Time Course of Sensory Substitution for Gravity Sensing in Visual Vertical Orientation Perception following Complete Vestibular Loss

Visual Abstract


Introduction
The vestibular organs in the inner ear are fundamental for several brain functions, including locomotion, gaze stabilization, and spatial navigation. Loss of the vestibular organs in animal models (Thomson et al., 1991;Macpherson and Inglis, 1993) or humans (Crawford, 1964) is immediately followed by a complete inability to stand and severe disorientation. These symptoms improve over a few days and, in a chronic state, patients with bilateral vestibular loss (BVL) can typically perform everyday tasks such as walking or driving but retain equilibrium issues, gaze instability (Gillespie and Minor, 1999;Zingler et al., 2008;Ward et al., 2013;Strupp et al., 2016) as well as cognitive (Hanes and McCollum, 2006;Bigelow and Agrawal, 2015;Popp et al., 2017) and navigation (Schautzer et al., 2003;Brandt et al., 2005) deficits.
Recovery from BVL appears to be based on sensory substitution. Following BVL, gaze stabilization reflexes that counterrotate the eye to compensate for head rotation rely increasingly on visual signals (Huygen et al., 1989) and neck proprioception (Huygen et al., 1991;Sadeghi et al., 2012). Accordingly, neurophysiological studies have shown that vestibular nuclei neurons develop an increased sensitivity to proprioceptive stimulation (Yates and Miller, 2009;Sadeghi et al., 2012;McCall et al., 2013) and efference copies (Sadeghi et al., 2012) following vestibular lesions.
Furthermore, the visual scene and objects around us generally remain perceptually invariant relative to the allocentric world, regardless of our head/body orientation (Van Beuzekom and Van Gisbergen, 2000;Van Beuzekom et al., 2001;De Vrijer et al., 2008. Clinical studies show that deficits in the vestibular system or in combining gravitational and visual signals due to brain injury compromise this visual stability (Brandt et al., 1994;Funk et al., 2010Funk et al., , 2011Baier et al., 2012a,b;Guardia et al., 2012). Along with other symptoms, this may result in disturbing episodes in which the world appears upside-down or sideways. For this reason, earth-vertical visual orientation [often referred to as subjective visual vertical (SVV)] tasks, that probe the ability to discern the orientation of visual stimuli relative to gravitational vertical, are a common technique to investigate vestibular function (Tarnutzer et al., 2009a,b;Clemens et al., 2011;Alberts et al., 2015Alberts et al., , 2016Alberts et al., , 2018, notably in clinical settings (Böhmer and Rickenmann, 1995;Valko et al., 2011). In these tasks, a subject is rolled ear-down in the dark and asked to orient a dimly lit bar vertically in space (to align it with gravity), or to report whether a visual bar is oriented left or right relative to vertical. With only the bar visible, this is accomplished with relatively small errors when the subject is rolled up to ;45° (Mittelstaedt, 1983;Van Beuzekom and Van Gisbergen, 2000;Kaptein and Van Gisbergen, 2004;De Vrijer et al., 2008). Results from electrical stimulation of vestibular afferents (Lewis et al., 2013) further point to a dominant role of vestibular over proprioceptive systems in creating a gravity-centered visual representation. Although these tasks are mainly used to probe orientation perception in humans, some studies (Daddaoua et al., 2008;Lewis et al., 2008) succeeded in training macaque monkeys to report SVV.
Strikingly, several studies (Tabak et al., 1997;Lopez et al., 2007;Grabherr et al., 2011;Alberts et al., 2015Alberts et al., , 2018Toupet et al., 2017;Bürgin et al., 2018) have shown that average SVV is unaltered in patients suffering from BVL, although these settings may become more variable (Bürgin et al., 2018). This suggests either that the vestibular system contributes little to SVV or that extravestibular gravity sensing can thoroughly substitute for otolith function following vestibular loss. Indeed, a number of studies have shown that graviception naturally occurs in the trunk (Mittelstaedt, 1992(Mittelstaedt, , 1996Vaitl et al., 2002;Trousselard et al., 2003;Clemens et al., 2011;Alberts et al., 2016). Model-based analyses (Clemens et al., 2011;Alberts et al., 2016) have postulated that the brain fuses vestibular and body graviception according to their respective reliabilities, the former typically being more accurate and consequently more heavily weighted. It has been proposed that sensory substitution involves a re-weighting of the remaining sensory modalities (Alberts et al., 2018).
However, the normal weight of vestibular cues has never been measured directly, and the time course of sensory reweighting after vestibular loss remains unknown since SVV studies in human patients have always been performed in chronic stages of BVL.
Here, we used an animal model to measure earth-vertical visual orientation perception following vestibular loss. We first tested animals before, and acutely after vestibular loss to test the hypothesis that vestibular cues normally dominate graviception; under this hypothesis, task performance should drop substantially after vestibular loss. Second, we monitored sensory reweighting over time, testing the hypothesis of full recovery and, for the first time, measuring the time course of this recovery. We find a vestibular dominance in graviception, and a progressive restoration of perceptual performance after BVL, with a time constant of approximately three weeks of retraining.

Animals
Three male rhesus macaques (animals N, P, and Z; aged seven, three, and six years, respectively), were used in the study. Animals were pair-housed in a vivarium under normal day/night cycle illumination. Animals were chronically implanted with a circular delrin ring for head restraint (Gu et al., 2006) and scleral search coils for monitoring eye position. Because of eye coil failures in animal N, an infrared optical eye tracker (ISCAN) monitored binocular eye movements in some experiments. Animals learned to perform a visual orientation discrimination task through standard operant conditioning procedures. After reaching baseline levels, animals N and Z received bilateral labyrinthectomies by opening the vestibular labyrinth, removing the neuroepithelia, and filling the vestibule with streptomycin powder (2%; Newlands et al., 2001;Liu et al., 2013;Yu et al., 2014). We chose to use animals that were previously trained at performing a slant discrimination task (Elmore et al., 2019) so as to minimize the number of animals used in our laboratory. All surgical procedures were conducted under isoflurane anesthesia. Experimental procedures were in accordance with National Institutes of Health guidelines and approved by a local Animal Studies Committee.

Setup and stimuli
Animals sat in a primate chair 30 cm from an LCD monitor mounted at eye level. Visual stimuli, generated with OpenGL, appeared through a circular aperture subtending ;29°of visual angle, which was constructed from black, nonreflective cardboard. The same material enclosed the area surrounding the animal's head and the monitor, such that only the stimulus was visible to the animal. Stimuli were slanted planar surfaces depicting static random dot stereograms (RDSs) and rendered as threedimensional red-green anaglyphs that filled the viewable region of the monitor and had uniform density on the screen (Sanada et al., 2012;Elmore et al., 2019). We used RDS planes to preclude the use of monocular cues for judging slant so that the task required three-dimensional visual perception. Both the chair and the monitor were attached to a three-dimensional rotating platform (Acutronics Inc.), which pitched the animal and monitor at discrete angles about the interaural axis. Custom Spike2 scripts controlled the trial structure (CED).

Earth-vertical visual orientation discrimination task
We trained animals to perform a SVV task adapted to rotation in the pitch plane. The layout of the task is shown in Figure 1. Each trial began with the presentation of a fixation target at eye level, centered in the viewable region of the LCD screen. Fixation was enforced with a 2°monocular window and 1°vergence window. After fixation for 300 ms, a RDS plane appeared for 800-1000 ms (Fig. 1A), and the animal had to maintain fixation during this period (otherwise the trial was aborted). Following stimulus presentation, the fixation point and RDS plane disappeared, and two choice targets appeared 8.6°above and below the center of the screen. Monkeys were trained (see below) to discriminate the orientation of the RDS stimulus relative to an allocentric earth-vertical orientation, defined as the plane whose normal is perpendicular to the direction of gravity. To receive a juice reward, the animal made a saccade to the upper target if he perceived the top of the presented RDS plane to be behind the earth-vertical plane and to the lower target if he perceived the top of the presented RDS plane to be in front of the allocentric earth-vertical plane. Correct answers (rewarded with a Figure 1. Overview of the task. A, Task flow diagram. Animals must fixate a central point for 300 ms, after which a RDS with a visual slant of 640°(in intervals of 5°) is shown for 1000 ms. Animals must next report the allocentric orientation of the stereogram by performing an upward or downward saccade. B, C, Coordinate transformations involved in the task. The animals must use gravity sensing (symbolized by a green pendulum) to determine the orientation of the slanted visual stimulus (blue) relative to the allocentric earth-vertical plane (red). If animals perform the task perfectly, then the PSE occurs when the visual stimulus is aligned with the earth-vertical plane, i.e., at 0°slant in B and À30°slant in C.
drop of water or juice) depend on the magnitude of the monkey's pitch, such that an accurate percept of one's body orientation is required for accurate task performance.
For each block of trials on a given day, the monkey was positioned at three different body pitches. A 0°(upright) body pitch was always included, as well as symmetrical positive and negative pitches (e.g., 630°). The orientation of the RDS planes with respect to the monitor ranged from À40°to 40°in 5°increments, and all planar slant values are presented once in random order at one pitch angle before the monkey's pitch was changed. The need to use large (5°) increments was dictated by two factors.
(1) The range of visual slant stimuli presented should always be the same and independent of the monkey pitch angle. Given that the largest body tilt we used was 6 30°, the visual slant stimuli needed to span a 6 40°range ( Fig.  1B). (2) The number of trials should be tractable, so the smallest possible difference in visual stimulus slant was set to 5°. This coarse sampling of slant orientations did not allow accurate threshold measurements.
Each pitch angle was repeated between 5 and 33 times (mean: 20.2; SD: 6.8) within each session. Importantly, since the monitor, animal and surrounding enclosure rotated together, the stimulus set was identical in a head-centered reference frame for every pitch angle. In the absence of visual verticality cues, the animal had to rely on other senses to judge the direction of gravity (Rosenberg and Angelaki, 2014). Furthermore, since the animals were held at each static tilt position for a long duration (while all slant angles were tested), a contribution of the semi-circular canals was negligible.

Behavioral training
All animals were first trained in a slant discrimination task while sitting upright (animal pitch = 0°; Elmore et al., 2019). Following neural recordings, animals were trained with the allocentric vertical task, initially with pitch magnitudes of 630°and 0°. A reference plane (similar to the reference line of Daddaoua et al., 2008) aligned with the allocentric vertical plane was flashed for 300 ms before stimulus presentation to aid the animal's decision. The contrast of this reference plane was then gradually reduced. Once animals became proficient to perform the allocentric vertical task at 630°p itch angles without the reference plane, we introduced them to novel pitch angles (610°, 615°, 620°, and 625°). Monkeys were trained prelabyrinthectomy with a staircase procedure until they acquire !85% compensation and then tested with the method of constant stimuli to determine baseline behavioral performance.

Data analysis
We quantified behavioral performance separately for each pitch angle by fitting the proportion of "upward" choices as a function of stimulus slant in a head-centered reference frame with a cumulative Gaussian function (Psignifit toolbox for MATLAB; Wichmann and Hill, 2001a,b) to compute bias/ point of subjective equality (PSE) and behavioral threshold (SD of cumulative Gaussian function), along with 95% confidence intervals. Note that, because the slant angle was increased by increment of 5°, small psychometric threshold (e.g., ,5°) could not be estimated reliably. Based on each week's data, PSE is then plotted as a function of body pitch angle and that week's data points are fit with a linear regression. The regression slope reflects the percent compensation. If the animal performs the task perfectly, then percent compensation equals 100%.

Sensory fusion model
We implemented a similar model as in Clemens et al. (2011) and Alberts et al. (2016). The model assumes that the brain estimates head pitch optimally based on the following information: (1) a gravity signal originating from the otoliths, encoding a pitch angle P OTO and subject to a Gaussian noise with SD s OTO ; (2) a proprioceptive gravity signal, encoding a pitch angle P PROP and subject to a Gaussian noise with SD s PROP ; and (3) a Gaussian a priori that the head is upright, centered on a pitch angle of 0°, with SD s PRIOR .
Based on this information, the optimal estimate of pitch is the following: Note that the three weights (w OTO , w PROP , and w PRIOR ) always sum to 1. Therefore, although the prior does not contribute directly to P FINAL in Equation 1 (since w PRIOR is multiplied by 0), it does so indirectly since a large prior weight necessarily reduces the vestibular and proprioceptive weights.
Furthermore, the SD of P FINAL is the following: Our sensory fusion model predicts that the final tilt estimate is proportional to the actual tilt of the animal (P STIMULUS ). Specifically, in intact animals, both P OTO and P PROP are equal to P STIMULUS , leading to: Following BVL, the otolith signal is set to 0, leading to: We assume that the PSE measured during the SVV task reflects the final pitch estimate (P FINAL ). As a consequence, the coefficient of proportionality (w OTO 1 w PROP in Eq. 4a and w PROP in Eq. 4b) corresponds to the percentage of compensation measured in Figures  2B, 3B,D,F.
We model the brain's adaptation to BVL as follows: immediately after BVL, we assume that the brain uses the same sensory weights as in intact animals. Next, we assume that the brain adapts by recomputing optimal weights w PROP and w PRIOR in the absence of vestibular inputs. These weights can be computed by removing the term 1/s 2 OTO from Equations 2b, 2c. We used a gradient ascent procedure to match the model's prediction to experimental data in animals N and Z. Specifically, we adjusted the parameters to reproduce that average percentage of compensation across these animals, before (0.895) and after (0.205) labyrinthectomy and following adaptation (0.81). Furthermore, we also adjusted them such that s FINAL matches the average psychometric threshold (3.6°) measured in Elmore et al. (2019). Note that our model is identical as those of Clemens et al. (2011) and Alberts et al. (2016), except that these models assumed that vestibular and proprioceptive noises are tilt dependent, i.e., s OTO and s PROP increase as a function of P STIMULUS . This allows reproducing response non-linearities that occur when SVV is measured across a wider range of tilt (6120°), where the PSE does not vary linearly as a function of head tilt. Here, we found that PSEs were linear across a tilt range of 630°, and therefore simply used constant noise parameters s OTO and s PROP , which prevent our model from being underdetermined.

Data availability
Data supporting the study is available from the authors on reasonable request. Black dots and error bars represent the mean PSE and the 95% interval of the mean. Color bands represent the 95% confidence interval of a linear regression between PSE and pitch angle. The percentage of compensation is the opposite of the slope of the regression line. The slope and 95% interval of the slope are indicated. C, Perceptual threshold as a function of pitch angle, obtained from the curve fits in A. Error bars represent the 95% confidence interval.

Results
Intact macaques learn to transform head-fixed visual slant in an allocentric, gravity-referenced frame We trained three macaque monkeys to discriminate the orientation of a stereoscopically displayed visual plane relative to an allocentric earth-vertical plane, as animals, together with the visual display, were tilted in pitch (Fig. 1). Thus, the task was a SVV task adapted to rotation in the pitch plane, and required converting egocentric visual slant into an allocentric reference frame using allocentric earth-vertical cues which could only arise from vestibular and/or proprioceptive graviceptive signals.
To quantify the animal's performance, we computed psychometric response curves as a function of egocentric visual slant ( Fig. 2A). If animals perform the task accurately, the curves recorded at each pitch position ( Fig. 2A, color code) should shift by an amount opposite to the animal's pitch ( Fig. 2A, broken vertical lines), such that the PSE, i.e., the visual angle at which a surface appears vertical, would correspond to the egocentric position of the allocentric vertical plane. In all animals, we observed that the curves shifted in an approximately compensatory manner ( Fig. 2A). Importantly, the visual stimuli, including the range of slant angles tested, were always identical (relative to the animal's head) and independent of the monkey's pitch angle. Thus, the observed behavioral shifts, which show a consistent and systematic dependence on monkey pitch tilt ( Fig. 2A), do not merely represent a response bias but a true reference frame transformation.
We computed the animal's percentage of pitch compensation by performing a linear regression between PSE and pitch (Fig. 2B); a slope of À1 corresponding to 100% compensation. All animals showed a systematic compensation for pitch (animal Z: 92 6 11%; animal P: 104 6 14%; animal N: 86 6 7% compensation). We did not find a dependence of perceptual threshold on the animal's tilt (Fig. 2C), but the experimental protocol was not optimized to measure threshold (see Materials and Methods, Data analysis). Thus, in the following analysis, we focus our comparisons to the shifts in psychometric functions.

Acute labyrinthectomy impairs gravity-referenced visual transformation
To ascertain the role of the vestibular system in sensing gravity, we surgically destroyed the vestibular labyrinths bilaterally in two animals (N and Z) and repeated the allocentric task at days 4-11 (animal N) and 14-21 (animal Z; Fig. 3A,B). In both animals, the ability to integrate gravity information was severely compromised. Individual psychometric curves only shifted to a limited extent to compensate for the animal's pitch (Fig. 3A) and, on average, the animal's percentage of tilt compensation decreased to ;20% (animal N: 20 6 8%; animal Z: 21 6 13% compensation;   3B). The effectiveness of labyrinthectomies was ascertained by the absence of the VOR (Fig. 4).

The ability to perform gravity-referenced visual transformation recovers following labyrinthectomy
Despite total loss of vestibular function, we found that the animals' performance gradually recovered over a period of weeks (Figs. 3C-F, 5A). We estimated the percentage of compensation as a function of the number of experimental testing/training days following surgery (which differed from the total days after surgery; Fig. 5B). We found that both animals recovered at a remarkably similar rate, mostly during the first few weeks, where the two curves were practically superimposed (Fig. 5C). On the third week of testing (animal N: 18-22 d; animal Z: 28-32 d following surgery), the percentage of compensation reached 48 6 10% (animal N) and 44 6 13% (animal P; Fig. 3C,D). Animal N (yellow) eventually reached a performance similar to baseline (Fig. 5C, error bar on the right side). The performance in animal Z (green) fluctuated somewhat between t = 30 and 60 d before reaching stable levels after 60 d. (Fig. 3E,F).
We fitted the recovery curves with exponential functions to estimate the time constant of the recovery as well as its asymptotic performance. In animal N, gravity compensation recovered with a time constant of 19.5 d to a fitted asymptotic value of 92%, which fell in the confidence interval of baseline value (Fig. 5C). In animal Z, exponential fitting was affected by the fluctuation of the animal's performance. When all experimental data were fitted, we estimated that the time constant of recovery was 33.4 d, and its asymptote 82%, within the animal's baseline performance. When excluding data points between 30 and 60 d (Fig. 5C, broken black line), the asymptotic performance was similar (81%), but the estimated time constant of recovery was shorter: 23.2 d; a value that resembled that of animal N and may be more representative of the animal's initial recovery.

Sensory re-weighting model
In order to compare our results with previous experimental and modeling studies, we adapted the sensory fusion models of Clemens et al. (2011) and Alberts et al. (2016; Fig. 6A). We assumed that pitch signals originating from the otoliths and proprioception (P OTO and P PROP ) are weighted (weights w OTO and w PROP ) to form a final pitch estimate. The brain also uses information about the a priori statistical distribution of head pitch, which is modeled as a Gaussian centered on 0 with SD s PRIOR . The weights of the sensory modalities (w OTO and w PROP ) and of the prior (w PRIOR ) are determined by the amount of sensory noise (s OTO and s PROP ) as well as the SD of the prior (s PRIOR ), more accurate (i.e., lower s ) modalities being more heavily weighted (see Materials and Methods). Note that these weights always sum up to 1 (Fig. 6B); therefore, the prior weight w PRIOR reduces the other sensory weights and biases the final estimate toward 0. During SVV, the percentage of compensation is equal to the total sensory weight (w OTO 1 w PROP in intact animals, w PROP following BVL).
We fitted the model's free parameters (s OTO , s PROP , and s PRIOR ) to reproduce the percentage of compensation in animals N and Z, as well as the average psychometric threshold, measured in Elmore et al. (2019;3.6°) and analyzed the prediction of the best-fitting model (s OTO = 4.3°, s PROP = 7.9°, and s PRIOR = 15.6°). In intact animals, the vestibular  Figure 2B. The colored bands represent the 95% confidence interval of the percentage of compensation. Exponential curve fits to the average percentage of compensation are shown by solid black curves; the broken black curve represents curve fitted in animal Z where data between 30 and 60 d is excluded. D, Average perceptual threshold (across all psychometric curves recorded within a 5-d time window) as a function of the number of testing days since labyrinthectomy. The colored bands represent the confidence interval of the mean. In A, C, D, the baseline performance before labyrinthectomy is represented by broken lines and confidence intervals on the right side of the plots. signal's weight is ;3.5 times larger compare to the proprioceptive weight (Fig. 6B, 0.73 vs 0.21). The sum of these weights (0.94; Fig. 6C) corresponds to the percentage of compensation in intact animals (Fig. 2B, 0.89 in animals N and Z). In contrast, after BVL, the vestibular modality disappears. We assumed that, in an acute stage, the sensory weights are identical as in intact animals. In this situation, the remaining proprioceptive weight of 0.22 matches the percentage of compensation measured immediately after labyrinthectomy (Fig. 3B). We finally modeled chronic BVL by re-computing the optimal weights once the vestibular modality is removed from the model. We found that proprioceptive weight (and the corresponding percentage of compensation) increases to 0.79 (Fig. 6B,C), in agreement with our experimental findings (Fig. 3F). The model also predicts (Fig. 6D) that loss of vestibular function should lead to an increase of the psychometric threshold from 3.6°(consistent with the value measured Elmore et al., 2019) to 6.4°(which, however, could not be measured accurately in this study). These results indicate that the fundamental assumptions that underlie most current models of verticality sensing are sufficient to explain the acute and chronic effects of BVL.

Discussion
We have measured the relative contributions of vestibular and proprioceptive cues in the estimation of gravity used for visual orientation perception in a SVV task adapted to the pitch plane. Macaques estimated the visual slant with respect to earth-vertical at different body pitch orientations. The animals and the visual display were tilted together relative the allocentric vertical, such that completing the task required combining egocentric visual slant with allocentric body tilt, which could only be sensed through graviception. We found that bilateral labyrinthectomy resulted in a dramatic decrease in the animal's ability to sense body tilt, dropping from 90% to 20%, that should represent the contribution of proprioception in labyrinth-intact animals (although this value may be slightly overestimated since animals were tested a few days following surgery). We also deduce that the contribution of otolithic graviception in intact animals should be about ;70%. We found that both animals improved over several weeks, reaching performance similar to baseline, with a time constant of approximately three weeks of testing/retraining. This indicates that extravestibular gravity cues, which may originate from body graviception, gradually compensate for vestibular loss. The rate of recovery was similar across animals and, remarkably, seemed to be governed by the number of experimental sessions, rather than the absolute time elapsed since the surgery. This indicates that compensation is promoted by exercising the task. Note that one may have expected that the animals start recovering as soon as they start locomoting in their cages. Therefore, it is possible that recovery may be task specific.
Optimal sensory fusion posits that multiple sources of information are combined according to their respective reliability. In healthy subjects, both vestibular and proprioception should therefore contribute to graviception, although the former should have a greater weight. Our study provides, for the first time, a direct measure of these weights. Indeed, the percentage of compensation measured immediately after labyrinthectomy should reflect the proprioceptive weight. Furthermore, the difference between the percentage of compensation measured before and immediately after labyrinthectomy should reflect the vestibular weight. Here, we found that the proprioceptive weight was ;20% in both animals, and the vestibular weight ;70%. Optimal sensory fusion also Figure 6. Sensory fusion model. A, Overview of the model. Vestibular (green) and proprioceptive (red) sensory signals encoding head pitch, as well as prior information (gray) are combined to generate a final pitch estimate P FINAL (orange). B-D, Model simulation. Panel B shows the optimal sensory weights in intact animals and following BVL. Same color code as in A. After BVL, the vestibular weight is shown in light green color to emphasize that the vestibular modality is silenced although its weight is unchanged. Panel C shows the total weight of sensory signals, which corresponds to the percentage of compensation. Panel D shows the variability of the final pitch estimate.
predicts that the brain should reweight proprioceptive signals following labyrinthectomy, in agreement with our findings. It is notable that the best-fitting parameters (i.e., the SDs of vestibular and proprioceptive sensory noise, as well as the prior distribution of head tilt; s OTO , s PROP , and s PRIOR , respectively) in our study resemble those found in Clemens et al. (2011) and Alberts et al. (2016). These two models assumed that sensory noise were tilt dependent; however, when average in the range of 630°u sed here, we find Clemens et al. (2011) and Alberts et al. (2016) predict a vestibular noise of 5.1°and 7.8°, respectively (vs 4.3°in our study); a proprioceptive noise of a 11.9°and 7.5°, respectively (vs 7.9°in our study), and a prior of 12.5°in both models (vs 15.6°in our study). This indicates that the optimal sensory fusion model generalizes well over species. Optimal estimation also predicts that, if other sensory cues are provided, their weight should increase following BVL. Accordingly, other studies (Lopez et al., 2007;Toupet et al., 2017;Alberts et al., 2018) found that visual cues about allocentric tilt had a greater impact on SVV following vestibular loss. Thus, optimal sensory fusion has proven to be a largely successful framework for studying verticality perception in healthy subjects (De Vrijer et al., 2008Tarnutzer et al., 2009a,b;Clemens et al., 2011;Alberts et al., 2016) or following vestibular deficits (Alberts et al., 2015(Alberts et al., , 2018. More generally, optimal estimation models (Oman, 1982;Borah et al., 1988;Merfeld, 1995;Glasauer and Merfeld, 1997;Laurens and Droulez, 2007;MacNeilage et al., 2007;Angelaki, 2011, 2017;Karmali and Merfeld, 2012) are now the predominant approach to model vestibular information processing.
Comparing our study to previous work in humans indicates that verticality perception follows similar principles in humans and macaque monkeys with a rather close degree of quantitative agreement. This similarity is also supported by previous SVV studies in macaques (Daddaoua et al., 2008;Lewis et al., 2008). In particular, Lewis et al. (2008) demonstrated that the dynamics of illusory tilt experienced when rotated in a centrifuge is similar as in humans (Clark and Graybiel, 1966), an observation that strongly supports optimal model theories (Merfeld, 1995;Bos and Bles, 2002;Laurens and Droulez, 2007).
The majority of SVV studies (Daddaoua et al., 2008;De Vrijer et al., 2008Kaptein and Van Gisbergen, 2006;Lewis et al., 2008;Clemens et al., 2011;Alberts et al., 2015Alberts et al., , 2016Alberts et al., , 2018Bürgin et al., 2018) have been performed by tilting subjects in the roll plane. These studies generally report a typical pattern of error where subjects overestimate small tilt angles (e.g., ,30°; "E-effect") and underestimate large tilt angles (e.g., 60-150°, "A-effect"). One study (Ebenholtz, 1970) measured SVV in humans during both pitch and roll tilt, and found similar patterns of overestimation and underestimation. Here, we designed a SVV task where animals reported the allocentric orientation of a visual plane while being tilted in the pitch plane. We used a restricted range of body tilt, in which the A-effect does not occur. Furthermore, we did not observe any E-effect. This likely happened because the E-effect during head tilt is linked to a reflexive eye counterrotation that stabilized the eye in space. In our task, we required animals to fixate a head-centered point, effectively suppressing the ocular counterrotation and likely the E-effect. Therefore, our results in intact animals, in which pitch angles up to 30°were sensed accurately, are consistent with previous SVV studies. Furthermore, our modeling results resemble those of previous studies that measured SVV in the roll plane in humans (Clemens et al., 2011;Alberts et al., 2016). This indicates that our results are likely representative of SVV tasks performed at small tilt angles (,30°), in the pitch and roll planes, in humans and macaque monkeys.
Note that, rather than bars, here we used planar surfaces because they elicit robust responses from higher visual areas that may play a key role in creating a gravitycentered visual representation (Rosenberg et al., 2013;Rosenberg and Angelaki, 2014). Further, rather than using a manual bar task in which subjects physically align a bar with the earth-vertical (as in previous studies; see Daddaoua et al., 2008;De Vrijer et al., 2008;Kaptein and Van Gisbergen, 2006;Lewis et al., 2008;Bürgin et al., 2018), we have used a two-alternative-forced-choice slant discrimination task, as more recent human studies have employed (De Vrijer et al., 2009;Clemens et al., 2011;Alberts et al., 2015Alberts et al., , 2016Alberts et al., , 2018. Finally, although SVV studies typically consider roll tilts, here, we have generalized the findings to pitch orientations. Perception of visual slant orientation during pitch body tilts involves computation of orientation in depth, whose neural mechanisms are likely more intensive computationally.
How does the brain transform egocentrically encoded visual information into the gravity-centered representation of the world we perceive? A few older studies suggested early visual cortex may be involved (Denney and Adorjani, 1972;Horn et al., 1972;Sauvan, 1999;Tomko et al., 1981) but were subsequently refuted (Daddaoua et al., 2014). It is now well accepted that visual signals are first encoded in retinal coordinates. Thus, somewhere in the neural representation of the three-dimensional visual scene, multisensory gravity cues must interact with visual orientation signals to generate an allocentric representation. Where and how a gravity-centered, allocentric representation is achieved in the brain is unknown. Visual adaptation experiments (Mitchell and Blakemore, 1972) and clinical studies (Brandt et al., 1994;Funk et al., 2010Funk et al., , 2011Guardia et al., 2012;Rousseaux et al., 2015) suggest that a gravity-centered representation of object orientation arises in higher cortical areas. One candidate area is the central intraparietal area, where many neurons encoded visual plane orientation in a gravity-centered or intermediate reference frame (Rosenberg and Angelaki, 2014). Another area, suggested by clinical work, is the human homolog of parieto-insular vestibular cortex (Brandt et al., 1994;Baier et. al, 2012b;Funk et al., 2010Funk et al., , 2011 or temporo-parietal cortex (Kheradmand and Winnick, 2017). This broad area in human cortex additionally includes the homolog of the macaque visual posterior sylvian area, also known as "parieto-temporal association area T3" (Jones and Burton, 1976;Grüsser et al., 1990a,b;Guldin et al., 1992;Guldin and Grüsser, 1998;Dicke et al., 2008). Finally, brainstem and thalamic regions have also been implicated (Baier et al., 2012a(Baier et al., , 2016(Baier et al., , 2017. It is important that future studies explore the neural circuits mediating SVV perception.