Hippocampal spatial view cells for memory and navigation, and their underlying connectivity in humans

First, we tested whether neurons responded to different locations “out there” in space in an object‐location memory task. The monkey had to remember where on a video screen a particular visual stimulus had been seen previously. We discovered that some primate hippocampal neurons responded differently to different locations in space “out there” on the screen, and some neurons combined this with what picture had been shown in that location previously (Cahusac et al., 1989; Rolls et al., 1989).

Next, we tested for spatial view versus place‐related representations of primate hippocampal and parahippocampal gyrus neurons in a spatial environment. Macaques were moved on a platform mounted on a free‐moving robot in an open laboratory or on wheels in a cue‐controlled 2 m × 2 m × 2 m spatial environment. In the cue‐controlled environment, there was one room cue on each of the four walls, and the room cues could be moved. The test conditions allowed factors that might account for spatial firing of the hippocampal neurons, including the spatial location where the monkey looked, the place where the monkey was, and the head direction of the monkey, to be analyzed (Rolls & O'Mara, 1995). We discovered that the majority of the neurons with spatial responses had spatial view responses, that depended on where the monkey was looking in the environment, but not on the place of the monkey in the environment or on head direction (Rolls & O'Mara, 1995).

In addition, some neurons with place‐related firing in the primate were discovered: some neurons responded, for example, to the place where the macaque was located, to movement to a place, or to spatial view depending on the place where the monkey was located (Rolls & O'Mara, 1995).

It is important to define the reference frames for spatial representations, as they are relevant to understanding the functions that can be performed. An egocentric frame of reference (relative to the head or body) is useful for actions made in nearby space. An allocentric frame of reference (i.e., world‐based coordinates) is useful for remembering the location of objects and rewards in the world, independently of the one's body or head orientation or eye position. In primates and other animals with good vision or with echolocation, this can be independent of the place where one is located, though in rodents, the allocentric representation is more likely to be of the place where one is located, as distance vision is poor, and reliance is instead on local somatosensory cues using the vibrissae, local odors, and so forth. The discovery that some hippocampal neurons respond to the location on a video screen in front of the macaque (Rolls et al., 1989), and can even reflect the object shown in a particular location on the screen (Cahusac et al., 1989), raised the issue of which coordinate frame is used by the primate hippocampus. Feigenbaum and Rolls (1991) analyzed whether these spatial view neurons utilize allocentric or egocentric spatial coordinates. They moved the video screen and the macaque relative to each other, and to different places in the room. Then, 46% of the spatial neurons had firing that occurred to the same spatial location viewed on the display, or viewed in the laboratory, when the macaque was rotated or moved to a different place in the room. Thus, these hippocampal cells had spatial representations in allocentric (i.e., world‐based) and not in egocentric (relative to the body or head) coordinates. Also, 10% of the hippocampal spatial neurons had firing that stayed in the same location relative to the monkey's head/body axis when the video monitor was displaced, or the macaque was rotated, or was displaced to a different place in the room. Thus, 10% of the neurons represented space in egocentric coordinates, that is, relative to the head.

Feigenbaum and Rolls (1991) in addition showed that there were two types of allocentric encoding. For the majority of the neurons, the spatial field was in terms of its location on the video screen, independently of the place of the screen relative to the monkey's head axis, and independently of the place of the macaque and the monitor in the room. This type of neuron was termed “local frame of reference” allocentric, in that spatial fields of these neurons were defined by the local spatial frame that was provided by the video monitor. For the second type of allocentric encoding, the spatial field was defined by the location in the laboratory toward which the monkey was fixating, and was relatively independent of position with respect to the monkey's body/head axis or to position on the face of the video screen. This type of neuron was termed “absolute” allocentric, in that their spatial view fields were defined by the location in the laboratory that the animal foveated (Feigenbaum & Rolls, 1991).

Allocentric encoding is also a property of rodent hippocampal place cells, but the encoding is of the place where the rodent (rat or mouse) is, not of where in space the rodent is looking. However, the parallel is that in both cases allocentric encoding is found, and this allocentric representation is important for hippocampal computation and potentially for remembering where objects have been found in the environment (Kesner & Rolls, 2015; Rolls, 2018a; Rolls, 2021a; Rolls, 2021b; Rolls & Wirth, 2018).

In rodents, place cells respond best during active locomotion (Foster et al., 1989; Terrazas et al., 2005). To test whether place cells might be more apparent in macaques during active locomotion, as active locomotion was thought to be a key issue in rodents (Foster et al., 1989) and may be relevant in primates (Thome et al., 2017), single hippocampal and parahippocampal gyrus neurons were recorded while monkeys very actively walked on all four legs around the test environment with the head and body free to turn (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998). Also, to provide a good opportunity for primate hippocampal spatial neurons to reveal how they encoded space, the simple cue‐controlled environment (Rolls & O'Mara, 1995) was changed to a much richer open laboratory environment approximately 5 × 5 m (illustrated in Figures 1 and 2, and with, e.g., windows on walls 1 and 2) within which the macaque had a 2.5 × 2.5 m area in which to walk and forage for food. The place of the monkey and the head direction were tracked continuously while the monkey walked round the environment, and the eye position (which refers to the horizontal and vertical eye directions with respect to the head), were recorded continuously to enable measurement of where the monkey was looking in the environment at all times. The monkey walked round the test area, foraging for food, to enable measurements of neuronal firing for a wide range of places, head directions, and spatial views in a very wide range of different combinations to allow analysis of the relative importance of place, spatial view, and head direction in what was encoded by the neurons (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998).

The firing of a hippocampal spatial view cell during active locomotion in this spatial testing environment is illustrated in Figure 1. The neuron fired primarily while the macaque looked at a part of wall three, as emphasized in Figure 1b,c that show a spot on a wall where the monkey was looking when the firing rate was greater than 12 action potentials per s, half of the maximum firing rate. Figure 1b illustrates the finding that the neuron responded while the monkey was looking from different places in the room at the spatial view field on wall 3. The range of different places and head directions over which the hippocampal neuron fired is illustrated in Figure 1c. Analyses showed that this neuron responded to where the monkey was looking in space relatively independently of the place where the monkey was located, and of head direction and eye position. Moreover, the spatial view fields of the neuron were similar when the monkey was actively walking, and also when he was stationary but actively exploring with eye movements different parts of the spatial environment (Georges‐François et al., 1999). Videos to illustrate the firing of spatial view cells are described in the Data Availability Statement at the end of the paper and are provided as Supplementary Material.

The firing of a different hippocampal neuron is shown in Figure 2, to provide evidence, with a different type of analysis, about how the firing is related to spatial view, and not to the place where the macaque is located, or to head direction, or to facing direction, or to eye position. The highest firing of the cell, with the macaque at the place and with the head direction shown in Figure 2a, occurred when the macaque looked 10° left. With the monkey in another place and with a different head direction, the highest firing was when the macaque was looking 30° right, but at the same spatial view (Figure 2b). Figure 2c shows the firing with the macaque at a different place (but the same head direction as in Figure 2b), and the firing was now when the monkey looked approximately 30° left. The spatial view field was at the same place on Wall one as in Figure 2a,b, illustrating the allocentric spatial view encoding provided by spatial view neurons. Examples of video animations to illustrate the firing of macaque hippocampal and parahippocampal spatial view cells are described in the Data Availability Statement and are provided as Supplementary Material.

These experiments show that it is the allocentric spatial view toward which the monkey looks that determines the neuronal responses, and not a particular place where the monkey was located, or head direction, or facing direction, or eye position, and this was confirmed with analyses of variance and with information‐theoretic analyses (Georges‐François et al., 1999; Rolls et al., 1998). It was found that on average the spatial view cells encoded considerably more (mutual Shannon) information about spatial view (0.47 bits) than about eye position (0.017 bits), head direction (0.005 bits), or place in the room (0.033 bits) (Georges‐François et al., 1999). This shows that the encoding by these primate hippocampal neurons may reflect some information about place, and so forth but is primarily about spatial view. The coding is allocentric in that, as illustrated in Figure 2, the neuronal firing occurs when the macaque is looking at a given location in the world, independently of the head direction, facing direction, eye position, and place of the monkey.

The spatial view fields of these hippocampal and parahippocampal spatial view neurons typically occupy a region of space that is approximately as large as 1/16 of all the four walls of the laboratory (Rolls et al., 1998). Each neuron responds to a different view, and the partly overlapping view fields thus provide precise information about the region of space being looked at. Interestingly, some single neurons had more than one spatial view field in this extensive and rich spatial environment (Rolls et al., 1998). Information theoretic measures showed that the information about spatial view increases almost linearly as the number of neurons in the sample increases, thus showing that each of the neurons makes an independent contribution within the population to representing allocentric space (Rolls et al., 1998). Given that Shannon mutual information is a logarithmic measure (Rolls, 2021a; Rolls & Treves, 2011), this evidence indicates that the number of spatial views (or the accuracy of the spatial representation) increases exponentially as the number of neurons in the ensemble increases. This is an important result in terms of how information is encoded by hippocampal and parahippocampal spatial view cells as well as by neurons in other brain areas that encode different types of information (Rolls, 2016a; Rolls, 2021a; Rolls & Treves, 2011). Moreover, this is a firing rate code, with much information present in the number of spikes from a single neuron (Panzeri et al., 1999; Rolls, 2016a; Rolls, 2021a; Rolls & Treves, 2011).

Many hippocampal and parahippocampal spatial view (or “space” or “view”) cells were found in these experiments (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998). In the initial sample of 352 neurons recorded under these conditions, the number of spatial view cells was 40, or 11.4% (Rolls et al., 1997b). This was in a single environment (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998), and of course, the proportion of neurons would be expected to be higher if testing included several different environments. The spontaneous firing rate of these neurons was low (mean 0.5 spikes/s), and their mean peak firing rate was 17 spikes/s (interquartile range 11–20 spikes/s), consistent with these being hippocampal pyramidal cells, which were in both the CA1 and CA3 regions, with also some spatial view cells in the parahippocampal gyrus.

The finding from rodents that place cells respond better during active locomotion than passive motion (Foster et al., 1989) made it important to investigate primate hippocampal neurons during active locomotion (see also Thome et al., 2017). Having said this, Rolls et al. found that primate hippocampal spatial view cells have similar responses during active locomotion as when the monkey is not locomoting, but is looking around and actively exploring the spatial environment with eye movements. This is shown by the fact that spatial view fields are present when the monkey is stationary as illustrated in Figure 2, or is walking as in Figure 1 (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b); are found when the monkey is tested only when stationary (Rolls et al., 2005; Rolls & O'Mara, 1995; Rolls & Xiang, 2005; Rolls & Xiang, 2006); and as illustrated in the videos described in the Data Availability Statement and provided in the Supplementary Material. Indeed, it is an interesting hypothesis that this active exploration of a spatial environment, by moving the eyes from location to location in a viewed spatial scene in primates, is analogous to the active exploration performed by a rodent when it is locomoting from one place to another place.

To assess whether a neuron in the primate, including human hippocampal system, responds to the place where the individual is rather than spatial view, or head direction, or facing direction, or eye position, extensive testing with contrasts of these different hypotheses is needed (Georges‐François et al., 1999). (Eye position refers to the horizontal and vertical angles of the eye in the orbit.) If the views visible from different places differ, showing that the firing depends on the place where the individual is located is insufficient, because so does the spatial view. To separate spatial view from place cells, neurons must be tested while the individual is in one place with all of the different spatial views visible from there. Further, the same neuron must also be tested when the individual is located in a different place, but with at least many of the same spatial views visible, as has been implemented in Rolls et al.'s recordings in macaques (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998). Indeed, although hippocampal neurons in squirrel monkeys were found to respond when the monkeys were in a particular location in a 3D chamber (Ludvig et al., 2004), where the monkeys were looking was not measured, so we cannot contrast spatial view with place coding in this case. Similarly, Ono et al. (1993) found that when a monkey sitting in a cab was moved, some neurons responded when the cab was in specific places in the room. However, they were not able to factor out place from spatial view encoding in the type of factorial design that is necessary. These points will need to be taken into account for future investigations of hippocampal neuronal activity in humans and other primates (cf. Ekstrom, 2015; Ekstrom et al., 2003; Fried et al., 1997; Kreiman et al., 2000; Miller et al., 2013), and recording simultaneously the eye position, head direction, facing direction in the environment, and head position is needed. Only investigations in which the same set of spatial views has been seen from each of the same set of different places can provide evidence about whether the neurons code for spatial view or for place, or perhaps for a combination, and that condition was met in the experiments described above by Rolls et al. Investigations in which different spatial views are present when the individual is in different places do not address the issue about whether the neuronal encoding is about the location being viewed or the place where the individual is located. Of course, if the individual is always in one place, as in some human imaging studies, and neuronal responses are different when different scenes are being viewed, the implication is that the encoding is something about the scene being viewed, and not the place where the individual is.

Having made this point clear, it is nevertheless of interest that for humans there is now some evidence for medial temporal lobe neurons with properties like those of spatial view cells (Ekstrom et al., 2003; Miller et al., 2013), even though direct measures of eye position were not conducted. For example, in the study by Ekstrom et al., cells were found to represent the interaction between the place and the view faced by the patient. It is also of interest that in humans some medial temporal lobe neurons reflect the learning of paired associations between views of places, and people or objects (Ison et al., 2015), and this implies that views of scenes are important for human hippocampal function. Consistent with this, human functional neuroimaging studies do show hippocampal or parahippocampal activation when scenes or parts of scenes are viewed even when the human is fixed in one place for neuroimaging (Brown et al., 2016; Burgess, 2008; Chadwick et al., 2010; Chadwick et al., 2013; Epstein & Kanwisher, 1998; Hassabis et al., 2009; Maguire, 2014; O'Keefe et al., 1998; Stern et al., 1996; Zeidman & Maguire, 2016). Further evidence on the functioning of the human hippocampus is considered in Sections 3.9 and 5.

In rodents, the representation of place by hippocampal place cells can be updated by self‐motion, for example, running in the dark (Jeffery et al., 1997; McNaughton et al., 1991; Quirk et al., 1990). In monkeys, the representation of the location in the scene encoded by spatial view cells can be updated by self‐motion, for example, by the monkey moving the eyes in the dark, or by the monkey turning or walking in the dark. This was shown in experiments on these spatial view cells, in which the view was obscured by black curtains, in which many of the cells could still respond when the macaque moved his eye position to look toward where the view was visible previously (Robertson et al., 1998) (see example in Figure 3). This idiothetic update also occurs when the monkey locomotes in the dark, and then looks to a spatial view location. Some drift of the spatial view field over a few minutes when the curtains were closed was typical, consistent with the hypotheses that self‐motion (idiothetic) updating was occurring, and that the visual view details of the scene normally define the spatial view field of a neuron. It may be remarked that after about the same time in the dark with the curtains closed, the experimenters also lost their updating of place, head direction, and where they were looking in space, and had to find their way out of the environment by feeling for the curtains, and then following them to find a light switch.

These experiments (Robertson et al., 1998; Rolls et al., 1997b) show that primate hippocampal and parahippocampal gyrus spatial view neurons can be updated by self‐motion for short periods by idiothetic information including eye position, head direction, and place movements made by the monkey, and that the drift related to the temporal integration of these signals can be corrected when the scene again becomes visible. These experiments also show that these hippocampal system spatial view cells are different from the much more visual perception‐related responses of inferior temporal visual cortex object and face cells, which stop responding when the object or face is removed from visibility (Rolls, 2003; Rolls & Tovee, 1994).

The neurons had only a small decrease of their response when the room was placed into darkness and/or the view details were obscured with curtains in CA1, the parahippocampal gyrus, and the presubiculum. On the other hand, CA3 neurons had a larger decrease (on average to 23% of their normal response) when the macaque looked toward the normally effective location in the environment but the view was not visible (Robertson et al., 1998). There may be partial recovery of information in the CA3 network using autoassociation (Hasselmo & Wyble, 1997; McNaughton & Morris, 1987; Rolls, 1987; Rolls, 1989a; Rolls, 1989b; Rolls, 2018a; Rolls, 2021a; Rolls & Treves, 1994; Treves & Rolls, 1994), and further recovery in the associative synapses from CA3 to CA1, as has been shown analytically (Schultz & Rolls, 1999) and by simulations (Rolls, 1995). Another contributory factor to the difference might be the direct perforant path input to the CA1 neurons (Rolls, 2016a; Rolls & Treves, 1998).

A major issue in computational neuroscience is how information is encoded by populations of neurons, compared to single neurons (Rolls, 2021a; Rolls & Treves, 2011). For example, how does the number of stimuli, for example, spatial view locations, that can be encoded increase with the number of recorded neurons? To investigate this, we applied Shannon mutual information theoretic techniques useful for analyzing neuronal responses (described in detail elsewhere (Rolls, 2021a; Rolls & Treves, 2011) with code made available (Rolls, 2021a)) to the responses of macaque hippocampal spatial view neurons (Rolls et al., 1998). First, it was found that different hippocampal spatial view neurons tended to have different spatial view fields. It was then found that the information from an ensemble of these neurons about the 16 viewed locations increases approximately linearly with the number of cells in the population, which in this experiment was 20 neurons (Rolls et al., 1998). This indicates that the neurons convey independent information about spatial view (up to this number of neurons), and therefore, as information is a log measure, that the number of locations that can be encoded increases exponentially with the number of neurons in the population (Rolls, 2021a; Rolls et al., 1998). In information theory terms, this means that the “signal correlations,” that is, the correlations between the response profiles of each neuron to the set of stimuli, are low (Rolls, 2021a). An important further result was that when the decoding procedure for how the neurons encode location that was used to measure the information was made very biologically plausible, decoded by just a dot product of the firing of the neuronal population, then the information was almost the same. This type of decoding could be performed by the simplest model of a neuron that linearly sums the activity of each of its inputs, which is the simplest operation a neuron could perform (Rolls, 2021a). This makes the whole analysis biologically plausible. Decoding procedures that use non‐biologically plausible algorithms (Diamanti et al., 2021; Panzeri et al., 2022) may overestimate the information that is actually available for use by neurons in the brain.

In this investigation, the hippocampal neurons were not recorded simultaneously, and it is just possible that if the firing of the different neurons was cross‐correlated in time for some but not other stimuli (locations), then extra information might be available from these so‐called “noise correlations” (Panzeri et al., 2022; Rolls, 2021a). However, although that remains to be tested for primate hippocampal neurons, this is rather unlikely, for when macaque inferior temporal cortex neurons responding to faces or objects are simultaneously recorded and the effects of any possible noise correlations are tested, any effects found are small, and very much less than the large amount of information available from the number of spikes, that is, from the measured firing rates (Aggelopoulos et al., 2005; Franco et al., 2004; Franco et al., 2007; Rolls, 2021a; Rolls et al., 2004; Rolls, Franco, et al., 2003; Rolls, Franco, et al., 2006; Rolls & Treves, 2011). Indeed, stimulus‐dependent (i.e., “noise”) correlations between neurons may typically reduce the amount of information that is available from a population of neurons depending on how it relates to the signal correlations (Cohen & Kohn, 2011; Kanashiro et al., 2017; Panzeri et al., 2022; Ruff & Cohen, 2016), rather than perhaps solve the binding problem as had been previously suggested (Engel et al., 1992; Kreiter & Singer, 1996; Singer, 1999).

A number of approaches have analyzed population encoding in especially situations where the space being encoded is low‐dimensional, in, for example, the motor system, and the question has been raised of whether non‐linear decoding by neurons, which might not be very biologically plausible, might be used (Ebitz & Hayden, 2021; Keemink & Machens, 2019; Kriegeskorte & Wei, 2021; Saxena & Cunningham, 2019). But low‐dimensional spaces are not typical for the encoding of information in the cortex, where we may, for example, be able to recognize 10,000 different objects, thousands of spatial locations across many scenes, in the order of 10,000 episodic memories in the hippocampus, and so forth (Rolls, 2018a; Rolls, 2021a; Rolls & Treves, 2011). In this situation, a very large number of synapses on each neuron is required, in the order of 10,000 as is found, and the 10,000 dimensionality of this space is sufficient for these numbers of objects, locations, or episodic memories to be stored and later retrieved correctly, from autoassociation attractor memories and from pattern association memories (Rolls, 2021a; Rolls & Treves, 1990; Treves & Rolls, 1991). For this much more usual type of neuronal encoding of high‐dimensional spaces, the decoding can be linear, which is biologically plausible, and the computationally useful nonlinearity is introduced by the operation of competitive networks (Rolls, 2016a; Rolls, 2021a), such as those believed to be present in the dentate gyrus and CA1 (Rolls, 2016a; Rolls, 2018a; Rolls, 2021a; Rolls & Mills, 2019; Rolls, Stringer, & Elliot, 2006), as well as in the neocortex (Rolls, 2021c).

Virtual environments can be used to investigate spatial representations, and have advantages that the recordings can be made more easily than during navigation through real environments, and disadvantages that there are no vestibular and proprioceptive inputs related to motion, and at least in macaques and humans the body movements are likely to be very different (Minderer et al., 2016).

Macaque hippocampal neurons that respond to the place where the individual is located have been described in a virtual navigation task (Furuya et al., 2014; Hori et al., 2005), and some neurons also appeared to be related to view, but eye position recording was not available.

Eye position was recorded from macaques in a virtual navigation task in a star maze with five landmarks and a reward hidden between two of the landmarks (Wirth et al., 2017). It was reported that 28% (53/189) of hippocampal cells fired when the animals looked at one or sometimes more than one of the landmarks, and that 83% of these cells had their responses modulated by place. Then, 17% of the hippocampal neurons responded when the macaque was in one or more places in the virtual environment without significant modulation by where the animal was looking. Some neurons responded to combinations of view, place, and task context (Wirth et al., 2017).

The identification of neuronal responses that depended on where the animal was looking in a virtual navigation task was very interesting (Wirth et al., 2017). So was the discovery that some of the neurons responded when the monkeys moved their eyes to a new spatial location just before the view appeared on the screen in the VR task (Wirth et al., 2017). This provided clear evidence for idiothetic update of view‐related responses, with the idiothetic movement in this case the eye movement.

The study was also interestingly extended by showing that when the star maze and goal remained the same, but five new landmark cues were substituted, the hippocampal neurons quickly learned to respond to the new visual cues used for the landmarks (Baraduc et al., 2019). An interpretation is that the topological chart of the environment previously learned for the maze which links the landmarks, goal, and places in the environment (Battaglia & Treves, 1998) could remain relatively unaltered, and new room landmarks cues could be rapidly associated onto the existing chart of the star maze.

Some neurons with spatial view properties have activity that is modulated by the place where the individual is located (Rolls & O'Mara, 1995; Wirth et al., 2017). For example, the finding that many of the primate hippocampal neurons that responded to the sight of landmarks in a star maze did so mainly from certain places might have been because the navigational task constrained the individual to stay within the star maze, so that looking at some landmarks from some places, and with a particular view of the star maze in the foreground, might have not provided the opportunity for the landmarks to be seen in the same way from all places (Wirth et al., 2017), in contrast to the testing in an open field used by Rolls et al. (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998). Effectively, it could be that in a maze, in which only certain combinations of spatial view, place, body turn, and so forth can occur, neurons can learn to respond only to whatever conjunctions of sensory input actually occur. Instead, in the real world and in an open field, neurons may be much better able to learn spatial view representations that are invariant with respect to place because the spatial view can be seen from many different places. In rodents, something like this does happen in a hairpin maze in which the places, head directions, and views are constrained by the maze (Derdikman et al., 2009). This prevents the formation of the usual 2D place fields of hippocampal neurons found in open field environments, and instead results in place cells firing at only some parts of the hairpin track being followed, and probably representing the few combinations of place, head direction and view afforded by the hairpin maze (Derdikman et al., 2009). Another and interesting possibility is that some of these neurons only responded to a landmark when it was at a particular bearing, which is how an “allocentric bearing to a landmark” cell would respond (Rolls, 2020). Another possibility is that testing in a maze emphasizes looking at certain landmarks from certain places, as this is important for the navigation task (Wirth et al., 2017), so the type of combination representation just described of view, place, and head direction may have been formed because that was the sensory input provided in the maze. Indeed, it is further postulated that in the same environment, if a primate (including a human) is navigating to a place, as contrasted to navigating to a series of landmarks, the task demands and the inputs reaching the individual may have a modulatory influence on the extent to which spatial view versus place encoding is evident in the hippocampal system. This would be interesting to explore in future research. Outside highly learned and fixed navigational tasks, spatial view cells that are relatively invariant with respect to place are likely to be more important, for then the viewed location toward which one is navigating can be recognized independently of the current place of the individual (Rolls, 2021b).

In another series of investigations during virtual navigation in macaques, it has been found that the spatial properties of hippocampal neurons can be influenced by whether a task is being performed, though in this case it was not clearly possible to separate place from viewed location encoding (Gulli et al., 2020). For comparison and in contrast, in real‐world tasks in which macaques had to learn associations between view locations and objects or rewards, it was found that the spatial encoding was stable, in that the viewed location in the laboratory to which a spatial view cell was responding first had to be identified, and then with that location included in the experiment, associations of objects or rewards with that spatial view location could be identified (Rolls et al., 2005; Rolls & Xiang, 2005).

In another virtual reality investigation in macaques, of 88 hippocampal neurons with selective spatial responses, 32 responded to view, 12 to place, and 44 to both (Tan et al., 2021; Yen & Tan, 2023).

Useful confirmation has also recently been obtained that relatively many macaque hippocampal neurons respond to the location “out there” in space toward which the animal is facing (22% of neurons), compared to only 5% of hippocampal neurons that encode the place where the macaque is located (Mao et al., 2021). Some neurons were classified as spatial view cells and others as “facing location” cells, but the environment being viewed was simple (a cylindrical arena with a drain on the floor and two touchscreens with food on the walls), and more spatial view cells are likely to be found in a rich spatial environment such as the open lab that we used (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998). Indeed, the reason that we moved to a rich open lab visual environment was that we expected to find, and did find, more spatial view cells than in a relatively simple spatial environment with only four cues in the testing arena (Rolls & O'Mara, 1995). Spatial view cells in our testing environments were found to respond to where the macaque was looking in space, and not to the location toward which the individual was facing, by testing these specific hypotheses (Georges‐François et al., 1999; Rolls et al., 1997b; Rolls et al., 1998; Rolls & O'Mara, 1995). A clear example showing that spatial view cells of the type that we have described code for where the monkey is looking in allocentric space and not where he is facing is shown in Figure 2. Further evidence is that in the dark, spatial view cells respond to a remembered spatial view location only when that location is being looked at, with facing location held constant (Robertson et al., 1998) (Figure 3).

In terms of brain computations, it is computationally useful for spatial view cells to respond to viewed allocentric locations in a natural scene that has many useful and clear landmarks, even if an individual is not facing those locations but is looking at them, because it is where objects or landmarks are in the real world, not where one is facing, that is important for memory of where objects are in the world and navigation to a location in the world (Rolls, 2021a). To make this point very clear, a representation that depends on facing direction (Mao et al., 2021) is not very useful because it is not invariant with respect to place. That is, to find an object or reward in a scene, the individual has to go to the place where the memory was formed, and then when facing in the correct direction, facing location neurons would fire and the object or reward associated with that could be retrieved from memory. In contrast, spatial view cells are allocentric, and invariant with respect to eye position, head direction, and place where the individual is located, so that wherever the individual is placed, a spatial view cell will respond, and potentially enable the recall of the object or reward at that viewed location (Georges‐François et al., 1999; Rolls & O'Mara, 1995). That true allocentric representation provided by hippocampal and parahippocampal spatial view cells is thus much more useful than a representation that is based on “facing location.” Of course, when primates navigate they may often be facing in the direction in which they are navigating. But that does not mean that allocentric spatial view is not being encoded, and experiments of the type illustrated in Figure 2 show clearly that allocentric spatial view is being encoded by spatial view neurons (Georges‐François et al., 1999; Rolls et al., 1997b; Rolls et al., 1998; Rolls & O'Mara, 1995).

For humans, there is evidence for medial temporal lobe and hippocampal neurons with properties like those of spatial view cells, for example, with responses to locations being viewed (from recordings in patients during neurosurgery) (Ekstrom et al., 2003; Miller et al., 2013). In the study by Ekstrom et al. (2003), some medial temporal lobe neurons were found to represent views of landmarks. In another study of human medial temporal lobe neurons, it was found that in a Treasure Hunt game, some neurons respond to the sight of remote locations rather than the subject's own place (Tsitsiklis et al., 2020). Just like macaque spatial view cells, these neurons in humans respond when the spatial location is seen with different bearings (showing that they are not “allocentric‐bearing‐to‐a‐landmark” neurons, but spatial view neurons). The locations in the human Treasure Hunt game were in at least some cases within the spatial environment that could be viewed. In the macaque testing, hippocampal spatial view neurons could respond when the macaque was distant from an effective part of the 3D environment (e.g., the location in the scene where a trolley was located), but also when the macaque was close to the effective part of the environment (e.g., at the place where the trolley was located, as illustrated by Rolls (1996a, 2021a)). This is thus somewhat comparable to the way in which the human visual “spatial target” neurons responded (Tsitsiklis et al., 2020). The results in humans (Tsitsiklis et al., 2020) thus appear to confirm the presence of spatial view cells in humans that were discovered in macaques (Feigenbaum & Rolls, 1991; Rolls et al., 1989; Rolls et al., 1997b; Rolls & O'Mara, 1995). In addition, some neurons have been recorded in humans that respond during navigation toward the location of a particular goal in a virtual environment (Qasim et al., 2019; Qasim et al., 2021). Further, in humans, some medial temporal lobe neurons reflect the learning of paired associations between views of places, and people or objects (Ison et al., 2015) (just as in macaques (Rolls et al., 2005)), and this implies that neurons coding for views of scenes are important for human hippocampal function.

Consistent with this, human functional neuroimaging studies do show hippocampal or parahippocampal activation when scenes or parts of scenes are viewed even when the human is fixed in one place for neuroimaging (Brown et al., 2010; Brown et al., 2016; Burgess, 2008; Chadwick et al., 2010; Chadwick et al., 2013; Epstein & Kanwisher, 1998; Hassabis et al., 2009; Maguire, 2014; O'Keefe et al., 1998; Zeidman & Maguire, 2016). That is further evidence that representations of space “out there” are key to hippocampal function in primates including humans. Further, using an adaptation paradigm while participants viewed snapshots of a virtual room which differed in place, spatial view, and heading, it was found that the pattern of hippocampal activity reflected both view‐based and place‐based distances, the pattern of parahippocampal activity preferentially discriminated between views, and the pattern of retrosplenial activity combined place and view information (Sulpizio et al., 2014). Using a somewhat similar approach, evidence for encoding of heading direction in the human presubiculum was found (Vass & Epstein, 2013), consistent with the head direction cells found in the macaque presubiculum (Robertson et al., 1999).

In the human parahippocampal gyrus, neurons have also been described in virtual navigation that represent egocentric directions toward “anchor points” (Kunz et al., 2021), which are similar to the landmarks described in Sections 5, 6 (Rolls, 2020). Some of these neurons also encoded the distance to the landmarks. Neurons of this type may be part of the interface to actions performed in space in which the human parietal cortex which has connectivity with the parahippocampal gyrus (Rolls, Deco, et al., 2022c; Rolls, Deco, et al., 2022d; Rolls, Wirth, et al., 2022) is important (Sections 5.3 and 8). In humans, an fMRI investigation showed that the distance to home may be represented in the hippocampus and retrosplenial cortex (Chrastil et al., 2015). Some other parahippocampal gyrus neurons encoded allocentric direction (such as “West”) (Kunz et al., 2021), and might be similar to the head direction cells described in Section 3.13 (Robertson et al., 1999).

To further elucidate the properties of primate hippocampal spatial view cells, they are now compared to object cells in the macaque inferior temporal visual cortex, and then to “concept” cells in humans, and then to head direction cells.

Hippocampal spatial view cells are quite different to inferior temporal visual cortex (IT) cells that respond to faces or objects (Aparicio et al., 2016; Arcaro & Livingstone, 2021; Booth & Rolls, 1998; Freedman, 2015; Freiwald, 2020; Freiwald et al., 2009; Hasselmo et al., 1989; Perrett et al., 1979; Perrett et al., 1982; Rolls, 2000; Rolls, 2021a; Rolls, 2021f; Rust & DiCarlo, 2010; Tsao, 2014) wherever they are moved to in a spatial environment (Aggelopoulos et al., 2005; Rolls, 2012; Rolls, 2016a; Rolls, Aggelopoulos, & Zheng, 2003). On the other hand, it is normally the combination of a set of features in a fixed position relative to each other in the world that activates spatial view neurons (Feigenbaum & Rolls, 1991). A helpful distinction is that objects can be moved to different places in the environment, and visual temporal cortex object‐selective neurons respond to an object independently of its location in a scene (Aggelopoulos et al., 2005; Rolls, Aggelopoulos, & Zheng, 2003). In contrast, parts of a spatial scene are fixed with respect to other parts of the scene, and cannot be moved independently with respect to the other parts. Thus, although hippocampal spatial view cells can respond to stimuli that are a fixed part of a spatial scene (e.g., table T2 in Figure 1 of Robertson et al. (1998)), the point is that this is a fixed part of a continuous spatial scene. Spatial scene representations may be learned by associating together features in a scene that have a fixed spatial relationship to each other (Rolls & Stringer, 2005; Stringer et al., 2005), and this is quite different from invariant visual object learning in which the features of a single object are associated together, because of regular association of the parts of a single object that are independent of the background scene and other objects in the scene that are not constant (Rolls, 2012; Rolls, 2016a; Rolls, 2021f; Stringer et al., 2007; Stringer & Rolls, 2008). The inputs to the hippocampal formation that help it to form spatial view representations may come from areas such as the occipital place area (Julian et al., 2016), and from scene processing areas in the macaque temporal cortex (Kornblith et al., 2013) and human PSA (Rolls, Deco, et al., 2022b).

Another difference from IT neurons is that many hippocampal spatial view neurons, because they represent parts of space, can respond even when the scene is not visible but that part of space is looked at (Rolls et al., 1997a); and can be updated idiothetically, that is by self‐motion, in that spatial view neurons respond when a macaque moves the eyes to a location in space even when no scene is visible, and in darkness (Rolls et al., 1997a) (see Section 3.5).

Another difference is that IT neurons respond well to visual stimuli in an object‐reward association task (Rolls, Aggelopoulos, & Zheng, 2003; Rolls et al., 1977), but hippocampal neurons have weak responses to objects in this type of non‐hippocampal‐dependent task, compared to the stronger object‐related responses that can occur in an object‐place, hippocampus‐dependent, task (Rolls & Xiang, 2005).

In primates, hippocampal neurons have been described that respond in an invariant way to the sight of individual faces (Quiroga et al., 2005; Sliwa et al., 2016). The neurons found in humans described as “concept cells,” an example of which is a neuron that responded to Jennifer Aniston, may respond not only to Jennifer Aniston, but also to other actors in the same movie, and the places with which they are associated (De Falco et al., 2016; Quiroga, 2012; Quiroga et al., 2005; Rey et al., 2015). Object‐place cells in macaques have some similar “concept” properties, in that they can be activated either by the object, or by the place, in object‐place memory tasks (Rolls et al., 2005; Rolls & Xiang, 2006). Similar properties have been described for human hippocampal neurons, in a task in which a human was associated with a place (Ison et al., 2015).

It can also be emphasized that spatial view hippocampal and parahippocampal cells are quite distinct from head direction cells, found in the primate presubiculum and parahippocampal gyrus (Robertson et al., 1999). For instance, if the head direction remains constant when the macaque is moved to different places in the environment where the spatial view differs, spatial view cells provide different responses. On the other hand, head direction cells have activity that remains constant for a particular head direction, even though the spatial view differs completely (Robertson et al., 1999).

In the rodent entorhinal cortex, grid cells that represent places by hexagonal place grids and are involved in idiothetic update of place have been described (Gerlei et al., 2021; Kropff & Treves, 2008; Moser et al., 2015). In macaques, a grid‐cell like representation in the entorhinal cortex has been found, but the neurons have grid‐like firing as the monkey moves the eyes across a spatial scene (Garcia & Buffalo, 2020; Killian et al., 2012; Meister & Buffalo, 2018; Rueckemann & Buffalo, 2017). Similar competitive learning processes to those suggested for rodents (Rolls, Stringer, & Elliot, 2006) may transform these primate entorhinal cortex “spatial view grid cells” into primate hippocampal spatial view cells (Rolls, 2021a), and may contribute to the idiothetic (eye movement‐related) update of spatial view cells (Robertson et al., 1998). The existence of spatial view grid cells in the entorhinal cortex of primates is predicted from the presence of spatial view cells in the primate CA3 and CA1 regions (Kesner & Rolls, 2015; Rolls, 2013; Rueckemann & Buffalo, 2017). Moreover, some of these “spatial view grid cells” have their responses aligned to the visual image (Meister & Buffalo, 2018), as predicted (Kesner & Rolls, 2015).

In the human entorhinal and cingulate cortex neurons with grid‐like response properties are found (Jacobs et al., 2013; Nadasdy et al., 2017), and there is neuroimaging evidence that is consistent with this (Julian et al., 2018; Nau et al., 2018). This is further evidence for the concept that representations of locations being viewed in space “out there” are a key property of spatial representations in the hippocampal system of primates including humans.

To perform idiothetic update of a spatial representation (such as that provided by spatial view or place cells), a self‐motion signal is needed to update the spatial representation. The idiothetic signal might usefully be a velocity of movement signal. This velocity signal might have its origin in vestibular signals about motion, in optic flow, and/or in corollary motor discharge (Bremmer, Duhamel, et al., 2002; Bremmer, Klam, et al., 2002). Neurons that do respond to self‐motion signals have been discovered in the primate hippocampus, in an investigation in which the monkey was moved while sitting on a robot with defined axial rotations and linear translations, and in a test situation in which optic flow visual motion cues could also be produced by rotating the whole environment round the monkey (O'Mara et al., 1994). The neurons respond to the velocity of whole body motion (O'Mara et al., 1994), which is idiothetic information. For instance, some neurons have larger responses for clockwise than for anti‐clockwise whole body rotation. Occlusion of the visual field showed that some of these neurons depend on visual input. For other neurons, there was no requirement for visual input, and these neurons probably responded to vestibular input. Other neurons responded to a combination of whole‐body motion and view or place. Of the 45 neurons with responses related to whole body motion (9.8% of the population of hippocampal neurons recorded), 13 responded to axial rotation only, 9 to linear translation only, and 20 neurons to axial rotation or to linear translation. The sign of the motion was important for some of the neurons, with different responses for clockwise versus anticlockwise rotation, or for forward versus backward linear translation, which are different velocities. Some neurons responded to a combination of whole body motion and either a local view (n = 2) or a place toward which the macaque was moving (n = 1).

Whole‐body motion neurons are likely to be a useful component of a memory system for memorizing spatial trajectories through environments for path integration that is useful in short‐range spatial navigation (O'Mara et al., 1994). They may provide self‐motion information useful to provide the idiothetic update of spatial view cells. Consistent with this discovery (O'Mara et al., 1994), neurons have more recently been found in the rat entorhinal cortex that have a linear response with linear running speed, and have been termed “speed cells” (Hinman et al., 2016; Kropff et al., 2015).

A second principal type of neuron in primates that provides idiothetic information useful for the update during self‐motion of spatial view neurons (Robertson et al., 1998), as well as for navigation, is head direction cells, well known in rodents (Cullen & Taube, 2017; Taube et al., 1990), which we discovered in the primate presubiculum (Robertson et al., 1999) (and they are probably elsewhere). Head direction neurons are likely to be important in updating spatial view cell representations of allocentric space and for navigation when the view details are obscured or in the dark (Rolls, 2020; Rolls, 2021b).

Head direction neurons continue to encode head direction even when the monkey is moved from a familiar room to a relatively unfamiliar corridor, and maintain their directionality for a few minutes in the dark, after which they drift (Robertson et al., 1999). This is important, for these cells can only maintain head directionality for a relatively short period without visual cues to lock them back into the correct directionality. Their inputs are derived from velocity signals produced in the vestibular nuclei in the brainstem and reach the parietal vestibular cortical areas (Cullen, 2019; Grusser et al., 1990; Ventre‐Dominey, 2014). The direction signal thus reflects a great deal of integration over time, and this is imprecise and noisy resulting in drift. This means that only short‐term idiothetic navigation (i.e., without visual cues) is possible. Vestibular signals influence neurons in a number of parietal cortex areas including VIP, with neurons that respond to head position (i.e., head direction) or head acceleration, in addition to the many neurons with head velocity tuning (Klam & Graf, 2003). Neurons that respond to vestibular inputs produced by head rotation or translation are also found in area 7a (Avila et al., 2019). The parietoinsular vestibular cortex may be especially important in the sense of direction (Chen et al., 2016).

A key issue is whether the primate including human hippocampus is for memory, or for navigation. There is emphasis on navigation for place cell function in rodents (Burgess et al., 2000; Burgess & O'Keefe, 1996; Hartley et al., 2014; O'Keefe, 1979; O'Keefe, 1991). However, the hippocampus is implicated in episodic memory in which the location, or temporal position in a sequence of a single episodic memory, is associated with, for example, the associated objects or rewards (Dere et al., 2008; Eichenbaum et al., 2012; Hasselmo, 2009; Kesner & Rolls, 2015; Rolls, 1990; Rolls & Mills, 2019; Treves & Rolls, 1994; Zeidman & Maguire, 2016). If the hippocampus helps to implement episodic memory, then object information would need to reach the hippocampus, where it might be combined with spatial view information to form for example, an episodic memory of a person or object seen in a viewed location.

To investigate the fundamental issue of whether object information, as well as spatial view information, is provided in the primate hippocampus, single hippocampal neurons were recorded during an object‐place memory task in which the monkeys had to learn associations between objects and where they were shown in an open laboratory (Rolls et al., 2005). Some neurons (10%) responded to an object independently of its location; other neurons (13%) responded to spatial view independently of the object shown; and some neurons (12%) fired to a combination of a particular object and the particular location where it was shown in the laboratory. Thus, in the primate hippocampus, there are separate as well as combined representations of objects and of their locations in space. These properties are needed in an episodic memory system, for associations between objects and where they are seen are prototypical for episodic memory. These discoveries provide evidence that a key requirement for a human episodic memory system, both separate and combined neuronal representations of objects and their locations “out there,” are present in the primate hippocampus (Rolls et al., 2005). These neurons might also be termed object‐spatial view neurons, to emphasize the difference from what is found in rodents. Neurons that correspond have now been described in rodents, but they, as expected, encode item‐place, not item‐spatial view, combinations (Komorowski et al., 2009). In the rodent investigation, the items were odors.

Consistent with these discoveries, the hippocampus receives projections from the temporal cortical areas specialized for objects or faces (Huang et al., 2021; Perrett et al., 1982; Rolls, 2000; Rolls, 2021a; Rolls, 2021f;Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022d), and neurons responsive for particular individuals were found both in the human medial temporal lobe (Kreiman et al., 2000; Quiroga, 2012) and the monkey hippocampus (Sliwa et al., 2016).

A feature of the theory of the hippocampus in episodic memory is that object and location memories should be capable of being formed in one trial, in order to be relevant to the timescale of episodic memory, and that the whole memory can be recalled from any part (Kesner & Rolls, 2015; Rolls, 1989b; Rolls, 1996b; Rolls, 2016a; Rolls, 2018a; Rolls, 2021a; Rolls & Kesner, 2006; Treves & Rolls, 1994). This was tested in macaques in a one‐trial memory task in which a viewed object had to be associated with its viewed spatial location. The task involved the storage of object‐location information, and then the recall of the object when the location was presented as a recall cue, and the recall of the location when the object was presented as a recall cue (Figure 4a). The design is similar to that of a one‐trial odor‐place recall memory task that is hippocampal‐dependent in rats (Day et al., 2003), and is quite different from a long‐term visual–visual associative memory task which is implemented in the perirhinal and related cortex (Fujimichi et al., 2010; Hirabayashi et al., 2013; Naya et al., 2001). Images of novel objects were used every day, and within a day, the same objects were used, so that the one‐trial recall task was difficult. Recordings were made from 347 hippocampal neurons during the performance of the object‐location memory storage and recall task (Rolls & Xiang, 2006). Some neurons performed object recall, when the recall cue was a viewed location (Figure 4b). Some neurons performed location recall, when the recall cue was an object (Figure 4c). The recall‐related firing is evident in stage 4, when the object or location was being recalled with no stimulus present on the screen. Details of the results are provided elsewhere (Rolls & Xiang, 2006). The findings provide evidence that the macaque hippocampus can provide for one‐trial object‐view association learning of the type that is prototypical for episodic memory (Rolls & Xiang, 2006). Rapid changes in neuronal response properties as a result of learning associations between items such as individuals and places have been confirmed in humans (Ison et al., 2015).

In humans, in an object‐place recall task in virtual reality, some neurons during recall also reflect the recall of the place when the object recall cue is provided (Miller et al., 2013). Further, it has been shown that time as well as space is encoded in the primate hippocampus, and that some hippocampal neurons recalled the time at which a specific object was seen (Naya & Suzuki, 2009).

Information about where rewards are located is a key attribute of an episodic memory system. The anterior hippocampus of primates (which corresponds to the ventral hippocampus of rodents) has inputs from brain areas such as the orbitofrontal cortex and amygdala that perform reward processing (Carmichael & Price, 1995; Huang et al., 2021; Pitkanen et al., 2002; Rolls, 2022; Rolls, Deco, et al., 2022d; Rolls, Deco, et al., 2022e; Stefanacci et al., 1996; Suzuki & Amaral, 1994). This connectivity is also proposed to provide information to the human hippocampus about reward location that is needed to provide the goals for navigation (Rolls, 2022; Rolls, Deco, et al., 2022e).

To analyze reward‐related input to the primate hippocampal system, neuronal activity was recorded during a reward‐spatial view association task in monkeys in which one location in each spatial scene on a video monitor, when touched, resulted in a fruit juice reward, and a second location resulted in a less preferred juice reward. The different scenes had different locations for the two reward types (Rolls & Xiang, 2005). Then, 18% of 312 hippocampal cells analyzed responded in different scenes to the location of the preferred reward, and 5% to the location of the less preferred reward (Rolls & Xiang, 2005). Of 44 neurons tested, 60% reversed the location to which they responded when the locations of the preferred rewards were reversed in the scenes, providing evidence that the reward‐place associations could be relearned in a few trials. Most (82%) of the 44 location‐reward neurons in the hippocampus did not respond to object‐reward associations in a visual discrimination task. Thus, the macaque hippocampus represents the reward associations of locations being viewed “out there,” and can store affective information as part of an episodic memory. This provides a way in which the current mood or reward/non‐reward state may influence the retrieval of episodic memories, which is of interest for psychiatric disorders in which sad memories may be emphasized because of altered functional connectivity of the orbitofrontal cortex with hippocampal memory mechanisms (Cheng et al., 2016; Cheng et al., 2018; Rolls, 2016b; Rolls, 2018b; Rolls, 2019b; Rolls et al., 2018; Rolls, Cheng, & Feng, 2020).

There is further evidence that neurons in the primate hippocampus are influenced by rewards. Wirth et al. (2009) described cells that encoded the reward outcome of a trial in an object‐place association task. These findings have been extended by showing that macaque hippocampal neurons reflect outcome information more than prefrontal cortex neurons (Brincat & Miller, 2015). In addition, the reward value of pictures is also represented in the primate hippocampus (Knudsen & Wallis, 2021; Knudsen & Wallis, 2022).

The results indicate that the primate hippocampus can learn associations between viewed locations and objects (Rolls et al., 2005) or rewards (Rolls & Xiang, 2005). Perhaps correspondingly but with a different representation of space, the responsiveness of rodent place cells can be influenced by where rewards are available (Hölscher et al., 2003; Tabuchi et al., 2003). The principle is that the hippocampus may encode information about where emotion‐related (rewarding or punishing) events happened; may be involved in the recall of emotions when particular places are seen later; may provide mechanisms by which the current mood can influence the memories that are recalled; and may be involved via value‐related inputs in whether memories are consolidated (Rolls, 2015; Rolls, 2018b; Rolls, 2022; Rolls, Deco, et al., 2022d; Rolls, Deco, et al., 2022e).

Connectivity is directed to the human hippocampus from a medial posterior part of the parahippocampal gyrus PHA1‐3 (which corresponds to TH in macaques) and the adjoining ventromedial visual areas (VMV1‐3) (Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022d) (Figures 5 and 6). This parahippocampal/VMV region (Sulpizio et al., 2020) is a parahippocampal place area (or PSA, PSA, as it responds to viewed scenes not the place where the individual is located) (Epstein, 2005; Epstein, 2008; Epstein & Baker, 2019; Epstein & Julian, 2013; Epstein & Kanwisher, 1998; Kamps et al., 2016; Natu et al., 2021; Sulpizio et al., 2020). Thus, my proposal is that the PSA is a route via which hippocampal spatial view cells receive their information about and selectivity for locations in scenes. This direct route (Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022d) is complemented by connectivity via the posterior cingulate cortex (Rolls, Wirth, et al., 2022).

The “ventromedial visual stream” pathway is shown in more detail in Figure 6, which summarizes how in humans visual information reaches the parahippocampal gyrus PHA1–PHA3 regions via ventromedial (VMV1‐3) and ventral visual complex regions (Rolls, Deco, et al., 2022b). In this stream there is effective connectivity from V1 > V2 > V3 > V4. Then V2, V3 and V4 have effective connectivity to the VMV regions, which in turn have effective connectivity to PHA1‐3, which in turn have effective connectivity directed to the hippocampal system (Figure 6, green arrows) (Rolls, Deco, et al., 2022b). In addition, V2 has effective connectivity to the transitional visual areas dorsal transitional visual area (DVT) and the ProStriate (ProS) region, which in humans are where in the HCP‐MMP atlas (Glasser, Coalson, et al., 2016; Huang et al., 2022) the retrosplenial place area is located (Sulpizio et al., 2020); and these regions in turn have effective connectivity to the PHA parahippocampal regions (Figure 6) (Rolls, Deco, et al., 2022b). In humans, the occipital place area OPA is located in V3CD, V3B, and IP0 (Sulpizio et al., 2020).

It is proposed that scene representations are built using combinations of ventral visual stream features that when overlapping in space are locked together by associative learning and can form a continuous attractor network to encode a visual scene (Rolls et al., 2008; Rolls, Deco, et al., 2022c; Rolls & Stringer, 2005; Stringer et al., 2005) using spatial view cells (Georges‐François et al., 1999; Robertson et al., 1998; Rolls, 2022; Rolls et al., 1997b; Rolls et al., 1998; Rolls & Wirth, 2018; Tsitsiklis et al., 2020; Wirth et al., 2017) in the parahippocampal scene (or place) area referred to above, which in turn connects to the hippocampus to provide the “where” component of episodic memory (Rolls, Deco, et al., 2022c; Rolls, Deco, et al., 2022d). This connectivity to the hippocampal scene system is considered further elsewhere (Rolls, 2022; Rolls, Deco, et al., 2022c; Rolls, Deco, et al., 2022d). It is proposed below that a contribution of the dorsal visual stream to scene processing is to provide idiothetic update of spatial view representations.

At first sight, this ventromedial visual cortical stream for “where,” scene, representations, may seem like an unusual proposal. The proposal is that “where” information, about locations in scenes that are encoded by hippocampal spatial view cells, reaches the hippocampus from the PSA in PHA1‐3 and VMV1‐3, which has much connectivity with early ventral visual stream cortical areas. Indeed faces are represented near to the PSA in the fusiform gyrus FFC (Natu et al., 2021; Pitcher et al., 2019; Weiner et al., 2017); ideograms (or logograms) of words are represented just lateral to faces in the visual word form area in the fusiform gyrus (Caffarra et al., 2021; Dehaene et al., 2005; Dehaene & Cohen, 2011; Yeatman & White, 2021); and cortical regions that represent objects are nearby and project forward into the inferior temporal visual cortical areas involved in invariant visual object recognition (Grill‐Spector et al., 2006; Rolls, 2021a; Rolls, 2021f). However, scenes are likely to be represented by spatially contiguous scene features that become associated together in the correct topological arrangement because of the statistics of the inputs (Rolls et al., 2008; Stringer et al., 2005), and thus visual scene representations are likely to be formed from visual features of the type that are represented in ventral stream visual areas.

Highly relevant to these points is that spatial view cells are found in the macaque parahippocampal gyrus as well as the hippocampus (Georges‐François et al., 1999; Robertson et al., 1998; Rolls et al., 1997b; Rolls et al., 1998; Rolls et al., 2005; Rolls & Xiang, 2005). Indeed, it is proposed and likely that many of the neurons in the parahippocampal scene (place) area are spatial view cells, and that this is how scenes are represented in the primate including human brain.

What has just been proposed in Section 5.1 raises the question of the role of the parietal cortex, traditionally regarded as the brain region involved in “where” representations (Pitcher & Ungerleider, 2021; Ungerleider & Mishkin, 1982), in the responses of hippocampal (and parahippocampal gyrus) spatial view cells. As shown in Figures 5 and 7, the hippocampal system does receive input from parietal cortex visual regions (Baker, Burks, Briggs, Conner, Glenn, Taylor, et al., 2018; Huang et al., 2021; Ma et al., 2022; Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022c; Rolls, Deco, et al., 2022d). Inputs are also received via the visuomotor parts of the posterior cingulate cortex (Baker, Burks, Briggs, Conner, Glenn, Manohar, et al., 2018; Rolls, Wirth, et al., 2022).

In more detail, the effective connectivity of the human dorsal visual cortical stream is shown in Figure 7 (Rolls, Deco, et al., 2022b). There is effective connectivity from parietal cortex regions to the PSA, as illustrated in Figure 7. Strong effective connectivity is directed from inferior parietal region PGp to the PSAs in PHA1‐3 (Figure 7) (Rolls, Deco, et al., 2022c). PGp receives its inputs from parietal area 7 regions and intraparietal regions (Figure 7) (Rolls, Deco, et al., 2022b) involved in visual motion analysis and in coordinate transforms from retinal to head‐based and then to world‐based (allocentric) coordinates (Rolls, 2020; Salinas & Sejnowski, 2001; Snyder et al., 1998). These coordinate transforms are fundamental for self‐motion update of scene representations, so that the spatial view neurons in the PSA can represent where in a scene the individual is looking independently of eye position, head direction, and even the place of the head in the environment, when the view details are obscured or in the dark (Robertson et al., 1998; Rolls, 2020). Given these two lines of evidence, it is proposed that the parietal cortex has the role of idiothetic update of the scene representations in the PSA and thereby in the hippocampus (Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022c). Thus, the hypothesis is that the “where” scene representations in the human ventromedial visual stream are built by combinations of ventral stream spatial features, and the viewed position in the scene is idiothetically updated by coordinate transforms to the allocentric level of scenes (Rolls, 2020) by the parietal cortex inputs to the PSA (Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022c). Optic flow is a signal that can be used in idiothetic update, and in addition it is known that optic flow regions such as V3, V6 and the MT+ complex have functional (Rolls, Deco, et al., 2022b; Sherrill et al., 2015) and effective (Rolls, Deco, et al., 2022b) connectivity with regions involved in navigation such as the hippocampus, retrosplenial cortex, and posterior parietal cortex, and that optic flow activates regions that are close to cortical scene regions in the occipital, retrosplenial, and parahippocampal scene regions (Sulpizio et al., 2020).

It is emphasized in this approach that the “path integration” that is required for idiothetic update in humans and other primates involves eye position as well as head direction and the place where the individual is located, and much takes place in the dorsal visual system regions in the cortex in the intraparietal sulcus and area 7 (Rolls, 2020; Rolls, Deco, et al., 2022c). The mechanisms by which these coordinate transforms are implemented in the parietal cortex (Rolls, 2020) are considered in Section 6. A real problem with hypothesizing that path integration of any type occurs within the hippocampus is that the energy landscape of any continuous attractor network representation of place or spatial view in the hippocampus that utilized idiothetic update (Rolls & Stringer, 2005; Stringer et al., 2002; Stringer et al., 2005) would be so distorted by association with the “what” and reward information used for episodic memory that it would be very poor at path integration, as the energy landscape would be too bumpy because of the associations (cf. Spalla et al., 2021).

What I am proposing is that the primate including human hippocampal “where” system has two components or parts. One is a ventromedial visual cortical stream scene system with spatial view cells formed by visual feature components locked together in the correct spatial arrangement using overlapping receptive fields and associatively modified synapses of recurrent collaterals of the pyramidal cells that learn from the statistics of the neuronal activity how close the features are in the scene (Rolls et al., 2008; Stringer et al., 2005). This is located in ventral visual stream areas such as the ventromedial visual areas VMV1‐3 and the PSA (Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022d). The second component is the dorsal visual stream areas extending into the intraparietal sulcus areas such as LIP, VIP and MIP, area 7, and PGp (in the HCP‐MMP atlas (Glasser, Coalson, et al., 2016; Huang et al., 2022)), which are involved in idiothetic update of spatial view cells by computing representations of allocentric view that are invariant with respect to retinal position, eye position, head direction, and the place where the individual is located (Rolls, 2020; Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022c). The connectivity from the hippocampus and parahippocampal gyrus to the parietal cortex may further provide a route for allocentric locations in space with associations with rewards, goals, or objects and remembered using hippocampal mechanisms to produce navigation and visuomotor actions in space that require transforms to egocentric coordinates (Rolls, Deco, et al., 2022c).

To add to the recently developing understanding of human hippocampal system inputs, Figure 8 shows the ventrolateral visual stream in humans progressing via V1 > V2 > V4 > FFC (which contains representations of faces, objects, and even words in the visual word form area laterally) > the anterior temporal lobe TE regions (Rolls, Deco, et al., 2022b) where invariant representations of objects and faces are built (Rolls, 2000; Rolls, 2021a; Rolls, 2021f). This pathway provides “what” inputs to the hippocampal memory system via parahippocampal area TF, which is lateral and anterior to the scene area in PHA1‐3 (Figure 8) (Rolls, Deco, et al., 2022b).

Figure 8 also illustrates some of the reward‐related inputs to the human hippocampal system with green arrows, and these are considered next.

First, reward value and emotion‐related inputs reach the human hippocampal system from the orbitofrontal cortex and anterior cingulate cortex partly via the perirhinal and entorhinal cortex (Figures 5, 8, and 9) (Ma et al., 2022; Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022d; Rolls, Deco, et al., 2022e), and perhaps even more directly (Huang et al., 2021). This important route for reward to gain access to the hippocampus in addition to “what” and “where” inputs is shown in the updated hippocampal schematic connection diagram in Figure 10, which shows how reward value information could be associated with “where” spatial view cell activity in hippocampal CA3, if not before.

The connectivity for these reward value and emotion‐related inputs to reach the human hippocampus is shown in more detail in Figures 8 and 9 (Rolls, Deco, et al., 2022b; Rolls, Deco, et al., 2022e). The sensory/perceptual cortical regions on the left of Figure 9 provide visual, taste, olfactory, and auditory input not coded in terms of their reward value to the orbitofrontal cortex (Rolls, 2019a; Rolls, 2019b; Rolls, 2021a). The representation of these signals in the orbitofrontal cortex is in terms of their reward value, as shown by stimulus–reward learning and reversal; and by satiation which selectively reduces the reward value (Berlin et al., 2004; Grabenhorst & Rolls, 2011; Hornak et al., 1996; Hornak et al., 2003; Hornak et al., 2004; Rolls, 2019b; Rolls, 2021e; Rolls et al., 1994; Rolls, Cheng, & Feng, 2020; Rolls, Vatansever, et al., 2020; Thorpe et al., 1983). Reward is represented especially in the human medial orbitofrontal cortex, and punishment and non‐reward in the lateral orbitofrontal cortex (Grabenhorst & Rolls, 2011; Rolls, Cheng, & Feng, 2020). The medial and lateral orbitofrontal cortex have effective connectivity to the pregenual anterior cingulate cortex, in part via the ventromedial prefrontal cortex (vmPFC) (Figure 9) (Rolls, Deco, et al., 2022e). The medial and lateral orbitofrontal cortex, and the vmPFC, then have effective connectivity with the hippocampal system (perirhinal and entorhinal cortex, and perhaps directly with the hippocampus) (Figures 8 and 9) (Rolls, Deco, et al., 2022e). The memory‐related part of the posterior cingulate cortex provides additional connectivity between these reward‐related regions and the hippocampal system (Figure 5) (Rolls, Deco, et al., 2022d; Rolls, Wirth, et al., 2022).

These neural connectivity investigations provide clear evidence on how reward value and emotion‐related information can reach the human hippocampal system. It is proposed that in the hippocampus, including in CA3, reward value information can be associated with spatial view information to enable the reward and emotional aspects of episodic memory to become part of the episodic memory (Rolls, 2022; Rolls, Deco, et al., 2022e). The return pathways from the hippocampus to the orbitofrontal cortex, pregenual anterior cingulate cortex, and vmPFC shown in Figures 5, and 8, 9, 10 provide a route for the reward and emotional value of an episodic memory to be recalled back from the hippocampus to the orbitofrontal and anterior cingulate cortex and vmPFC (Rolls, 1995; Rolls, 2021a; Rolls & Treves, 1994; Treves & Rolls, 1994), from which it can influence other brain regions involved in actions (Rolls, 2021a; Rolls, Cheng, & Feng, 2020; Rolls, Deco, et al., 2022e).

This connectivity described in humans also shows how a key component of navigation, the goal or reward toward which the navigation is directed, reaches the human hippocampal system, where navigation may be guided by a remembered sequence of spatial view locations encoded by spatial view cells (Rolls, 2021a; Rolls, 2021b; Rolls, Deco, et al., 2022e).

In addition, the orbitofrontal cortex and pregenual anterior cingulate cortex have connectivity (yellow in Figure 6) directed to the human cholinergic basal nucleus of Meynert which projects to the neocortex, and septal region which projects to the hippocampus, and these pathways are proposed to influence memory consolidation including consolidation into long‐term semantic memory (Rolls, 2022; Rolls, Deco, et al., 2022e).