Investigations into vestibular processing in the cerebral cortex date back to 1949, when recordings were performed in cats (35). Early studies utilized electrical stimulation of the vestibular afferent nerves in anesthetized animals (mostly monkeys) to identify regions, such as areas 2v and 3a, that respond to vestibular input [Reviewed in Guldin and Grüsser (25)]. Subsequent single-unit recordings in awake animals during whole-body motion produced by spinning chairs, motion platforms, and centrifuges revealed cortical areas encoding vestibular information related to self-motion. Several such regions have been identified in macaque monkeys. To date, robust vestibular representations of self-motion have been found in numerous areas, including the dorsal medial superior temporal sulcus (MSTd) (28, 36, 37), the ventral intraparietal area (VIP) (38–42), the visual posterior sylvian area (VPS) (43), parietal insular vestibular cortex (PIVC) (27, 40, 44, 45), the smooth eye movement region of the frontal eye field (FEFsem) (46), area 7a (47), and the posterior cingulate cortex (PCC) (48). These areas constitute a network for vestibular processing in the brain [Figure 1; (11, 25, 49)].

Over the past two decades, more sophisticated experimental designs utilizing a 6-degree-of-freedom motion platform that can produce arbitrary movements in both translation and rotation in 3D space have enabled refined studies of the vestibular representations in these areas (Figure 2A). In these experiments, animals are passively moved in darkness to activate the vestibular system. These vestibular areas exhibit distinct response properties. The proportion of neurons responsive to translational vestibular stimulation is highest in the PIVC (76.4%), followed by the VPS (72.3%), FEFsem (71.87%), PCC (68%), MSTd (64%), VIP (49%), and 7a (40%) (28, 40, 44, 47, 48). Chen and et al. analyzed the temporal properties of vestibular signaling across areas to clarify the pathway of vestibular signals flowing in the cerebral cortex and found that the PIVC leads the response, followed by the VPS and VIP, with the MSTd exhibiting the longest latency to external stimuli (40). This result may suggest that vestibular information flows across areas, consistent with the notion that vestibular signals reach the PIVC first after thalamic relay from the vestibular nucleus (VN).

Vestibular signals originate in the peripheral organs that are sensitive to acceleration stimuli (50). Accordingly, electrophysiological recordings in vestibular afferents show predominantly acceleration coding, particularly during translation of the head or whole body in darkness (51). However, in the central nervous system, including the brainstem, cerebellum, and cerebral cortex, researchers have found that vestibular neurons exhibit much more plentiful temporal signals, including acceleration, velocity, jerk (a derivative of acceleration), and even position, reminiscent of the motor cortex (52). Specifically, experimenters first found that neurons in different regions exhibit single-and double-peaked responses, corresponding to velocity and acceleration components in the stimulus, respectively [(53, 54); also see review by Cheng and Gu (11)]. These findings suggest that progressive integration (e.g., velocity) and differentiation (e.g., jerk) of acceleration-dominant peripheral vestibular signals occur in the central nervous system and may serve distinct functions.

In addition to the varied temporal dynamics, the central vestibular signals have spatial selectivity. A widely used metric for quantifying the strength of directional tuning is the directional discrimination index (DDI) (55), which assesses the contrast in firing rates between the preferred and null directions of motion regarding the variability of the neuron’s response. Based on this measure, vestibular spatial tuning strength could be assessed and compared across regions. For example, the PIVC, VPS, and VIP seem to carry the strongest spatial tuning, while the MSTd and FEFsem show robust, albeit slightly weaker, responses [(46); also see review by Cheng and Gu (11)].

Temporal and spatial properties are not two completely independent issues. For example, double-peaked neurons typically show one early-peak response in the preferred self-motion direction, while a second late peak appears in the anti-preferred direction. Laurens and et al. developed a three-dimensional (3D) spatiotemporal model based on responses across time and across different self-motion directions to systematically analyze spatiotemporal properties, aiming to characterize how vestibular signals are transmitted from otolith afferents (OAs) through the VN to cortical areas (56). In general, the model fits the experimental data well and reveals that OAs exclusively encode acceleration and jerk signals, while VN and cortical neurons show mixed selectivity for velocity, acceleration, jerk, and position. Thus, the model quantitatively describes to what extent vestibular signals undergo progressive integration as they propagate along the vestibular pathway. Consistent with the experimental observations that the PIVC appears to show the earliest response in the cortex, the model also reveals that neurons in this region have the highest proportion of acceleration signals among all cortical areas examined to date. In contrast, the MSTd is unique in preferentially encoding velocity over acceleration (Figure 1). Finally, other areas encode complex spatiotemporal signals with a rough balance between acceleration and velocity or a slight preference bias toward acceleration. In addition to velocity and acceleration, some cortical regions carry even higher-order temporal signals. For example, Liu and et al. found robust vestibular coding in single neurons in the PCC, and some neurons exhibit strong jerk and position signals (48), potentially related to the hippocampal navigation system.

Characterizing the temporal properties of vestibular signals in each brain area is key to understanding how self-motion is represented and utilized in guiding behavior. The relative weighting of different dynamics may relate to each area’s anatomical connectivity and functional role. For example, some vestibular regions, such as the VPS, PIVC and PCC, project back to the VN according to anatomical evidence (57). Thus, velocity signaling in these pathways may support gaze stabilization and other reflexes with coordinated eye, head, and body movements during natural navigation (58, 59). Similarly, the velocity dominant signals in the MST may also be related to gaze function when the eyes maintain steady pursuit of moving targets when accompanied by self-motion (49). Instead, the acceleration signals in many areas have been shown to be related to heading perception during self-motion (32, 33). In contrast to velocity and acceleration, jerk signals have rarely been studied to date, and their functions remain unclear. Some behavioral studies have examined the impact of jerk signals on subjects’ judgments on discrimination of self-motion strength, yet the conclusions are mixed (60–62). Researchers have indicated using computational modeling that jerk signals exist in some areas of the brain, including the PIVC and PCC (48, 56). Since these areas are within the Papez circuit and are strongly connected to regions involved in unpleasant emotional feelings, such as nausea, dizziness, and pain (63–65), we speculate that the jerk signals may be related to motion sickness during strong and unpredicted self-motion stimuli. Finally, some position signals have been reported in PCC neurons. Interestingly, PCC neurons show a bias for pitch and roll over yaw axes of rotation (48). A previous study also demonstrated that head direction cells in mice maintain fixed azimuthal tuning relative to the horizontal plane, even when the animal’s head orientation changes relative to gravity (66). These properties suggest that position signals in rotation may be critical for anchoring head direction signals to earth-centered coordinates during natural navigation. In summary, vestibular signals contain rich temporal dynamics, and different signals may have distinct functions that require further investigation.

In addition to the vestibular response, many cortical areas are also multisensory and respond to other sensory stimuli, such as optic flow, a cue heavily involved in self-motion perception and path integration during spatial navigation (67). Across regions, the proportion of neurons selective for optic flow is largely varied (Table 1) and is highest in dorsal visual pathways (e.g., the MSTd, FEFsem, and VIP) and lowest in vestibular dominant areas (the VPS, PIVC, and PCC). Unlike the dominant acceleration coding in the peripheral vestibular pathway, visual motion is represented predominantly as velocity signals in the brain (28, 68, 69). Across areas, spatial selectivity for optic flow is MSTd > VIP, FEF > VPS (11, 40, 46), which has an opposite tendency to vestibular flow.

The coexistence of vestibular and visual signals on individual neurons evokes challenges for cue integration when considering the congruency of their temporal and spatial properties. For example, while vestibular signals contain a wealth of temporal information, visual channels predominantly code velocity (70–72). Thus, how different temporal signals are integrated across modalities requires further consideration and investigation (see the following sections). Regarding spatial selectivity, the direction of preference in both modalities can be either congruent or incongruent. Importantly, there is a clear bimodal distribution, such that neurons tend to show highly congruent or nearly opposite heading preferences. This pattern is prevalent across areas, suggesting that it is a general rule in the brain. Interestingly, the VPS is dominated by opposite cells (43). Thus, while congruent cells are thought to be beneficial for multisensory integration (73), the roles of opposite cells are less understood (74).

In addition to the study of vestibular translational motion, some research has been conducted to descriptively analyze how neurons in various vestibular brain regions respond to rotational stimuli. The proportion of responsive neurons to vestibular rotational stimuli varies across different brain regions [MSTd (89%), VPS (75.9%), PCC (59%), VIP (44%), PIVC (49.4%), and 7a (31% for yaw rotation)] (44, 53, 55, 47, 48). Interestingly, the MSTd, which plays a crucial role in visual self-motion perception, has a significantly high proportion of cells responding to vestibular rotational stimuli. Moreover, the DDI is notably higher under vestibular rotation conditions than vestibular translation conditions. Notably, almost all cells under rotational conditions are vestibular-visual opposite cells (55), suggesting that the MSTd does not integrate these signals to produce a robust perception of self-rotation. To date, no conclusive explanation has been obtained from these observations. Due to the absence of behavior-related neural correlation studies, the significance of the proportions of neurons responding to rotation in various brain regions remains further identification in future experiments.

When observing a spatial variable (e.g., position or velocity), a reference frame can be defined as a specific viewpoint or perspective. The reference frame of the peripheral vestibular system is centered on the head. However, the vestibular reference frame in the central nervous system may be more complex because it must adapt to different functions, particularly when interacting with other sensory signals that are based on different reference frames. In particular, there are several possibilities:

Researchers have measured the spatial tuning functions of single neurons while intentionally manipulating the head, body, and eyes to identify reference frames. For example, Shaikh and et al. measured vestibular tuning in response to different whole-body translation directions while the head was at various fixed positions relative to the trunk (e.g., ±30°). If the spatial tuning functions remain unchanged and are not dependent on different head-on-trunk positions, it suggests that the recorded neuron represents a body-centered coordinate system. In contrast, if the spatial tuning curves exhibit systematic shifts (e.g., ±30°) with the head-on-trunk positions, it indicates that the neuron encodes a head-centered coordinate system. Using a similar approach, previous studies have successfully measured whether vestibular tunings shift with eye, head, or body orientations by varying eye-on-head, head-on-body or body-in-world positions. The central vestibular signals across regions and neurons could be represented in either of the reference frames except the eye-centered reference frame (Table 2; Figure 1).

As described in the previous sections, vestibular signals are thought to provide idiothetic information for heading perception. Psychophysical experiments have been conducted to test this rationale by using motion platforms or turntables to physically stimulate the vestibular system of humans and nonhuman primate subjects. Subjects are required to report their perceived self-motion direction, for example, heading to the left or right relative to a reference of straight forward. Numerous studies have shown that humans and nonhuman primates can judge their heading directions with a high degree of precision based on vestibular cues alone, particularly monkeys after overwhelming training [Figure 2B; (16, 30, 31, 82–84)]. In particular, the ability of subjects to discriminate heading directions is typically quantified by psychometric functions under a two-alternative forced-choice (2-AFC) experimental paradigm. Participants were found to reliably discriminate deviation from the straight forward direction within a few degrees. For example, macaques can detect deviations of 1°-3.5° based on vestibular cues alone (16), which is close to visual discriminability (31).

To explore neural correlates of vestibular heading perception, researchers have recorded neuronal activity in a number of areas, e.g., the MSTd, that have previously been shown to be modulated by physical motions (26, 37). These modulations in the extrastriate cortex have been shown to arise from a vestibular origin, since labyrinthectomy removes these signals (16). Gu and et al. further recorded vestibular responses while monkeys performed a heading discrimination task on a moving motion platform (16). Aided by receiver operating characteristic (ROC) analysis, researchers can construct neurometric curves to assess an ideal observer’s heading performance based solely on the firing rates of individual neurons. This method, under the framework of signal detection theory (85), allows the direct comparison of heading performance between the ideal observer and the animal. In the MSTd, researchers have found that only a small proportion of neurons exhibit a comparable threshold with the animals, while most of the other neurons show weaker discriminability. This pattern suggests that vestibular heading perception relies on integrating signals across populations of neurons (73) and probably across areas (86). This pattern has also been observed in subsequent studies in other areas, including the VN (87), cerebellum (88), VIP (54), otolith afferents (89), and PIVC (90), suggesting that pooling activity from populations of neurons is a prevalent mechanism in the brain.

In addition to the direct comparison between neuronal and psychophysical sensitivity, simultaneous neural recordings during discrimination tasks allow us to examine neuron-behavior correlations on a trial-by-trial basis. In particular, under identical stimulus conditions, fluctuations in neuronal activities are expected to significantly covary with the behavioral choice [the so-called choice probability or choice correlation, (85)] if the neurons are involved in the decision process. Such significant choice correlations have indeed been discovered in many areas, including the MSTd, VIP, VN, cerebellum and PIVC (16, 54, 88, 90). However, the exact functional implications of choice correlations should be interpreted with caution because numerous factors may lead to significant choice correlations without implying causal link functions, such as noise correlations and top-down feedback signals (91–93). Indeed, although significant choice correlations have been found to be prevalent in the central nervous system, they are essentially lacking in the peripheral afferents, presumably because feedback signals rarely reach the periphery (89).

Precise heading perception relies on multisensory cues. While vestibular or visual signals provide useful information about self-motion, they alone could be confounded in many situations (94). Thus, integrating cues can significantly overcome these limits by improving perception.

One significant benefit from cue integration is that it enables finer perceptual sensitivity than that based on either single cue (29). This is indeed the case in vestibular and visual integration: the psychophysical threshold in heading discrimination is reduced when congruent visual and vestibular cues are provided (Figure 2B), and the improvement level is close to the prediction of Bayesian optimal integration theory (31, 82, 95). The underlying neural circuit mediating multisensory heading perception remains unidentified. There are two proposed possibilities that can probably originate back to the weak and strong fusion framework proposed by Landy et al. in multisensory depth perception (96). First, in the early integration model, multisensory information about heading converges somewhere in early/mid stage of sensory cortices before sending to higher level decision areas for evidence accumulation, which is analogous to the strong-fusion framework. In contrast, the late integration model stipulates that unisensory signals are initially transmitted to downstream decision-related regions where integration occurs, reminiscent of weak fusion.

Stein et al. performed pioneering studies exploring principles of multisensory integration at the single-neuron level and discovered that efficient multisensory integration requires spatial and temporal coherence (97, 98). Given this, one might expect that visual and vestibular signals would have similar spatiotemporal properties to maximize the cue integration effect. However, recent findings in neurophysiology challenge this intuition.

There are at least two challenges in the spatial domain: the reference frame and spatial congruency of tuning curves. As mentioned above, different sensory modalities originate from different peripheral organs represented in distinct spatial reference frames. One intuition is that the different reference frames should unify somewhere in the central nervous system for cross-cue combination (99), yet this hypothesis is not supported by experimental evidence. In particular, while a head-centered coordinate of visual receptive fields in parietal areas, such as the VIP, has been reported previously [(100); but see (101)], tuning curves modulated by motion directions are mainly coded in an eye-centered reference frame in a number of cortices, including the MSTd (80, 81), VIP (41), V6 (102), and FEFsem (81). In contrast, vestibular signals are mainly based on a head-to body-centered reference frame in the MSTd, VIP, and FEFsem (77, 80, 81), although some neurons encode a head-centered coordinate that is modestly shifted to an eye-centered coordinate in the MSTd (80). Thus, most of the experimental findings in neurophysiology indicate that the visual and vestibular reference frames for self-motion perception are largely separated. While computational modeling works demonstrate that population readout of eye-centered visual and extraretinal signals could recover head-centered self-motion information, evidence of such head-centered coding in single neurons in sensory cortices remains lacking.

Another challenge is the presence of a large population of neurons that prefer nearly opposite self-motion directions indicated by vestibular and optic flow cues. Neurons representing conflict information exist across numerous sensory cortices [(43, 46); also see review by Cheng and Gu (11)] and sensory modalities, for example, horizontal disparity and motion parallax cues indicating depth (103), suggesting that this pattern is prevalent in the brain. Unlike congruent cells, neuronal tuning in opposite cells would be reduced when exposed to congruent sensory inputs. Thus, the opposite cells are unlikely to contribute to improved heading precision during cue combination (31, 73). The exact functions of this type of neuron require future studies with some proposed functions, including (1) distinguishing object motion and self-motion (74, 103, 104) and (2) serving as computational units for causal inference between multisensory integration and segregation (105).

Regarding the temporal domain, we have mentioned that visual signals are velocity-dominant in sensory cortices, whereas vestibular signals show varied temporal components in the central nervous system. According to the temporal-congruency principle in multisensory integration, it is expected that congruent visual and vestibular signals (i.e., velocity) should facilitate cue integration. The MSTd seems to be an ideal substrate. In particular, the MSTd neurons largely exhibit congruent visual and vestibular temporal dynamics, that is, velocity (28, 31). During the cue-combined condition, responses across the modalities are well-aligned and subsequently facilitate enhancement.

Hence, a specific population of neurons in the MSTd appears to meet both the spatial and temporal congruent principle in multisensory integration and thus may be beneficial for multisensory heading perception. However, recent experimental findings challenge this view. First, causal link experiments with electrical stimulation and chemical inactivation applied in MSTd fail to produce significant effects on the animals’ judgment of heading direction under vestibular conditions, while the effect is highly significant under visual conditions [Figure 3; (106)]. As a comparison, inactivating the PIVC significantly impairs the animals’ vestibular performance [Figure 3; (107)]. Thus, the MSTd seems to play a critical role in visual but not vestibular heading perception. Second, researchers have found that in a heading discrimination task based on a reaction time version (108), a human subject’s performance in the combined vestibular, visual and cue conditions is consistent with a model in which the brain accumulates vestibular acceleration and visual velocity evidence. Later, when macaques performed the multisensory heading discrimination task, Hou et al. (32) and Zheng et al. (33) recorded single neurons in the LIP and the saccade region of the FEF (FEFsac), two brain areas thought to be involved in perceptual decision-making tasks (109). They found that populations of neurons in both regions exhibit similar temporal dynamics. In particular, under the vestibular-only condition, ramping activities are more aligned with a process when momentary acceleration evidence is accumulated. In contrast, visual response dynamics are more aligned with the accumulation of velocity signals. Comparing the two stimulus conditions, vestibular ramping activity leads the visual by a few hundred milliseconds, which roughly corresponds to the lag between the acceleration and velocity peak in the motion profile used in the experiments. This result was further substantiated through causal inference experiments in which the researchers systematically manipulated the bandwidth of the Gaussian stimulus profile. By modulating the acceleration peak time while keeping the velocity peak constant, they could shift the acceleration profiles independently of velocity. Under these conditions, they observed that the ramping visual neuronal responses were unchanged, whereas vestibular response timing shifted in accordance with the changes in acceleration peak (32). Thus, either human psychophysical experiments based on reaction time tasks or monkey physiological experiments indicate that the brain may employ incongruent vestibular and visual temporal signals for heading judgment. This finding suggests that another possible model (the late integration model) is more likely to be true (Figure 3).

In macaque experiments, researchers have used a fixed duration paradigm in which the animals experience heading stimuli for a fixed amount of time (e.g., 2 s). Thus, it is unclear how downstream areas reconcile the discrepancy between vestibular and visual heading cues in the temporal domain. A few different strategies could integrate sensory evidence over time (Figure 4). One is to keep collecting sensory information up to the end of a trial (termed the “final readout” strategy). Alternatively, information may be continuously collected until the moment when upstream areas provide maximal information (“peak readout” strategy). Considering the two visuo-vestibular temporal integration models as aforementioned (vestibular acceleration and visual velocity versus velocity for both modalities), there are a few outcomes in the behavioral performance during the bimodal stimulus condition based on all possible combinations. First, for the “final-readout” strategy, either the visuo-vestibular temporal congruent or incongruent model would produce identical behavior because all information is ultimately collected by reaching the same level at the end of a trial. For the “peak readout” strategy, however, information is enhanced more in the temporal-congruent model at the moment when both cues reach their peak response, unlike in the temporal-incongruent model in which sensory cues are not aligned. Thus, the key question is which method the brain uses to perform the multisensory heading task.

To disentangle these possibilities, researchers have systematically varied the temporal offset between the visual and vestibular inputs and examined the behavioral performance. The trick is to artificially align the visual velocity peak in the motion profile with the vestibular acceleration peak by leading the visual input for a few hundred milliseconds. Interestingly, monkeys demonstrated better performance under this manipulation [Figure 4A; (33)]. Such a result first rules out the “final readout” strategy that predicts identical performance regardless of temporal offset between single cues. Second, the result is consistent with the prediction from the “peak readout” strategy with the visuo-vestibular temporal incongruent model. Thus, the behavioral performance under temporal-offset manipulated experiments suggests that the brain used vestibular acceleration and visual velocity signals to perform the heading task. Importantly, the behavioral pattern is consistent with neural activities simultaneously recorded in the LIP and FEFsac. In particular, information capacity in the population of sensory-motor association neurons is highest during the manipulation of visuo-vestibular sensory inputs, precisely explaining the behavior.

Result from the temporal offset manipulation experiment could also fit into the causal inference model (34). In particular, the brain still integrates the vestibular and visual signals when the two stimuli inputs are offset within a small amount (e.g., <500 ms), as reflected by the reduced psychophysical threshold compared to the condition when no offset is introduced. However, when the offset is large, for example, 750 ms, psychophysical threshold is instead increasing, suggesting that the brain is not integrating the two cues any more, but rather rely on one of the two cues.