, demonstrating the highest perceptual correlation. Here we consider some potential problems before comparing these results with previous work.
A major problem with research using bistable stimuli in animal models is that behavioral responses may not be based on the dimension of interest. For example, our monkeys could report their perception of texture orientation independently of their perception of lattice coherence. Two aspects of the data suggest that this is not the case. First, in accordance with previous reports, changing the relative orientation angle of the separating array components systematically changed the likelihood of coherent perception. Secondly, on average, the effect is the same for patterns that contain or do not contain texture signals. Taken together, these observations suggest that monkey responses consistently reflect their perception of connectedness/transparency.
Another potential problem is that we have not optimized the grating motion parameters for the specific situation. In many previous works comparing neuronal and psychophysical sensitivity, stimuli were selected individually for each registered unit [31, 32, 34, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45]. Here we have used the same two directions of movement of the lattice pattern regardless of the adjustment of the orientation of each cell. This design allowed us to study how sensitivity changed with overlap between lattice movement and preferred orientation, however, it did not provide an a priori basis for determining whether cells prefer coherent or transparent lattices. Therefore, we rely on empirical criteria, using the response of each cell to the textured mesh, to assign preference and zero labels to each category of mesh movement. Although unlikely, this could systematically skew the results of our sensitivity analysis and CP signal detection, potentially overestimating any measure. However, several aspects of the analysis and data discussed below suggest that this is not the case.
First, assigning preferred (null) names to stimuli that elicited more (less) activity did not affect the distinguishability of these response distributions. Instead, it only ensures that the neurometric and psychometric functions have the same sign, so they can be compared directly. Second, the responses used to calculate CP (the “wrong” trials for textured gratings and all trials for gratings without texture contrast) were not included in the regression analysis that determined whether each cell “prefers” connected or transparent sports. This ensures that selection effects are not biased towards preferred/invalid designations, resulting in a significant selection probability.
The studies of Newsom and his colleagues [36, 39, 46, 47] were the first to determine the role of MT in the approximate estimation of the direction of motion. Subsequent reports have collected data on MT participation in depth34,44,48,49,50,51 and speed32,52, fine orientation33 and perception of 3D structure from movement31,53,54 (3D sustainable forests). to rule). We extend these results in two important ways. First, we provide evidence that MT responses contribute to perceptual segmentation of visuomotor signals. Second, we observed a relationship between the MT mode orientation selectivity and this selection signal.
Conceptually, the present results are most similar to the work on 3-D SFM, as both are complex bistable perception involving movement and depth ordering. Dodd et al.31 found a large selection probability (0.67) in the monkey task reporting the rotational orientation of a bistable 3D SFM cylinder. We found a much smaller selection effect for bistable grid stimuli (about 0.55 for both monkeys). Since the assessment of the CP depends on the selection coefficient, it is difficult to interpret the CP obtained under different conditions in different tasks. However, the magnitude of the selection effect that we observed was the same for zero and low texture contrast gratings, and also when we combined low/no texture contrast stimuli to increase power. Therefore, this difference in CP is unlikely to be due to differences in selection rates between datasets.
The modest changes in MT firing rate that accompany perception in the latter case seem puzzling compared to the intense and qualitatively different perceptual states induced by 3-D SFM stimulation and bistable grid structures. One possibility is that we underestimated the selection effect by calculating the firing rate over the entire duration of the stimulus. In contrast to the case of 31 3-D SFM, where differences in MT activity developed around 250 ms in trials and then steadily increased throughout the trials, our analysis of the temporal dynamics of selection signals (see 500 ms after stimulus onset in both monkeys. In addition , after a sharp rise during this period, we observed fluctuations in CP during the remainder of the trial. Hupe and Rubin55 report that human perception of bistable rectangular arrays often changes during long trials. Although our stimuli were presented for only 1.5 seconds, our monkeys’ perception could also vary from coherence to transparency during the trial (their responses reflected their final perception at cue selection.) Therefore, a reaction time version of our task, or plan in which the monkeys can continuously report their perception, is expected to be have a larger selection effect.The last possibility is that the MT signal is read differently in the two tasks. Although it has long been thought that CPU signals result from sensory decoding and correlated noise,56 Gu and colleagues57 found that in computational models, different pooling strategies, rather than levels of correlated variability, can better explain CPU in dorsal media-superior temporal neurons. sheet change orientation recognition task (MSTd). The smaller selection effect we observed in MT probably reflects the extensive aggregation of many low-informative neurons to create perceptions of coherence or transparency. In any case where local motion cues were to be grouped into one or two objects (bistable gratings) or separate surfaces of common objects (3-D SFM), independent evidence that MT responses were significantly associated with perceptual judgments, there were strong MT responses. it is proposed to play a role in the segmentation of complex images into multi-object scenes using visual motion information.
As mentioned above, we were the first to report an association between MT pattern cellular activity and perception. As formulated in the original two-stage model by Movshon and colleagues, the mode unit is the output stage of the MT. However, recent work has shown that mode and component cells represent different ends of a continuum and that parametric differences in the structure of the receptive field are responsible for the tuning spectrum of the mode components. Therefore, we found a significant correlation between CP and PI, similar to the relationship between binocular mismatch adjustment symmetry and CP in the depth recognition task or orientation setting configuration in the fine orientation discrimination task. Relations between documents and CP 33 . Wang and Movshon62 analyzed a large number of cells with MT orientation selectivity and found that, on average, the mode index was associated with many tuning properties, suggesting that mode selectivity exists in many other types of signals that can be read from the MT population. . Therefore, for future studies of the relationship between MT activity and subjective perception, it will be important to determine whether the pattern index correlates similarly with other task and stimulus selection signals, or whether this relationship is specific to the case of perceptual segmentation.
Similarly, Nienborg and Cumming 42 found that although near and far cells selective for binocular mismatch in V2 were equally sensitive in the depth discrimination task, only the near-preferring cell population exhibited significant CP. However, retraining monkeys to preferentially weight distant differences resulted in significant CPs in more favored cages. Other studies have also reported that training history depends on perceptual correlations34,40,63 or a causal relationship between MT activity and differential discrimination48. The relationship we observed between CP and regimen direction selectivity likely reflects the specific strategy that the monkeys used to solve our problem, and not the specific role of mode selection signals in visual-motor perception. In future work, it will be important to determine if the learning history has a significant impact on determining which MT signals are weighted preferentially and flexibly to make segmentation judgments.
Stoner and colleagues14,23 were the first to report that changing the brightness of overlapping grid regions predictably affected the coherence and transparency of human observer reports and directional adjustments in macaque MT neurons. The authors found that when the brightness of the overlapping regions physically matched transparency, observers reported more transparent perception, while MT neurons signaled movement of raster components. Conversely, when the overlapping brightness and transparent overlap are physically incompatible, the observer perceives coherent movement, and MT neurons signal the global movement of the pattern. Thus, these studies show that physical changes in visual stimuli that reliably influence segmentation reports also induce predictable changes in MT arousal. Recent work in this area has explored which MT signals track the perceptual appearance of complex stimuli18,24,64. For example, a subset of MT neurons have been shown to exhibit bimodal tuning to a random point motion map (RDK) with two directions that are less spaced apart than a unidirectional RDK. Bandwidth of cellular tuning 19, 25 . Observers always see the first pattern as transparent movement, even though most MT neurons exhibit unimodal adaptations in response to these stimuli, and a simple average of all MT cells gives a unimodal population response. Thus, a subset of cells exhibiting bimodal tuning may form the neural substrate for this perception. Interestingly, in marmosets, this population matched PDS cells when tested using conventional grid and grid stimuli.
Our results go one step further than the above, which are critical to establishing the role of MT in perceptual segmentation. In fact, segmentation is a subjective phenomenon. Many polystable visual displays illustrate the ability of the visual system to organize and interpret persistent stimuli in more than one way. Simultaneously collecting neural responses and perceptual reports in our study allowed us to explore the covariance between MT firing rate and perceptual interpretations of constant stimulation. Having demonstrated this relationship, we acknowledge that the direction of causality has not been established, that is, further experiments are needed to determine whether the signal of perceptual segmentation observed by us is, as some argue [65, 66, 67], automatic. The process again represents descending signals coming back to the sensory cortex from higher areas 68, 69, 70 (Fig. 8). Reports of a larger proportion of pattern-selective cells in MSTd71, one of the main cortical targets of MT, suggest that extending these experiments to include simultaneous recordings of MT and MSTd would be a good first step towards further understanding the neural mechanisms of perception. segmentation.
A two-stage model of component and mode orientation selectivity and the potential effect of top-down feedback on choice-related activity in machine translation. Here, mode direction selectivity (PDS – “P”) in the MT step is created by (i) a large sample of direction selective input data consistent with specific mode velocities, and (ii) strong tuning suppression. The directionally selective (CDS) component of the MT (“C”) stage has a narrow sampling range in the input direction and does not have much tuning suppression. Untuned inhibition gives control over both populations. The colored arrows indicate the preferred device orientation. For clarity, only a subset of the V1-MT connections and one component mode and orientation selection box are shown. In the context of interpreting our feed-forward (FF) results, the broader input setting and strong tuning inhibition (highlighted in red) in PDS cells induced large differences in activity in response to multiple movement patterns. In our segmentation problem, this group drives decision chains and distorts perception. In contrast, in the case of feedback (FB), perceptual decisions are generated in upstream circuits by sensory data and cognitive biases, and the greater effect of downstream FB on PDS cells (thick lines) generates selection signals. b Schematic representation of alternative models of CDS and PDS devices. Here the PDS signals in the MT are generated not only by the direct input of V1, but also by the indirect input of the V1-V2-MT path. The indirect paths of the model are adjusted to give selectivity to texture boundaries (grid overlapping areas). The MT layer CDS module performs a weighted sum of direct and indirect inputs and sends the output to the PDS module. PDS is regulated by competitive inhibition. Again, only those connections are shown that are necessary to draw the basic architecture of the model. Here, a different FF mechanism than that proposed in a could lead to greater variability in the cellular lattice response to PDS, again leading to biases in decision patterns. Alternatively, greater CP in PDS cells may still be the result of bias in the strength or efficiency of FB attachment to PDS cells. The evidence supports the two- and three-stage MT PDS models and the CP FF and FB interpretations.
Two adult macaques (macaca mulatta), one male and one female (7 and 5 years old respectively), weighing from 4.5 to 9.0 kg, were used as objects of study. Prior to all sterile surgery experiments, animals were implanted with a custom-made recording chamber for vertical electrodes approaching the MT area, a stainless steel headrest stand (Crist Instruments, Hagerstown, MD), and eye position with a measured scleral search coil. (Cooner Wire, San Diego, California). All protocols comply with United States Department of Agriculture (USDA) regulations and the National Institutes of Health (NIH) Guidelines for the Humane Care and Use of Laboratory Animals and have been approved by the University of Chicago Institutional Animal Care and Use Committee (IAUKC).
All visual stimuli were presented in a round aperture against a black or gray background. During recording, the position and diameter of this hole were adjusted in accordance with the classical receptive field of neurons at the electrode tip. We used two broad categories of visual stimuli: psychometric stimuli and tuning stimuli.
The psychometric stimulus is a grating pattern (20 cd/m2, 50% contrast, 50% duty cycle, 5 degrees/sec) created by superimposing two rectangular gratings drifting in a direction perpendicular to their direction (Fig. 1b). It has been previously shown that human observers perceive these grid patterns as bistable stimuli, sometimes as a single pattern moving in the same direction (coherent movement) and sometimes as two separate surfaces moving in different directions (transparent movement). components of the lattice pattern, oriented symmetrically – the angle between the lattices is from 95° to 130° (drawn from the set: 95°, 100°, 105°, 115°, 120°, 125°, 130° °, throughout the session Isolation angle neurons at 115° were not preserved, but we include psychophysical data here) – approximately 90° or 270° (pattern orientation). In each session, only one corner of the interlattice lattice was used; during each session, the orientation of the pattern for each trial was randomly selected from two possibilities.
In order to disambiguate the perception of the grid and provide an empirical basis for the reward for action, we introduce random point textures into the light bar step 72 of each grid component. This is achieved by increasing or decreasing (by a fixed amount) the brightness of a randomly selected subset of pixels (Fig. 1c). The direction of texture movement gives a strong signal that shifts the observer’s perception towards coherent or transparent movement (Fig. 1c). Under coherent conditions, all textures, regardless of which component of the texture lattice covers, are translated in the direction of the pattern (Fig. 1c, coherent). In the transparent state, the texture moves perpendicular to the direction of the grating it covers (Fig. 1c, transparent) (Supplementary Movie 1). To control the complexity of the task, in most sessions the Michelson contrast (Lmax-Lmin/Lmax+Lmin) for this texture mark varied from a set of (-80, -40, -20, -10, -5, 0, 5). , 10, 20, 40, 80). Contrast is defined as the relative brightness of a raster (so a contrast value of 80% would result in a texture of 36 or 6 cd/m2). For 6 sessions in monkey N and 5 sessions in monkey S, we used narrower textural contrast ranges (-30, -20, -15, -10, -5, 0, 5, 10, 15, 20, 30), where psychophysical characteristics follow the same pattern as full-range contrast, but without saturation.
Tuning stimuli are sinusoidal grids (contrast 50%, 1 cycle/degree, 5 degrees/sec) moving in one of 16 equally spaced directions, or sinusoidal grids moving in these directions (consisting of two opposite 135° angles superimposed sinusoidal gratings on top of each other). in the same direction of the pattern.