Introduction

As signals propagate along the visual hierarchy, individual neurons represent increasingly complex aspects of the visual environment. In principle, these later representations are better suited to support complex perceptual tasks than the simpler representations that precede them, and may thus have a more direct relationship with perceptual experience. This relationship is best assessed through the simultaneous recording of neural responses and behavioral reports of perceptual decisions (Newsome et al., 1989; Parker and Newsome, 1998). In many areas of the brain, the activity of individual neurons approaches or even exceeds the perceptual sensitivity of the subject and, moreover, predicts behavioral choices even on trials where the visual stimulus is ambiguous (Britten et al., 1992; Prince et al., 2000; Uka and DeAngelis, 2006; Nienborg and Cumming, 2006, 2014). Such results have been taken as evidence of an area’s participation in the formation of a perceptual decision, regardless of the exact origin of choice-correlated activity (Cumming and Nienborg, 2016). Whether sensory noise is fed forward to causally influence downstream areas that integrate evidence into a decision (Shadlen et al., 1996), or whether choice-related activity is fed back from decision circuits to sensory areas to support hierarchical probabilistic inference (Nienborg and Cumming, 2009; Haefner et al., 2016), most explanations of choice-related activity suggest it is a reflection of the decision-making process.

A particularly successful application of this approach has been in the investigation of the neural representation of binocular disparity. Both the primary (V1) and secondary (V2) visual cortex are sensitive to visual disparities, but there is a clear shift from absolute to relative disparity sensitivity in V2 compared with V1, more closely aligning with perception (Thomas et al., 2002). Nienborg and Cumming (2006) further demonstrated that V2 neurons are more sensitive to disparity than V1 neurons, and predict behavioral choices in response to ambiguous stimuli, while V1 neurons do not. We wondered whether a similar V1-V2 distinction might exist in the representation of visual form. We previously found that V1 and V2 neurons can be distinguished based on their responses to naturalistic texture stimuli nearly as well as based on their sensitivity to relative disparity (Thomas et al., 2002; Freeman et al., 2013). Further, multiple observations provide indirect evidence for a relationship between the response of populations of V2 neurons and the perception of naturalistic textures (Freeman and Simoncelli, 2011; Freeman et al., 2013; Ziemba et al., 2016; Ziemba and Simoncelli, 2021). Despite being recorded under anesthesia, V2 responses, but not V1 responses, predicted human psychophysical performance on a naturalistic texture discrimination task (Freeman et al., 2013). Here, we directly test the strength of this link between V2 responses and the perception of naturalistic image structure by measuring neuronal and behavioral sensitivity simultaneously in the same observers.

We found that V2 neurons were substantially more sensitive to naturalistic image structure than V1 neurons, confirming our previous results in anesthetized animals (Freeman et al., 2013; Ziemba et al., 2018, 2019). However, average sensitivity in both V1 and V2 was far from behavior. Correspondingly, we found inconsistent evidence for a relationship between neuronal responses and behavioral choice in either V1 or V2. When compared with previous data collected from the same two monkeys performing an orientation discrimination task, we found that choice-correlated neural activity was unstable across tasks and observers, and untethered from both neuronal sensitivity and signatures of nonsensory modulation. Our results provide further evidence for the functional differentiation of V1 and V2 in the neural representation of naturalistic image structure, but suggest that further elaboration of these signals in downstream areas may more directly support perception of these visual features (Movshon and Simoncelli, 2014; Okazawa et al., 2015; Okazawa et al., 2017). Finally, we conclude more broadly that the presence of choice-correlated activity in the sensory cortex is unlikely to be a universal indicator of a neural population’s participation in the formation of a perceptual decision.

ResultsNeuronal sensitivity to naturalistic visual structure in awake monkey V1 and V2

To study the neural basis of perception of naturalistic image structure, we created a set of model-based synthetic texture stimuli that have been well studied in perceptual and neurophysiological research (Portilla and Simoncelli, 2000; Balas et al., 2009; Freeman and Simoncelli, 2011; Freeman et al., 2013; Okazawa et al., 2015; Ziemba et al., 2016; Okazawa et al., 2017; Ziemba et al., 2018, 2019; Herrera-Esposito et al., 2021a; Ziemba and Simoncelli, 2021; Lieber et al., 2023; Lee et al., 2024). The model measures the marginal and joint statistics across the outputs of a simulated population of V1 simple and complex cells, in response to a photographic image of a natural texture (Fig. 1A; Portilla and Simoncelli, 2000). The same statistics are then measured from an input image of Gaussian white noise and iteratively adjusted until they match the statistics from the original image. If only the second-order, or spectral, statistics from the original image are imposed, then the result is what we refer to as a “noise” image. If additional higher-order joint statistics across the output of V1-filters are matched to the original, then the resulting “naturalistic” texture image often more closely resembles the original image (Fig. 1A; Portilla and Simoncelli, 2000). Starting the synthesis from a different input white-noise image results in different samples of statistically matched texture. When the response over many samples is averaged, V1 neurons respond with similar firing rates to naturalistic and noise images, whereas V2 neurons generally respond with higher firing rates to naturalistic images (Freeman et al., 2013; Ziemba et al., 2018, 2019). These experimental observations were obtained from anesthetized macaque monkeys, and here we sought to test whether this distinction between V1 and V2 selectivity persists under conditions where subjects are awake and behaving.

Figure 1.

Tuning to “naturalness” in V1 and V2 of awake monkeys. A, Schematic of texture synthesis procedure. An original black and white photograph is decomposed into the responses of a population of V1-like filters tuned to different orientations and spatial frequencies. Both the linear responses, and their squared energy, are computed, representing simple and complex cells, respectively. The second stage computes local correlations across the V1 responses tuned to different orientations, spatial frequencies, and positions and spatially averages them. To produce a synthetic texture, the model analyzes an image of Gaussian white noise and iteratively adjusts it to more exactly match the statistics of the original image. If the image is matched for the full set of correlations, then we refer to it as a “naturalistic texture,” and if only matched for statistics capturing the spectral (second-order) content of the original image, we refer to it as “spectrally matched noise.” B, Mean responses of an example V2 neuron to multiple samples of naturalistic (dark blue) and noise (light blue) images from 5 different texture families (top). Modulation index computed from these responses for 5 texture families (middle). C, Distribution of modulation index values for a population of V1 (left) and V2 (right) neurons recorded from two awake, fixating macaque monkeys. Downward arrows indicate distribution mean, μ.

We installed a recording chamber with access to both V1 and V2 in two macaque monkeys trained to perform a texture discrimination task. During each recording session, we lowered an electrode into the visual cortex and isolated a single unit in either V1 or V2. We hand-mapped receptive fields to determine the receptive field center while the monkey fixated on a central spot. For the subsequent measurements, the animal maintained fixation while several naturalistic and noise textures were shown, centered on the receptive field and presented within an aperture 4 deg in diameter. We presented multiple samples of naturalistic and spectrally matched noise images from 5 different texture families, each presented for 100 ms and separated by 100 ms of gray screen (Fig. 1B). We calculated a modulation index by dividing the difference in response to naturalistic and noise stimuli by the sum. This index captures the strength of preference for naturalistic images for each texture family (Fig. 1B, bottom).

Examining the average modulation index over all 5 texture families for the entire population of recorded neurons, we found a similar pattern to that observed in anesthetized macaque cortex (Fig. 1C). Specifically, there was little overall preference in the V1 population for naturalistic or spectrally matched noise stimuli and the average modulation index was near zero. In contrast, V2 neurons were mostly driven to higher firing rates by naturalistic stimuli, leading to more positive modulation indices. This pattern was consistent across both monkeys (monkey 1 V1: mean = 0.00, n = 137; monkey 1 V2: mean = 0.08, n = 250; monkey 2 V1: mean = 0.02, n = 143; monkey 2 V2: mean = 0.09, n = 192).

Unlike the average modulation index across textures in the original report recorded in anesthetized animals (Freeman et al., 2013), V1 neurons were significantly shifted toward positive values (P = 0.012; t-test on signed modulation; mean = 0.01, n = 280; Fig. 1C, left). However, here we used only the 5 texture families that yielded the highest average modulation index in V2 from the set of 15 in the original study. Since the average modulation for each texture family was correlated across V1 and V2, the average modulation index in anesthetized V1 for these 5 textures was also significantly shifted toward positive values (P = 0.006; t-test on signed modulation; mean = 0.03, n = 102). There was no significant difference between average modulation index recorded previously from anesthetized V1 and here from awake V1 (P = 0.1; t-test on signed modulation). However, the modulation index in V2 was larger in the anesthetized monkey (mean = 0.22, n = 103) compared with awake V2 recordings made here (P < 0.001; t-test; mean = 0.09, n = 442; Fig. 1C, right). This may in part reflect regression to the mean, as the stimuli used here were chosen on the basis of the high V2 modulation index in the anesthetized experiments. Importantly, in both anesthetized and awake monkeys, V2 neurons had a significantly higher modulation index compared with those in V1 (P < 0.001; t-test).

We selected the texture family that evoked the most discriminable responses between naturalistic and noise stimuli for the subsequent behavioral task, regardless of whether the neuron preferred naturalistic or noise stimuli (for example, the second texture family in Fig. 1B). Previous work indicates that neuronal sensitivity along the discrimination axis, rather than overall firing rate, best predicts the strength of choice-correlated activity (Krug et al., 2016). Therefore, we chose the texture family (out of the 5) most likely to reveal a relationship between the responses of the particular neuron under study and the subject’s behavior in the discrimination task.

We focus on the question of how well a given sensory neuron can support discrimination on the basis of the strength of higher-order image statistics. This task most strongly differentiates V2 from V1 neurons (Freeman et al., 2013; Movshon and Simoncelli, 2014; Ziemba et al., 2019; Lieber et al., 2023). A related, but distinct, question is how well a neuron can support discrimination of different sets of higher-order statistics—a task that is perhaps more relevant to natural vision (Ziemba et al., 2016; Ziemba and Simoncelli, 2021; Herrera-Esposito et al., 2021b). However, we expect these different capacities might be related. Although many naturally occurring visual textures can be well-discriminated on the basis of spectral statistics alone (Ziemba et al., 2016; Herrera-Esposito et al., 2021a, 2021b), naturalistic textures with identical spectral statistics can still be easily discriminated on the basis of their higher-order statistics (Fig. 2A). To examine the relation between neuronal detection and discrimination capabilities, we presented a subset of V1 and V2 neurons with samples from 5 pairs of texture families that differed in their higher-order statistics but were exactly matched for their spectral statistics (Fig. 2A). We computed the discrimination sensitivity between each pair and found that average neuronal sensitivity was higher in V2 compared with V1 (Fig. 2B; P = 0.0013, t-test). Moreover, the magnitude of the average modulation index of these V2 neurons (Fig. 1C) predicted the strength of discrimination sensitivity between naturalistic textures (Fig. 2C; r = 0.55, P = 0.002, Spearman correlation; for V1: r = 0.09, P = 0.75). Together, these results suggest that higher-order image statistics may be a useful cue for challenging texture discrimination tasks, and that neuronal detection and discrimination abilities for these statistics are correlated within area V2, but not in area V1.

Figure 2.

Neuronal sensitivity to the higher-order statistics of natural textures predicts naturalistic texture discrimination A, Example samples from pairs of texture families that have been matched to each other for their spectral statistics but differ in their higher-order statistics. Images in each column share spectral statistics. B, Distributions of the average discrimination sensitivity across the 5 texture family pairs for V1 (left) and V2 (right). Downward arrows indicate distribution mean, μ. C, Average neuronal discrimination sensitivity plotted against modulation index between naturalistic and noise textures for V1 and V2 neurons. Naturalistic-preferring V2 neurons have higher discrimination sensitivity (r = 0.6, P = 0.008; V1: r = −0.5, P = 0.45, Spearman correlation). V2 neurons that prefer noise also tend to have higher discrimination sensitivity (r = −0.55, P = 0.07; V1: r = −0.02, P = 0.97).

Figure 3.

Naturalistic texture discrimination task. A, After the subject acquired fixation for 250 ms, two choice targets appeared (one noise and one naturalistic). After another 500 ms, a stimulus was presented in the neuron’s receptive field (blue circle). Subjects judged whether the statistics of this stimulus were closer to noise or naturalistic texture. After a 500 ms presentation, the stimulus disappeared and the subject communicated their decision with a saccade toward one of the two choice targets. Rewards were given for correct answers, defined relative to naturalness value of 0.5, with stimuli at this boundary rewarded randomly. B, Example stimuli along naturalness axis, for one texture family. C, Behavioral performance of monkey 1 over many sessions. Left, average psychometric performance across all sessions. The points represent measured behavior. The line represents a fit of the signal detection theory model of behavior. Right top, distribution of guess rate over all sessions. Middle, distribution of criterion bias over all sessions. Bottom, distribution of sensitivity over all sessions (defined as the reciprocal of the standard deviation of the fitted noise). D, Same as (C) but for monkey 2.

Naturalistic texture discrimination

In each recording session, after examining the tuning and determining the texture family yielding the highest sensitivity for the recorded neuron, we trained monkeys to perform a texture discrimination task where they judged the “naturalness” of a peripherally presented patch of texture (Fig. 3A). On each trial, after the animal attained stable fixation, we presented two choice targets to the left and right of fixation in the upper visual field. One choice target was a naturalistic texture matched to an original natural photograph for the joint statistics of the outputs of differently tuned V1-like filters (Portilla and Simoncelli, 2000). The other choice target was spectrally matched noise, lacking the higher-order statistics of the naturalistic texture. 500 ms after choice target onset, we presented a target stimulus centered on the receptive field of the simultaneously recorded single unit in V1 or V2. We generated the target stimulus by synthesizing intermediate textures whose statistics were linearly interpolated between the naturalistic and spectrally matched noise endpoints (Fig. 3B; Freeman et al., 2013; Vacher et al., 2020). When no higher-order statistics were included in the synthesis (naturalness = 0) the stimulus was spectrally matched noise, and when the higher-order statistics were fully imposed (naturalness = 1) the stimulus resembled naturalistic texture. Stimuli with different levels of naturalness were synthesized using the same image seed, and this seed was randomized across trials so any particular spatial feature provided no cue for completing the task. Only discrimination of the higher-order statistics themselves allowed for high performance. The monkey indicated whether the target was more naturalistic or noise-like by making an eye movement to one of the choice targets following stimulus offset after 500 ms. Target stimuli above 0.5 naturalness were rewarded for a saccade to the naturalistic choice target, and those below 0.5 naturalness were rewarded for a saccade to the noise target.

Both subjects performed the task well. When the target had a naturalness value of 0 or 1, performance was nearly perfect, and progressively declined as naturalness approached the experimenter-induced decision boundary at 0.5 (Fig. 3C,D). We fit the animals’ behavioral data for each session with a model in which choices arise from comparing a learned decision criterion to a noisy estimate of naturalness, with occasional “lapse” trials. Sensitivity, the slope of the resulting psychometric function at the fitted decision criterion, corresponds to the reciprocal of the fitted noise standard deviation. Both subjects had similar high sensitivity and little bias, despite having to learn the 0.5 naturalness boundary for each family (Fig. 3C,D).

Comparing neuronal and perceptual sensitivity for naturalistic visual structure

Tuning for naturalness was, on average, modest in both V1 and V2. We assigned single units from each area according to their preference for naturalistic or spectrally matched noise images (see methods), and examined the average responses to all stimuli for each group. The groups were of similar size in V1, but in V2 there were more than twice the number of naturalistic preferring neurons (Fig. 4A). These naturalistic preferring V2 neurons on average fired ∼15 impulses per second more for naturalistic stimuli compared with noise. This response differential was ∼8 impulses per second for noise-preferring V2 neurons, and was ∼6 in V1 neurons regardless of preference.

Figure 4.

Comparison of neural and behavioral sensitivity. A, Average naturalness tuning for populations of V1 (green) and V2 (blue) neurons preferring natural (dark) or noise (light) textures, plotted separately for monkey 1 (top) and monkey 2 (bottom). B, Neurometric and psychometric functions obtained for sessions yielding the closest match between neuronal and behavioral sensitivity for V1 (top) and V2 (bottom) . Black line represents the signal detection theory model fit to behavioral responses. Colored symbols represent ideal observer analysis applied to the responses of single neurons. Neuronal sensitivity was obtained by fitting the signal detection theory model to these neurometric data. C, Neurometric and psychometric functions recorded from sessions yielding ratios of neuronal to behavioral sensitivity that were more typical of the V1 (top) and V2 (bottom) populations. D, Distribution of the ratio between single neuron and behavioral sensitivity. Upright distributions represent naturalistic-preferring neurons, and upside down distributions represent noise-preferring neurons. Ordinate axes in each panel indicate proportion across the entire population, and downward arrows indicate the median ratio for naturalistic- (filled) and noise-preferring (empty) neurons. In both monkeys, V2 neurons (bottom) were significantly closer to the sensitivity of behavior than V1 neurons (top), and naturalistic-preferring neurons in V2 were significantly closer to the sensitivity of behavior than noise-preferring neurons. There was no difference between naturalistic- and noise-preferring neurons in V1.

Although both subjects learned to perform the task well and showed consistent sensitivity across sessions, the sensitivity of single neurons to naturalness varied widely. To estimate neuronal sensitivity, we performed an ideal observer analysis on the distribution of spike counts to different levels of naturalness. We applied a decision criterion at the median spike count elicited by stimuli at 0.5 naturalness, and took any response above this criterion as a neuronal “decision” to report natural. When plotted as a function of stimulus naturalness, the proportion of natural responses traces out a neurometric function (Fig. 4B,C). The slope of this function serves as a measure of the sensitivity of the neuron to changes in naturalness, and can be directly compared to sensitivity measured from the psychometric function. Examining the most sensitive neurons demonstrates that on occasion neuronal sensitivity approached that of the animal’s choices, especially in V2 (Fig. 4B). However, more typically, the sensitivity of the monkey far exceeded that of single neurons in both V1 and V2 (Fig. 4C).

Both subjects showed a similar pattern in the relationship between neuronal and behavioral sensitivity. Sensitivity in V2 was significantly closer to behavior than in V1 (Fig. 4D; P < 0.001, Wilcoxon rank sum test). On average, V1 neurons were 18 times less sensitive than the animal’s behavior while V2 neurons were about 10 times less sensitive than behavior. In V1, there was no difference between neurons that preferred naturalistic to those that preferred noise (median = 17.0 versus 20.4; P = 0.93, Wilcoxon rank sum test). However, in V2 in both monkeys, neurons that preferred naturalistic images had sensitivity significantly closer to behavior (median = 7.4 versus 17.1; P < 0.001, Wilcoxon rank sum test). Although we have previously hypothesized that V2 neurons may play a role in the perception of naturalistic texture (Freeman et al., 2013), this 7-fold difference between behavioral and neuronal sensitivity is substantially larger than in many studies that have compared behavior with a potential neural correlate (Britten et al., 1992; Nienborg and Cumming, 2006, 2014; Goris et al., 2017). However, unlike most of these studies, we had limited ability to optimize the texture discrimination task to the particular selectivity of the recorded neuron because of the complexity of the space of possible naturalistic textures. A previous study that recorded populations of V1 and V4 units while monkeys performed a fixed orientation discrimination task found average behavioral to neuronal sensitivity ratios greater than 20 (Jasper et al., 2019). In contrast, when a monkey performs an orientation task adjusted to the preference of a recorded V1 neuron, average behavioral to neuron sensitivity ratios are close to one (Nienborg and Cumming, 2014; Goris et al., 2017). This comparison suggests that if we performed a more comprehensive characterization of the naturalistic texture selectivity of V2 neurons and optimized the task stimuli accordingly, neuronal sensitivity might be much closer to behavior.

Inconsistent choice-correlated activity across subjects in V1 and V2

We wondered whether we could find evidence for the potential participation of our recorded population in the formation of perceptual decisions about naturalistic texture despite the discrepancy between neuronal and perceptual sensitivity. In particular, we wondered whether the increased sensitivity of V2 neurons would manifest in a greater tendency to predict perceptual decisions on a trial-by-trial basis. To examine this, we computed the “choice probability” for the responses of each neuron to the ambiguous, 0.5 naturalness condition (Fig. 5A; Britten et al., 1996). This quantity corresponds to the probability that a neuron fires more spikes preceding a behavioral decision associated with its preferred stimulus. A choice probability of 0.5 indicates that a neuron’s response contains no information about choice and a value of 1 indicates perfect prediction of choice.

Figure 5.

Choice probability for naturalistic texture discrimination. A, Mean responses from an example V2 neuron conditioned on whether the monkey chose “naturalistic” or “noise” on a given trial (top). Distribution of responses when naturalness = 0.5 conditioned on the behavioral choice (bottom). B, Distribution of choice probability for naturalness = 0.5 for V1 (top) and V2 (bottom) neurons recorded from monkey 1 (left) and Monkey 2 (right). Filled bars represent neurons with choice probability significantly different from 0.5 as determined by permutation test. Mean choice probability in V1 and V2 was significantly larger than 0.5 in monkey 1 (P < 0.05; t-test), but not monkey 2 (P > 0.05; t-test). C, Mean choice probability for the two stimulus conditions flanking naturalness = 0.5, plotted against choice probability for naturalness = 0.5 for V1 (top) and V2 (bottom) neurons recorded from monkey 1 (left) and monkey 2 (right).

The two monkeys differed in the pattern of their choice-correlated activity. In monkey 1, there were many neurons with strong choice-correlated activity (Fig. 5A) and the mean choice-probability in both V1 and V2 was significantly larger than 0.5 (Fig. 5B, left). However, we found no evidence that mean choice probability differed from 0.5 in monkey 2 in either V1 or V2 (Fig. 5B, right). While the inconsistency between the two monkeys is not unprecedented, we were surprised to find little difference in the strength of choice-correlated activity in V1 and V2, given the very weak neuronal sensitivity to texture naturalness in V1 neurons compared with V2. To examine the consistency of this choice-correlated activity, we also calculated the average choice probability for the two stimulus conditions adjacent to the ambiguous, 0.5 naturalness condition. For monkey 1, we found no significant correlation in choice probability across conditions in V1 (r = 0.12; P = 0.38), but we found a significant relationship in V2 (r = 0.28; P = 0.007; Fig. 5C). Despite this, choice probability was still significantly larger than 0.5 in V1 for neighboring conditions (mean = 0.055; P = 0.001; t-test). For monkey 2, we found no significant correlation in either V1 (r = 0.11; P = 0.34) or V2 (r = 0.09; P = 0.38). However, choice probability was actually significantly larger than 0.5 exclusively when computed from the adjacent conditions of V1 in monkey 2 (mean = 0.54, P < 0.001; t-test). In summary, choice-correlated activity was a stable and consistent feature of neural responses only in V2 of monkey 1.

The stimuli in our texture discrimination task differ from most previous experiments for which significant choice-correlated activity has been observed. While most studies use dynamic, noisy stimuli, we presented a single, static sample of texture for 500 ms during each trial (but see Kosai et al., 2014). Our previous results indicate that differences in statistically matched texture samples account for a substantial amount of neuronal variability in V1 and V2 (Ziemba et al., 2016). To examine whether texture sample variability might drive our choice probability observations, we presented different texture samples multiple times over the course of a session. We calculated a grand choice probability value by combining neural responses across multiple presentations of the same texture sample that led to different behavioral responses. The results of this analysis were similar to our observations without taking different samples into account for both monkey 1 (V1: Mean = 0.54, P = 0.003, t-test; V2: Mean = 0.53, P = 0.04), and monkey 2 (V1: Mean = 0.50, P = 0.81, t-test; V2: Mean = 0.52, P = 0.04).

We wondered whether more sensitive neurons might have a stronger association with choice. We found little evidence for any such consistent relationship in our data. In monkey 1, more sensitive V1 neurons tended to also exhibit lower choice probability (r = −0.26, p = 0.048, Spearman correlation; Fig. 6, top left). While this pattern is opposed to that most often observed, it mirrors results from this monkey performing an orientation discrimination task (Goris et al., 2017). Data from V1 of monkey 2 trended in the opposite direction (r = 0.22, p = 0.059, Spearman correlation; Fig. 6, top right). However, there was little evidence for these relationships when examining noise- or naturalistic-preferring neurons independently (all p > 0.09). Furthermore, there was no evidence for a relationship between neuronal sensitivity and choice probability in V2 of either monkey (monkey 1: r = −0.024, p = 0.8; monkey 2: r = −.009, p = 0.93).

Figure 6.

Relationship between neuronal sensitivity and choice-correlated activity. Choice probability for the naturalness = 0.5 condition against the signed neuronal sensitivity of V1 (top) and V2 (bottom) neurons recorded from monkey 1 (left) and monkey 2 (right). Negative sign indicates a higher firing rate associated with low naturalness stimuli. Red lines indicate the best fitting linear relationship for positively or negatively signed neurons.

Figure 7.

Trial-by-trial relationship between behavioral choice and neural activity. A, The choices of the monkey plotted against responses of an example V2 neuron for all conditions where naturalness ≠ 0.5 (left), and when naturalness = 0.5 (right). The diameter of each symbol indicates the number of responses of that magnitude observed. Lines represent the fit of a logistic regression analysis performed separately across (left, solid) and within (right, dashed) stimulus conditions. B, The slope of the logistic regression analysis performed across conditions plotted against the analysis performed within conditions for monkey 1 (left) and monkey 2 (right). Insets show example across (solid) and within (dashed) fits within quadrants 2–4. Green symbols represent V1 and blue symbols represent V2 neurons. Symbol outlined in black indicates example neuron shown in (A). C, Previously reported orientation discrimination task. D, Slope of logistic regression across vs within conditions (as in (B)) for the orientation discrimination task (same two monkeys).

Inconsistent choice-correlated activity across tasks in V1 and V2

Given the complex relationship between neuronal sensitivity and choice-correlated activity, we developed a more nuanced way of examining this relationship to facilitate comparisons across subjects and different tasks (Zaidel et al., 2017). Combining the logic of choice probability with the power of logistic regression instead of ROC analysis, we applied the same method to quantify the strength of both sensitivity and choice-correlated activity. For each neuron we used logistic regression to predict the choice of the animal based on the observed spike count on each trial. First, we found the slope coefficient when estimating the model across stimulus conditions (but excluding the ambiguous 0.5 naturalness condition; Fig. 7A