Orientation discrimination requires coactivation of on- and off-dominated visual channels

Abstract

Orientation sensitivity depends on the cortical convergence of on- and off-center subcortical neurons. Off-center inputs are faster and stronger than their on-center counterparts: How does this asymmetry affect orientation discrimination? We tackled this question psychophysically with grating stimuli that either increased or decreased luminance. The gratings were of low contrast in order to avoid the complicating influences of nonlinearities such as response saturation, masking, and aftereffects. Gratings were presented in either of two locations, and subjects indicated the perceived location. Stimuli were randomly timed, and response correctness and reaction time were recorded. We found the following: (a) Contrast sensitivity was insignificant for a range of contrasts around zero. (b) Outside this range, contrast sensitivity for contrast decrements exceeded that for increments by an average of 15%. (c) Reaction times for contrast decrements were up to 45 ms less than for increments. (d) These findings are reproduced by a signal-detection model which incorporates recent physiological findings: Neurons in primary visual cortex are hyperpolarized at rest; these neurons respond more to darks than to lights; and off-dominated cortical neurons have shorter latencies than their on-dominated neighbors. (e) We tested orientation discrimination by splitting a grating into two components, one containing the light bars and the other the dark, and presenting the two components asynchronously. Discrimination was optimal when light bars preceded dark bars, consistent with coactivation of on- and off-center cortical inputs. We conclude that the ability to discriminate between orientations is intimately connected with the properties of subcortical channels.

Introduction

Attempts to link perception with neural function have a long history. Landmark studies in the late 1960s used classical psychophysical approaches such as adaptation and masking to study the spatial tuning and orientation selectivity of human vision (Blakemore & Campbell, 1969; Campbell & Kulikowski, 1966), revealing results that closely matched emerging findings about neurons in primary visual cortex (Hubel & Wiesel, 1968). In a similar vein, we were motivated to look for perceptual consequences of recent neurophysiological studies showing an asymmetry between responses to light increments and decrements of equal magnitude (Komban et al., 2014; C. I. Yeh, Xing, & Shapley, 2009). As Phillips and Wilson (1984) have pointed out, experiments using suprathreshold stimuli require careful interpretation. We have therefore taken a different approach, using low-contrast stimuli to avoid nonlinearities such as response saturation, masking, and afterimages. 

The asymmetry between responses to light increments and decrements has been shown psychophysically (Bowen, Pokorny, & Smith, 1989; Dannemiller & Stephens, 2001; Komban et al., 2014; Komban, Alonso, & Zaidi, 2011; Krauskopf, 1980; Lu & Sperling, 2012), in visual evoked potentials (Zemon, Gordon, & Welch, 1988), in multiunit recordings (Kremkow et al., 2014; Xing, Yeh, & Shapley, 2010), and in single cortical neurons (Komban et al., 2014; Kremkow, Jin, Wang, & Alonso, 2016; Lee, Huang, & Fitzpatrick, 2016; Liu & Yao, 2014; Samonds, Potetz, & Lee, 2012; Veit, Bhattacharyya, Kretz, & Rainer, 2014; Wang et al., 2015; C. I. Yeh et al., 2009). Each of these studies has found that responses to darks are stronger, faster, subserved by greater numbers of neurons, or more precisely retinotopically mapped, than responses to lights. The asymmetry is present in species from tree shrews to humans, suggesting that it may play a useful visual role. The measured response difference between lights and darks is, however, often small compared with the magnitude of the responses themselves. We wondered whether this small difference is due in part to the use of suprathreshold stimuli, leading to nonlinearities such as response saturation. We therefore measured the light/dark asymmetry in human subjects presented with low-contrast stimuli. 

The presence of this asymmetry in cortical responses suggests that it may influence an important visual capability: orientation discrimination. This ability is thought to rely on orientation-selective neurons (Wilson & Wilkinson, 2004). The cortical origin of orientation selectivity is still controversial (Vidyasagar & Eysel, 2015), but the strongest line of evidence indicates that it depends on the convergence of on- and off-center subcortical inputs onto single cortical neurons (Jin, Wang, Swadlow, & Alonso, 2011b; Reid & Alonso, 1995; Tanaka, 1983). Given that on- and off-dominated neurons convey responses to lights and darks, respectively, it is to be expected that the asymmetry between channels will affect orientation discrimination. We tested this idea by breaking gratings into light and dark bars, which were presented asynchronously. 

We therefore had two aims. The first was to measure response asymmetries between lights and darks in the absence of nonlinearities such as response saturation, masking, and afterimages. The second aim was to see what part this asymmetry plays in orientation discrimination. We have previously presented a summary of parts of this work (Freeman, Luo-Li, & Alais, 2015). 

Methods

Subjects

Sixteen human subjects (11 women, five men) took part in these experiments; they were aged 18 to 49. Subjects had normal vision in that visual acuity was at least 6/6 in each eye and stereo threshold was 1 min or better. One subject (GL-L) is an author of this article; all other subjects were unaware of the aims of the study and were paid for their time. All procedures conformed with the Declaration of Helsinki. 

Equipment

Two sets of equipment were used, distinguished mainly by the type of visual stimulator. Eleven subjects were tested with a cathode-ray-tube (CRT) monitor. Concerns have been raised about the fitness of such a monitor for testing asymmetries in responses to lights and darks (Gawne & Woods, 2003); we therefore used a liquid-crystal display (LCD) monitor for the remaining subjects. The two equipment sets are described in turn. 

CRT monitor

Stimuli were presented on a CRT monitor (Philips 105S, Philips, Amsterdam, the Netherlands) driven by an ATI Radeon HD 5770 video card. The card was controlled, and responses collected, with the Psychophysics Toolbox software (Brainard, 1997; Pelli, 1997) extended by a low-level kernel driver (Kleiner, Brainard, & Pelli, 2007). The monitor had a spatial resolution of 77 pixels/° and a frame rate of 75 Hz. Stimuli were monochromatic (x = 0.340, y = 0.329) with a mean luminance of 42 cd/m². Luminance was modulated with 10 bits/gun, yielding a contrast resolution of 1/(0.5 × 2¹⁰) = 0.0020. Larger contrasts were linearized using a lookup table. Subjects used a chin rest which stabilized eye-to-monitor distance at 1.14 m, and they signaled their responses by using two buttons on an RTbox (Li, Liang, Kleiner, & Lu, 2010). Experiments were conducted in a darkened room so that the only light visible to the subjects was from the monitor. 

LCD monitor

The second set of equipment differed from the first in the following respects. Stimuli were presented on a VIEWPixx monitor (VPixx Technologies, Saint-Bruno, Canada). Monitor resolution was 73 pixels/° at 120 Hz, mean luminance was 48 cd/m², and chromaticity was (x = 0.294, y = 0.312). Eye-to-monitor distance was 0.92 m, and subjects signaled responses with a ResponsePixx button box (VPixx). 

Comparison between setups

Figure 1 shows consecutive frames from each of our monitors. The LCD monitor offers at least two advantages over the CRT: It is faster (120 Hz compared with 75 Hz), and the decay time of a video frame is the same as the onset time. We therefore quantitatively compared the psychophysical data obtained from the two monitors. An analysis of variance on the data in Figure 5C showed no effect from the monitor (factors: subject, odd powers of contrast up to 5, monitor)—α = 0.05, r² = 0.95, F(1, 141) = 0.0071, p = 0.79—nor did such an analysis on the data in Figure 8C (factors: subject, contrast and its square, contrast polarity, monitor)—α = 0.05, r² = 0.92, F(1, 141) = 0.35, p = 0.55. Accordingly, the data are combined without distinguishing between monitors. The fitness of our methodology for measuring light/dark asymmetries is taken up in the Discussion. 

Figure 1

 

Video-frame timing on the stimulus monitors. Two stimulus monitors were used, one a cathode-ray tube and the other a liquid-crystal display. The graphs show the voltage recorded by a photosensor placed at the middle of the screen in the absence of any stimulus. Two video frames are shown for the cathode-ray-tube monitor and three for the liquid-crystal display. The liquid-crystal display's backlight successively lit eight horizontal strips down the screen, producing the eight discontinuities visible within each frame.

Experiment 1: Detection of lights and darks

We performed two experiments. The first compared the detection of contrast increments with detection of decrements. 

Stimuli

Figure 2A shows the spatial form of the stimulus, a raised Gabor: 0.5 × contrast × (1 + cosine) × Gaussian. More precisely, the stimulus is  

  ,

where x and y are the distances across the visual field (in degrees) in the horizontal and vertical directions, respectively. Contrast is defined as  

  ,

where l_center is luminance at the center of the stimulus and l_background is background luminance. The spatial frequency of the cosine function was f = 3 c/° and the standard deviation of the Gaussian function was σ_stim = 0.3° in both the horizontal and vertical directions. On each trial, the stimulus was presented with equal probability in two locations separated by 0.8°. Contrast on each trial was sampled from a Gaussian probability density with zero mean and a standard deviation of 0.05. Thus, as shown in Figure 2A, each trial contained a contrast increment or decrement: The subject's task was to indicate on which side the stimulus was delivered. The area containing the stimulus was 2.5° wide and high, bordered by a black line 0.25° wide.  

Figure 2

 

Stimuli in Experiment 1. (A) On each trial, a Gabor patch was presented on either the left or right side of the viewing area, with either positive or negative contrast. The subjects' task was to indicate the side on which the patch appeared. (B) The time course during a single trial. The test stimulus was presented between 1 and 2 s from the start. Incorrect responses were signaled with auditory feedback. (C) If the subject did not respond within 1 s of the test onset, an auditory prompt was presented and a further second allowed for the response. Reaction times in these prompted trials were discarded.

Procedure

The time course of a single experimental trial is shown in Figure 2B. It started with an interval whose duration was sampled from a uniform probability density spanning the range 1 to 2 s. The stimulus had sudden onset and faded linearly over 0.2 s. Subjects responded in the 1 s following the start of the stimulus or, if they failed to respond (usually because the stimulus was too faint), a medium-pitch auditory prompt was delivered and they were given another second in which to respond (Figure 2C). In both unprompted and prompted cases a low-pitch signal was sounded if their choice was incorrect. The next trial then started. Each run of trials lasted 60 s, and subjects rested between runs if they wished. 

Experiment 2: Orientation discrimination

The second experiment tested the ability to discriminate between gratings of differing orientation. This experiment differed from the previous one as follows. As shown in Figure 3A, the Gaussian envelope of the Gabor stimuli had a standard deviation of 0.45° and the Gabors were split into light bars (0.5 × contrast × (1 + sine) × Gaussian) and dark bars (−0.5 × contrast × (1 − sine) × Gaussian). The two polarities were presented asynchronously, for one video frame each, as shown in Figure 3B. On each trial, light and dark bars were aligned with each other but tilted from vertical by 2° (Figure 3C). The subjects' task was to indicate whether the tilt was clockwise or counterclockwise. 

Figure 3

 

Stimuli in Experiment 2. (A) Gabors were broken into a sum of components lighter and darker than the background. (B) Light and dark bars were presented asynchronously. The duration of a video frame was fixed at 13 ms because only the cathode-ray-tube monitor was used for this experiment. (C) On a specific trial, both light and dark bars were tilted 2° from vertical. The subjects' task was to indicate whether the tilt was clockwise or counterclockwise from vertical.

Model

We modeled the results of Experiment 1 as follows. Simple receptive fields in primary visual cortex typically have multiple subfields of two types: On-subfields respond to positive contrast and off-subfields to negative contrast (Hubel & Wiesel, 1962). The two polarities of subfield typically differ in strength, so that a neuron can be either on- or off-dominant (Komban et al., 2014). We assume four neuronal populations, with on- and off-dominant neurons at each of the two stimulus locations. Let p be the membrane potential relative to the action-potential threshold. As illustrated in Figure 4B, membrane potential in the absence of a stimulus is assumed to be  

 

where the resting potential p_r is negative, indicating that the neuron is hyperpolarized relative to threshold. The noise component n represents a Gaussian probability density with zero mean and standard deviation σ. Given that our stimuli are weak, we assume that membrane potential depends linearly on contrast c:  

 

where Display Formula is contrast sensitivity. On- and off-dominant neurons have positive and negative sensitivities Display Formula respectively. As shown in Figure 4C, the action-potential rate Display Formula , is proportional to the part of membrane potential that exceeds the ,threshold:  

 

where  

 

and k is the (constant) conversion factor from potential to impulse rate. Reaction time t is obtained by integrating the action-potential rate to a criterion number of action potentials m:  

 

where t_min is the minimum reaction time (and therefore includes fixed sensory and motor delays).  

Figure 4

 

Model for results of Experiment 1. (A) The model assumes that four populations of cortical neurons are involved: on- and off-dominated neurons on each side of the viewing area. (B) Probability density of membrane potential, relative to the action-potential threshold, for each population. Unstimulated neurons (green) have a mean potential which is less than the action-potential threshold. A stimulus with positive contrast shifts the on-dominated neurons (red) rightward and the off-dominated neurons (blue) leftward from the resting position. (C) Impulse rate is obtained by thresholding the membrane potential.

Figure 5

 

Psychometric functions, Experiment 1. (A) The proportion of correct responses as a function of stimulus contrast for a single subject. (B) The psychometric function in (A) is made more monotonic by plotting the probability that the subject chooses the lighter side during stimulation: This has the effect of inverting the points at negative contrast. (C) Psychometric functions for all subjects. (D) The circles show the psychometric function obtained by pooling data across subjects, and the line shows the best fit of the signal-detection model described in the text.

The subject chooses the side of the viewing area containing the population that first reaches the criterion, and reaction time is set equal to the time at which the population reaches this criterion. When all four populations have zero impulse rate (typically because of a very low contrast), the subject chooses either side with equal probability, and reaction time is assumed to be greater than 1 s. The model was solved by Monte Carlo simulation. For each contrast a total of 5,000 trials were run, each using a noise sample independent of other trials. The model was fitted to the data by minimizing least squared error. Denoting the gap between resting potential and action-potential threshold as one unit of potential difference, the resulting parameters were  

 

 

 

 

 

 

 

 

 

 

Results

Experiment 1: Detection of lights and darks

It is clear that psychophysical and neurophysiological responses to light decrements are not simple mirror images of the responses to light increments (Bowen et al., 1989; Dannemiller & Stephens, 2001; Komban et al., 2014; Krauskopf, 1980; Kremkow et al., 2014, 2016; Lee et al., 2016; Liu & Yao, 2014; Lu & Sperling, 2012; Samonds et al., 2012; Veit et al., 2014; Wang et al., 2015; Xing et al., 2010; C. I. Yeh et al., 2009; Zemon et al., 1988). We aimed to describe this asymmetry by minimizing response nonlinearities and thereby determining how the asymmetry affects orientation discrimination. The stimuli in our first experiment, shown in Figure 2A, consisted of Gabors with a raised carrier: A raised cosine function of distance was multiplied by a Gaussian profile so that each presentation either incremented or decremented light, but not both. The Gabor appeared on either the left or right of the midline, and the subjects' task was to indicate on which side it appeared. 

Contrast increments had fast onset and faded offset, as shown in Figure 2B, in order to favor on-dominant neurons. Contrast decrements had the same timing, but with the opposite polarity, to preferentially stimulate off-dominated neurons. We recorded the proportion of correct responses as well as reaction time. Subjects frequently failed to see test stimuli with the lowest contrasts and were therefore prompted for a response, as shown in Figure 2C. These prompted responses contributed to the measurement of proportion correct but were not included with the reaction-time data. 

Psychometric functions

The results for one subject are shown in Figure 5A. Contrast on each trial was sampled from a Gaussian probability density centered on 0, and the psychometric function was computed by compiling responses into bins that contained, as nearly as possible, equal numbers of samples. As expected, the proportion of correct responses increases with contrast magnitude. It is convenient to make this U-shaped function more monotonic, as shown in Figure 5B. The vertical axis here plots the probability that the subject chooses the lighter side of the bordered area. Figure 5C and D show data from all subjects and responses pooled across subjects, respectively. The line in Figure 5D is the prediction of a signal-detection model that will be described later. 

The psychometric functions in Figure 5 have two notable features. First, there is a plateau centered on zero contrast. This feature is explored in more detail in Figure 6A, which shows the slope of the psychometric function (obtained by differencing neighboring bins in the data of Figure 5C). Figure 6B shows the mean over subjects, with 95% confidence intervals. The error bar at zero contrast includes zero contrast sensitivity, indicating that the sensitivity is not significantly different from zero. Thus, there is a range of contrasts in which the subject has little or no information about the stimulus. Neurons in primary visual cortex are hyperpolarized at rest (Tan, Chen, Scholl, Seidemann, & Priebe, 2014) and therefore require a contrast magnitude greater than some fixed minimum to generate action potentials. It seems likely that the plateau in our psychometric functions indicates the range of contrasts in which membrane potential has not reached the action-potential threshold. 

Figure 6

 

Contrast sensitivity, Experiment 1. (A) Contrast sensitivity was calculated from the psychometric functions in Figure 5C by taking the difference between neighboring contrast bins. Each line represents a single subject. (B) Mean contrast sensitivity across subjects, with 95% confidence intervals. The error bar at a contrast of zero includes zero contrast sensitivity, indicating that subjects are insensitive to a range of contrasts about zero.

The other notable feature in both Figures 5 and 6 is the asymmetry between the left and right sides of the curves. This is shown more clearly in Figure 7, where the proportion of correct responses is shown as a function of contrast magnitude for a single subject (Figure 7A), for all subjects (Figure 7B), and pooling data across all subjects (Figure 7C). Red and blue curves derive from contrast increments and decrements, respectively. In general, the blue curves lie above the red ones, indicating that our subjects were more sensitive to contrast decrements than to increments of the same magnitude. An analysis of variance on the data in Figure 7B showed that the difference in proportion correct was significant (factors: subject, powers of contrast magnitude from 1 to 3, contrast polarity), α = 0.05, r² = 0.94, F(1, 140) = 8.7, p = 0.0038. Figure 7D was obtained by finding the contrast at which the proportion of correct responses was 0.75 and dividing this threshold for contrast increments by that for decrements: The ratio averages 1.15 over all subjects. 

Figure 7

 

Psychometric functions versus contrast magnitude, Experiment 1. (A–B) The psychometric function for (A) a single subject and (B) all subjects. The function for contrast increments is shown in red, and for decrements in blue. (C) Circles show pooled data, and lines give the predictions of the signal-detection model. Subjects are more sensitive to decrements than to increments: The thresholds for the increment and decrement curves, measured as the contrast at which the (interpolated) proportion correct is 0.75, are 0.043 and 0.037, respectively. (D) Contrast threshold for contrast increments divided by that for decrements. Each circle represents one subject.

Chronometric functions

Along with the proportion of correct responses, we measured the reaction times of our subjects. Figure 8A shows reaction times for one subject, where each circle gives the time taken to respond to a single presentation of the test stimulus. For contrasts close to zero, subjects were often not aware that the test had been delivered and therefore had to be prompted to respond. These points give the reaction time to an auditory prompt 1 s after the visual stimulus, and reaction times greater than 1 s have therefore been discarded. For the remaining points, reaction time decreases with contrast magnitude, as expected. 

We were particularly interested in comparing reaction times to lights and darks. Figure 8B through D make this comparison by plotting the data in various ways against contrast magnitude. They show plots in which the data are binned by contrast magnitude and each point provides mean reaction time. Figure 8B gives the results for a single subject, Figure 8C gives results for all subjects with one line for each subject and contrast polarity, and Figure 8D shows pooled data. The last plot shows that reaction times for contrast decrements are less than for increments; an analysis of variance on the data in Figure 8C showed that the difference is highly significant (factors: subject, contrast magnitude and its square, contrast polarity), α = 0.05, r² = 0.92, F(1, 141) = 46, p = 2.9 × 10⁻¹⁰. This result matches previous psychophysics (Komban et al., 2011, 2014) and the finding that off-dominated neurons in cat primary visual cortex respond faster than do on-dominated neurons (Komban et al., 2014). The lines in Figure 8D, which give the predictions of the model to be described later, are based on single-neuron properties such as these. 

The reaction-time difference between contrast increments and decrements is shown for all subjects in Figure 9A and for the mean across subjects in Figure 9B. There are two points of particular interest in this last plot. First, the biggest difference between lights and darks is obtained with a contrast magnitude of 0.05, and the difference declines with higher contrasts. Second, the maximum reaction-time difference we obtain is about 45 ms. This is a substantially larger difference that that found in single-neuron studies (Komban et al., 2014), possibly because we used smaller contrasts. 

Figure 8

 

Chronometric functions, Experiment 1. (A) Reaction time was measured along with the proportion of correct responses. This graph shows reaction time for a single subject, with responses to contrast increments in red and to decrements in blue. (B–C) Mean reaction time for (B) a single subject and (C) all subjects. (D) The pooled data, given by the circles, show that responses to contrast decrements are faster than those to increments. The lines show the best-fitting model.

Figure 9

 

Reaction-time differences between contrast increments and decrements, Experiment 1. (A–B) Reaction time for contrast decrements was subtracted from that for increments; the difference is shown for (A) all subjects and (B) the mean across subjects. The 95% confidence intervals in (B) show that reaction times for decrements are significantly less than for increments across most of the contrast range.

Model

Our psychophysical demonstrations of response asymmetries between lights and darks have clear neural correlates. We find, for example, that the contrast sensitivity for darks is greater than that for lights. This matches the finding that the areas of cat (Jin et al., 2008) and monkey (Xing et al., 2010; C. I. Yeh et al., 2009) visual cortex dominated by off-responses are larger than on-dominated areas. We therefore wished to see whether a signal-detection model based on known physiology could reproduce our psychometric and chronometric functions. 

The model assumes two neuronal populations—on- and off-dominated cortical cells—on each side of the viewing area. Membrane potential in each of these four populations is assumed to have a Gaussian probability density. Unstimulated, the mean of this density lies below the action-potential threshold, as shown in Figure 4B. Stimulated, the mean changes by the product of contrast and contrast sensitivity, where the sensitivity is positive for on- and negative for off-dominated neurons. Figure 4B shows the resulting shifts in the density. Impulse rate is obtained by half-wave rectifying the membrane potential, as shown in Figure 4C. On any trial, a sample is taken from the thresholded density in each of the populations and integrated over time toward a criterion number of impulses. The subject chooses the side of the viewing area containing the population that first reaches the criterion, and reaction time is set equal to the time at which the population reaches this criterion. 

The model was fitted to the observations by optimizing five parameters: contrast sensitivity for on- and off-dominated neurons, standard deviation of the probability density, criterion level, and minimum reaction time. The results are shown by the lines in Figures 5D, 7C, and 8D. The fits are good in that the fraction of variance accounted for by the model is 99.4%, 98.7%, and 98.2%, respectively. It seems, therefore, that a relatively simple model can account for both the proportion of correct responses and the reaction time, provided that contrasts are small. 

Experiment 2: Orientation discrimination

Human subjects are highly sensitive to contour orientation, in that they can reliably judge the alignment of two contours to within about 0.5° (Westheimer, Shimamura, & McKee, 1976). This capability is thought to depend on orientation-selective neurons in primary visual cortex (Wilson & Wilkinson, 2004). It is known that on- and off-center geniculate relay cells can converge onto the same cortical neuron (Alonso, Usrey, & Reid, 1996; Tanaka, 1983), and there is increasing evidence that this convergence provides the foundation for orientation selectivity (Jin et al., 2011b; Reid & Alonso, 1995). Given the timing differences between on- and off-pathways that we have already described, we wondered whether orientation discrimination might also depend on the relative timings of contrast increments and decrements. 

To test this idea, we decomposed a grating into light and dark bars as shown in Figure 3A and presented these two components asynchronously at the same location (Figure 3B). On each trial, both sets of bars were tilted 2° clockwise or counterclockwise from vertical, as shown in Figure 3C, and the subjects' task was to indicate the direction of tilt. The pooled results for 10 subjects are shown in Figure 10A. Seven psychometric functions are shown, with one for each onset asynchrony between light and dark bars. Orientation discrimination is best when the light and dark bars are presented at about the same time. 

Figure 10

 

Experiment 2. (A) The subjects' task in this experiment was to discriminate between gratings with differing tilts. On each trial, the light and dark components of a grating were presented asynchronously, as shown by the legend at right. Psychometric functions calculated for the pooled data are shown. (B) The effect of onset asynchrony was determined by interpolating the psychometric functions to find the contrast threshold at a proportion correct of 0.75. Contrast sensitivity is the reciprocal of this threshold. Error bars give 95% confidence intervals determined by bootstrap resampling of the pooled data. Performance is best when light bars precede dark bars by 13 ms. (C) Chronometric functions for the pooled data. (D) Circles show contrast sensitivity at a reaction time of 0.55 s, and error bars give 95% confidence intervals. Performance is again best for asynchronies close to zero.

We quantified this effect by finding the (interpolated) contrast at a proportion correct of 0.75, then taking the reciprocal of this contrast to obtain contrast sensitivity. The result is shown in Figure 10B, along with 95% confidence intervals. Two aspects of this graph are notable. First, discrimination is substantially poorer when the asynchrony between light and dark bars is large in magnitude than when it is close to zero. Second, there is a clear asymmetry between the left and right sides of the graph: Discrimination is better when light bars lead dark bars. In particular, the two points with an asynchrony magnitude of 13 ms do not fall within each other's error bars, indicating that this difference is significant. Given that on-pathways have a longer response latency than do off-pathways, this result suggests that orientation discrimination is optimal when on- and off-signals reach cortex at about the same time. 

Reaction time provides further evidence for this finding via the chronometric functions in Figure 10C. The tuning curve in Figure 10D was calculated by finding the contrast at which reaction time equaled 0.55 s. This measure of discriminability again falls with increasing magnitude of asynchrony. The tuning curve is asymmetric, but in this case the asymmetry narrowly fails to reach significance. 

Discussion

Our results demonstrate a robust difference between responses to lights and darks, and the influence of this difference on a fundamental visual capability: orientation discrimination. The key new findings are as follows. 

•  
  There is a range of contrasts centered on zero for which the stimulus is invisible
•  
  Previous studies have shown that responses to light decrements are stronger than responses to increments. Other studies have found that responses to darks are faster than to lights. We have shown in a single experiment that responses to light decrements have an advantage over increments in both sensitivity and speed
•  
  These data are reproduced by a signal-detection model in which off-dominated neurons have higher contrast sensitivity than their on-dominated neighbors but both classes of neuron are hyperpolarized at rest.
•  
  Optimal orientation discrimination requires that on- and off-inputs to visual cortex be activated at about the same time.

We review our methodology and then discuss each of these results in turn. 

Methodology

Several concerns have been raised about the use of video stimulators in vision experiments (Zele & Vingrys, 2005), particularly in studies of response asymmetries to lights and darks (Gawne & Woods, 2003). These concerns include the time taken to paint, and the asymmetry of the luminance profile during, a single video frame (Zele & Vingrys, 2005) and the interval between the offset of one frame and the onset of the next (Gawne & Woods, 2003). We have partially addressed these concerns by including a relatively fast (120 Hz) stimulator. More conclusively, we find that the peak latency difference between dark and light responses averages 45 ms, which covers more than three frames in the case of the CRT monitor and more than five frames for the LCD. This rules out the possibility that our findings are artifacts of intraframe effects or of interactions between consecutive frames. 

Another methodological issue is the definition of contrast. The use of Michelson contrast, Display Formula , is problematic when comparing luminance increments with decrements: The denominator for luminance patterns lying above background is greater than for patterns in which luminance is less than the background. This could result in light/dark asymmetries due to the contrast definition rather than to visual processing. We have been careful, therefore, to use Weber contrast, Display Formula . In this case, stimulus luminance is proportional to contrast, so that light/dark asymmetries cannot be ascribed to the contrast definition.  

Psychometric function

The psychometric function in Figure 5D is unconventional in that it displays responses to both positive and negative contrasts. The upper right quadrant, for which contrast is positive, is more familiar in that it is sigmoidal in shape (Nachmias & Sansbury, 1974; Naka & Rushton, 1966). What accounts for the low contrast sensitivity at near-zero contrast? Foley and Legge (1981) provided evidence for a noisy signal followed by a threshold but did not suggest a neural locus for the threshold. Dean (1981) showed that neurons in cat primary visual cortex do not respond to contrasts below a fixed magnitude, and it has recently been found (Tan et al., 2014) that unstimulated neurons in the primary visual cortex of awake monkeys are hyperpolarized relative to the action-potential threshold. It seems likely, therefore, that the contrast-response function has a low gradient near the origin because of resting hyperpolarization in primary cortex. 

Light/dark asymmetries

Asymmetries between responses to lights and darks, which thread through our results, are now well established in the literature. Zemon et al. (1988) recorded visually evoked cortical potentials in human subjects presented with positive- and negative-contrast checkerboards, and found that contrast gain was greater in the latter case. Bowen et al. (1989) used a stimulus with a sawtooth time course to show that psychophysical detection was better when the rapid change in luminance was downward than when it was upward. More recently, three studies have found that off-dominant neurons are more prevalent than on-dominant in primary visual cortex of cat (Jin et al., 2008) and monkey (Xing et al., 2010; C. I. Yeh et al., 2009). Off-center inputs to cortex have a lower response latency than do their on-center counterparts (Jin et al., 2011a), a difference that is reflected in the findings that off-dominant cortical neurons respond faster that do on-dominant (Komban et al., 2014) and that psychophysical reaction times are less for darks than for lights (Komban et al., 2011, 2014). A number of other studies (Dannemiller & Stephens, 2001; Krauskopf, 1980; Kremkow et al., 2014, 2016; Lee et al., 2016; Liu & Yao, 2014; Lu & Sperling, 2012; Samonds et al., 2012; Veit et al., 2014; Wang et al., 2015) also show that responses to negative contrasts have an advantage over those to positive contrasts. Two studies (Chichilnisky & Kalmar, 2002; Smith, Whitney, & Fitzpatrick, 2015) present contrary findings, but they derive from stimulation in the peripheral visual field, suggesting that distance from the center of vision may be an important factor. 

Where do light/dark asymmetries originate in the visual system? Komban et al. (2014) have suggested the photoreceptor-bipolar synapse as the source. The synapses between photoreceptors and on- and off-center bipolar calls are metabotropic and ionotropic, respectively, and the former type of synapse is known to be slower (Snellman, Kaur, Shen, & Nawy, 2008). Direct evidence for or against the proposal that light/dark asymmetries originate at this synapse is lacking. Another possible source is light adaptation. Increased (decreased) luminance decreases (increases) visual sensitivity (Shapley & Enroth-Cugell, 1984), and much of the sensitivity change occurs within 100 ms of a luminance change (Hayhoe, Benimoff, & Hood, 1987; T. Yeh, Lee, & Kremers, 1996). The relationship between sensitivity and luminance is a power law with an exponent between −0.5 and −1 (Shapley & Enroth-Cugell, 1984). Figure 7C indicates a sensitivity increase of 15% when luminance drops by 8% (at a contrast magnitude of around 0.04). This implies a power law with an exponent more negative than −1, meaning that if adaptation produces asymmetries between light and dark responses it cannot be solely responsible. 

Orientation discrimination

We have also shown that orientation discrimination depends on the timing of light and dark components of the stimuli being discriminated. This makes sense given the known physiology of the geniculocortical synapse. Orientation-selective neurons in primary visual cortex receive convergent input from spatially separated populations of on- and off-center geniculate neurons (Jin et al., 2011b; Reid & Alonso, 1995; Tanaka, 1983). Given that on-pathways to cortex are slower than off-pathways (Jin et al., 2011a), light components of a stimulus apparently need to be delivered before dark components in order for the two response components to arrive at the cortex at about the same time. Only then are orientation-selective neurons optimally stimulated. This finding raises an obvious puzzle: Natural stimulation, in which light and dark components of a stimulus are delivered simultaneously, will result in suboptimal stimulation. Is the asymmetry of light and dark responses a bug or a feature in the visual system? 

One answer to this question may be that natural viewing typically involves durations substantially longer than the asynchronies used in our experiment. Another answer may lie in motion sensitivity. The on- and off-center geniculate-neuron populations that converge on a cortical neuron are spatially separated and have differing timing, providing the essential substrate for motion-direction selectivity: A stimulus moving in one direction will almost certainly evoke a cortical response that differs from that evoked by movement in the opposite direction. Indeed, a model assuming that cortical inputs are separated in both space and time reproduces a number of published observations on motion-direction selectivity (Hesam Shariati & Freeman, 2012). Neurophysiological confirmation of this idea is not yet available. In the meantime, the possibility remains that light/dark asymmetries play a major role in the foundation of motion-direction selectivity. 

Stimulus appearance

We have shown that responses to lights are weaker than responses to darks. How might this influence the appearance of a pattern containing both lights and darks? Peli (1997) presented subjects with two Gabors of differing phase and adjusted the physical contrast of the patterns until they had the same apparent contrast. The match was best when the Michelson contrasts of the patterns were equal. Michelson contrast, using our terminology, is  

  .

 

The denominator here is less for darks than for lights, suggesting an advantage for responses to darks relative to lights. We therefore replotted the data in Figures 7C and 8D in terms of Michelson contrast instead of Weber contrast. The results, illustrated in Figure 11, show that the advantage of darks over lights is still significant—left side (factors: subject, powers of contrast magnitude from 1 to 3, contrast polarity), α = 0.05, r² = 0.94, F(1, 140) = 9.2, p = 0.0029; right side (factors: subject, contrast magnitude and its square, contrast polarity), α = 0.05, r² = 0.93, F(1, 141) = 269, p = 1.8 × 10⁻³⁴—but effect sizes are reduced by this procedure (left side: threshold ratio = 1.10; right side: maximum reaction-time difference = 36 ms). This means that our subjects respond similarly to differing patterns if contrast is defined so as to reinforce the contribution of lights. Conversely, the result in Figure 11 suggests that Peli's observations on stimulus appearance can be at least partly explained by the response asymmetry between lights and darks. 

Figure 11

 

Michelson contrast. The data in Figures 7C and 8D are replotted in terms of Michelson (rather than Weber) contrast. The advantage of responses to darks over responses to lights is reduced. This result helps to explain the apparent contrast of patterns containing both light and dark (Peli, 1997).

Low-contrast stimulation

We have used very small contrast magnitudes in our measurements, mostly less than 0.1. The use of low contrast has at least two advantages. First, it minimizes nonlinearities such as response saturation, masking, and afterimages. Using low contrast therefore leaves us with three essential nonlinearities: the “dead zone” in sensitivity around zero contrast, and the higher gain and lower latencies seen with contrast decrements. The finding that these nonlinearities cannot be removed by using weak stimuli suggests that they play a critical role in visual function. A second advantage of using low contrasts is that they correspond with the most prevalent contrasts in the natural environment (Brady & Field, 2000). Frequency histograms of contrast in natural scenes show a steady decline away from zero contrast. Our visual environments contain large but informative areas of low contrast such as surfaces, textures, and shadowed regions. Our low-contrast results therefore throw light on the information available from these ubiquitous features of the visual scene. 

Acknowledgments

The authors declare no competing financial interests. Funding was provided by the Australian Research Council (DP130102336). We thank Paul Martin for his valuable comments on an earlier version of this article. 

Commercial relationships: none. 

Corresponding author: Alan W. Freeman. 

Email: [email protected]. 

Address: School of Medical Sciences, The University of Sydney, Sydney, Australia.