Distributed neural codes of the 3D position in the marmoset frontal cortex and hippocampus Authors: Chuanyao Felix Wei, Jiankang Xi, Jingyi Xie, Xiaojia Zhu, Haotian Yan, Xin Wang, Zhengtuo Zhao, Cirong Liu, Chun Xu Published: 2026 ABSTRACT The neural encoding of the animal’s position is essential for spatial navigation and cognition, yet the three-dimensional (3D) positioning system in the primate brain remains poorly understood. By combining wireless multi-channel recordings from common marmosets with a free-forging task in the 3D space, we identified 42% and 24% neurons in the frontal cortex (FC) and hippocampus (HPC) as place cells, respectively. Place cells in FC and HPC encoded the 3D position with similar spatial information, sparsity and stability, and exhibited stronger functional connections than non-place cells. Notably, their place fields were preferentially distributed in the center and border of the 3D space, and were recapitulated by an RNN-AutoEncoder model simulating the FC-HPC network. Collectively, these findings support a hypothesis that a widely distributed neural network involving frontal cortex and hippocampus constitutes a 3D positioning system in primates possibly by local recurrent circuits and long-range connections. Introduction The spatial cognition empowers animal survival by enabling animals to recognize and memorize one’s own position and navigate in a physical space. The cognitive map of the environment in the brain supports the spatial memory and navigation by various function cells including place, head direction, grid and border cells, which are extensively studied in the rodent hippocampal formation1–5 and have been recently characterized in the primate hippocampus6–8. Most studies on spatial cognition used rodents and focused on the 2D space. However, many species such as primates, live and locomote in a 3D volumetric world. It is a long-standing question whether a similar positioning system for 3D space exist in the primate brain, particularly the primate hippocampus. Pioneering studies have begun to identify place cells (PCs) encoding 3D positions in the hippocampus of bats9 and rats10. Although primate hippocampal cells have been found to encode the 1D/2D position6,7, the existence and functional relevance of PCs encoding 3D positions have been challenged by recent studies revealing a predominance of view cells over place cells7,8,11. In the 3D space, the encoding features of PCs in rats and bats may be modulated either similarly or differently by the gravity, locomotion patterns and experiences9,10,12–17. For instance, the geometry of PC’s place fields is elongated in rats but not in bats9,10. It remains elusive how the 3D positioning system operates in the primate hippocampus including the potential modulation by gravity and locomotion patterns such as climbing and walking. Accumulating evidence suggests that PCs exist outside the hippocampus such as the orbitofrontal cortex18, piriform cortex19 and the somatosensory cortex in rodents20. Furthermore, PCs were identified in frontal cortex including prefrontal and premotor areas in both macaques21 and humans22,23. Thus, the widely distributed navigational cells in the network of frontal cortex and hippocampus may underscore their critical roles in free-foraging behaviors and goal-directed behaviors23–28. Yet, it remains unclear how the 3D positioning system is organized in the primate brain, particularly in the frontal-hippocampal network29. In this study, we have setup a wireless multi-channel recording system in freely moving marmosets, and tested the hypothesis that the 3D space is widely represented in a distributed network of frontal areas and hippocampus in primates29. When marmosets engaged in a 3D-space free-foraging task, we recorded neuronal activities in the hippocampus and frontal areas including prefrontal and premotor areas. We identified PCs encoding 3D positions in both hippocampus and frontal cortex, which encompassed 42% and 24% neurons in these brain areas, respectively. PCs in these two areas shared similar properties such as spatial information, sparsity, and stability. Their place fields were rarely elongated along the gravity axis. Interestingly, their place fields were enriched at the center and border of the 3D space, which was recapitulated by an RNN-AutoEncoder model with frontal-hippocampal connections. Our results demonstrate a distributed 3D positioning system in the marmoset brain circuity as well as a strong neural representation of the 3D space in the frontal cortex. Results Single-unit recording from marmosets moving freely in a 3D space To investigate the 3D spatial cognition in primates, marmosets were subjected to a free foraging task in a 3D maze in which climbing frames were available on the wall and on the second floor (see Methods, Fig. 1A). We tracked the positions of marmoset’s body and head using deeplabcut and DANNCE (3-dimensional aligned neural network for computational ethology) algorithms30,31 (Fig. S1, systematic error estimated at 7 mm, ~5% of adult animal’s body size). We measured behavioral variables of marmoset’s locomotion including the speed, head direction (pitch and azimuth) and facing position (defined by the head direction projected onto the wall), as exemplified by a passage of 2.5-s locomotion (Fig. 1B). Marmosets were actively foraging over 30% of the time of each session (20-40 min) with travel distance over 200 m (Fig. S2A and S2B), which allowed for sufficient coverage of the 3D space in the maze (Fig. S2C–S2F).Fig. 1Single-unit recording from marmosets moving freely in a 3D space.A Scheme illustrating the experimental arena setting. Marmosets were engaged in free forging behaviors in a 3D space where walls were covered by climbing frames. The second floor (1.2 m × 1.2 m) in the middle was ~0.5 m high. Animal’s behavioral trajectories were tracked by four surrounding cameras. B Examples of behavioral tracking analysis during a 2.5-s epoch: the 3D trajectory of animal’s body position (top left); instantaneous speed (top right), horizontal head direction (azimuth, bottom left), and pitch head direction (bottom right). C Scheme illustrating a 64-channel wireless recording system using Blackrock equipment (Method 1). The movable 32-ch tetrodes (right inset) were inserted into FC and HPC and connected with a CerePlex Exilis for wireless data transmission. The whole implant was protected by a titanium chamber fixed on the skull and a protective plastic helmet that could be covered by fluorescent sponges (left inset). D Scheme illustrating a 256-channel wireless recording system using Spikegadgets equipment (Method 2). The 128-ch Ultraflexible electrodes (inset) were inserted into FC and HPC and connected with a HH256 headstage, a microSD for data storage and a backpack battery on the animal’s back. The whole implant was protected in a similar way as in (A). E Left, example waveforms of three colored single units identified in a single tetrode. Right, spikes from individual units were sorted into clusters in Peak-Valley feature spaces. Unit isolation was evaluated by calculation of J3 and Davies Bouldin validity index (DB). F Summary of the sorting quality evaluated by the statistics of J3 and DB for tetrodes with and without single units (orange, tetrodes with units, n = 62; gray, tetrodes without units, n = 21). High J3 and low DB values indicate good isolation. Single-unit recording from marmosets moving freely in a 3D space. A Scheme illustrating the experimental arena setting. Marmosets were engaged in free forging behaviors in a 3D space where walls were covered by climbing frames. The second floor (1.2 m × 1.2 m) in the middle was ~0.5 m high. Animal’s behavioral trajectories were tracked by four surrounding cameras. B Examples of behavioral tracking analysis during a 2.5-s epoch: the 3D trajectory of animal’s body position (top left); instantaneous speed (top right), horizontal head direction (azimuth, bottom left), and pitch head direction (bottom right). C Scheme illustrating a 64-channel wireless recording system using Blackrock equipment (Method 1). The movable 32-ch tetrodes (right inset) were inserted into FC and HPC and connected with a CerePlex Exilis for wireless data transmission. The whole implant was protected by a titanium chamber fixed on the skull and a protective plastic helmet that could be covered by fluorescent sponges (left inset). D Scheme illustrating a 256-channel wireless recording system using Spikegadgets equipment (Method 2). The 128-ch Ultraflexible electrodes (inset) were inserted into FC and HPC and connected with a HH256 headstage, a microSD for data storage and a backpack battery on the animal’s back. The whole implant was protected in a similar way as in (A). E Left, example waveforms of three colored single units identified in a single tetrode. Right, spikes from individual units were sorted into clusters in Peak-Valley feature spaces. Unit isolation was evaluated by calculation of J3 and Davies Bouldin validity index (DB). F Summary of the sorting quality evaluated by the statistics of J3 and DB for tetrodes with and without single units (orange, tetrodes with units, n = 62; gray, tetrodes without units, n = 21). High J3 and low DB values indicate good isolation. We chronically implanted tetrodes or ultraflexible electrodes into the hippocampus (HPC) and frontal cortex (FC) of four marmosets by the MRI-guided localization of brain areas (see Methods, Fig. S3), and then performed wireless single-unit recordings from freely moving marmosets using two data acquisition methods. The first was based on a Blackrock CerePlex Exilis connected with tetrodes and a Cereplex 64-channel data acquisition system (Fig. 1C), and the other was based on a SpikeGadgets HH256 connected with ultraflexible electrodes32 and a backpacked data logger (see Methods, Fig. 1D). The whole implant was protected by a titanium base and a protective plastic helmet with fluorescent sponges, which could serve as head direction markers. Marmosets exhibited very similar locomotion capabilities when using these two data acquisition methods (Fig. S4). In summary, we have recorded 43 sessions from three marmosets during the spatial foraging task in the 3D maze, and collected 919 well-isolated single units recorded from both tetrodes and ultraflexible electrodes (Fig. 1E and 1F, Fig. S5A–S5C). Based on their waveform features such as the peak-to-valley latency (Fig. S5D and S5E), we identified 248 and 213 putative pyramidal cells from FC and HPC, respectively, and included them in the following data analysis (Fig. 1E and 1F, Fig. S5). Identifications of PCs encoding 3D positions in FC and HPC As exemplified in Fig. 2A, B (see more examples in Fig. S7), we identified spatially modulated PCs by quantifying the significance of each cell’s spatial information for 3D-position encoding33. Interestingly, the PCs in FC (104/248, 42%) were about twice as many as those in HPC (51/213, 24%). The PCs in FC and HPC exhibited similar spatial information and sparsity in the 3D space (Fig. 2C, D). These features were encoded similarly along each dimension of the 3D space, but the spatial information of HPC neurons were slightly higher in the Z axis (gravity axis) than in X and Y axis (Fig. S8). Furthermore, PCs in both FC and HPC showed similar stability across time, as assessed by the correlation analysis between the first and second halves of one session (Fig. 2E and 2F, see examples in Fig. S7). Among PCs, only 6.7% and 3.9% were modulated by the theta-band wave in FC and HPC, respectively. The theta index for FC cells was higher than that for HPC cells (Fig. 2G), among which the CA2/3 PCs showed higher theta index than CA1 PCs (Fig. S9D). In addition to single-cell level of space representation, both the FC and HPC populations of each session can also decoded the 3D position above the chance level (Fig. S10).Fig. 2Identifications of PCs in marmoset FC and HPC.A, B Example recording of a 3D PC in FC (A) and HPC (B). Left, 3D trajectories with red dots representing spikes. Cell ID was shown at above. Middle, heatmap showing firing rates (normalized by the max-min value). Max value was shown above. Scale bar was shown below. Right, place field with a dot representing the centroid. C, D Box plot of the spatial information (C) and sparsity index (D) of PCs (green dots) and non-PCs (gray dots) in FC (left darker shading) and HPC (right lighter shading). Single-unit numbers were shown below. E Example of firing rate maps during the whole recording session, the first half and the second half session (normalized by the max value shown above). The stability index for the HPC cell (upper panels) was 0.71, and that for the FC cell (bottom panels) was 0.39. F Box plot of the stability index of PCs and non-PCs in FC (left darker shading) and HPC (right lighter shading). Each dot represents one cell, n = 104 FC PCs, 144 FC non-PCs, 51 HPC PCs and 162 HPC non-PCs. G Box plot of the theta index of PCs and non-PCs in FC (left darker shading) and HPC (right lighter shading). Each dot represents one cell, n = 104 FC PCs, 144 FC non-PCs, 51 HPC PCs and 162 HPC non-PCs. (C, D, F and G) The box plot shows values at min, 25%, 50%, 75% and max. Wilcoxon rank sum two-sided tests: the p-values in order are 0.30, 0.0018, 2.5 × 10−5, 0.56, 0.012, 1.9 × 10−4, 0.095, 3.9 × 10−14, 1.5 × 10−14, 5.1 × 10−5, 0.36 and 0.20 respectively. Identifications of PCs in marmoset FC and HPC. A, B Example recording of a 3D PC in FC (A) and HPC (B). Left, 3D trajectories with red dots representing spikes. Cell ID was shown at above. Middle, heatmap showing firing rates (normalized by the max-min value). Max value was shown above. Scale bar was shown below. Right, place field with a dot representing the centroid. C, D Box plot of the spatial information (C) and sparsity index (D) of PCs (green dots) and non-PCs (gray dots) in FC (left darker shading) and HPC (right lighter shading). Single-unit numbers were shown below. E Example of firing rate maps during the whole recording session, the first half and the second half session (normalized by the max value shown above). The stability index for the HPC cell (upper panels) was 0.71, and that for the FC cell (bottom panels) was 0.39. F Box plot of the stability index of PCs and non-PCs in FC (left darker shading) and HPC (right lighter shading). Each dot represents one cell, n = 104 FC PCs, 144 FC non-PCs, 51 HPC PCs and 162 HPC non-PCs. G Box plot of the theta index of PCs and non-PCs in FC (left darker shading) and HPC (right lighter shading). Each dot represents one cell, n = 104 FC PCs, 144 FC non-PCs, 51 HPC PCs and 162 HPC non-PCs. (C, D, F and G) The box plot shows values at min, 25%, 50%, 75% and max. Wilcoxon rank sum two-sided tests: the p-values in order are 0.30, 0.0018, 2.5 × 10−5, 0.56, 0.012, 1.9 × 10−4, 0.095, 3.9 × 10−14, 1.5 × 10−14, 5.1 × 10−5, 0.36 and 0.20 respectively. We further measured more features of spatial cognition using data acquired from sessions in which head directions were tracked (184 FC cells and 146 HPC cells). We found that similar numbers of FC and HPC cells were modulated by the speed and head directions (FC vs HPC: 51% vs 51% speed, 38% vs 30% azimuth, 14% vs 12% pitch; Fig. S11A–11F). Consistent with previous studies8, we found more cells encoding facing positions (50% in FC and 36% in HPC) than those encoding 3D positions (Fig. S11G and S11H). In summary, many FC and HPC cells showed conjunctive encoding for spatial variables including 3D position, facing position, speed, and head pitch and azimuth (Fig. S11I and S11J). Interestingly, we found that PCs showed highly correlated position encoding even when animals were in different locomotion patterns (e.g., walking on and hanging beneath the mezzanine), suggesting a stable spatial tuning independent of the locomotion patterns (Fig. S12). Taken together, these results suggest that PCs in FC and HPC represent the 3D position similarly at the single-cell level but differ at the size of neuronal population. Similar place field properties between PCs in FC and HPC To compare place field properties of PCs in FC and HPC, we measured the number and volume of place fields and their spatial distributions in the 3D space (see Methods). We found that the PCs in FC and HPC showed similar numbers and volume sizes of place fields–most PCs exhibited a single place field with small volume size (Fig. 3A–C). Of note, the place field size exhibited an exponential distribution – most were small (fine-tuned) and a few were large (~20% of the entire 3D space) (Fig. 3D). These results suggest that PCs in FC and HPC support both the exactitude and vastitude of spatial encoding in the 3D space.Fig. 3Characterizations of place fields of PCs.A Box plot of the number of place fields in each 3D PC in FC (blue dots, n = 104) and HPC (yellow dots, n = 51). B Box plot of the volume of each place field in PCs in FC (blue dots, n = 164) and HPC (yellow dots, n = 90). C Box plot of the total volume of all place fields in each 3D PC in FC (blue dots, n = 104) and HPC (yellow dots, n = 51). A–C The box plot shows values at min, 25%, 50%, 75% and max. Wilcoxon rank sum two-sided test: p = 0.44, 0.50 and 0.45 in order. D Histogram of the volume of place fields of PCs in FC (left, blue bars) and HPC (right, yellow bars). Gray lines were the exponential fit. Inset showed the scatter plot of the same data with log-scaled Y axis. Straight lines were the linear regression (p = 5.2 × 10−8 and 0.0016). E, G Examples of colored place fields of some PCs in FC (E) and HPC (G). F, H Upper panels, the width of each place field (gray line) of PCs in FC (F) and HPC (G) projected onto X, Y and Z axis, respectively. Place fields in each panel were sorted by the centroid location (colored dot) projected onto the axis, respectively. Lower panels, the density plot of the centroid location in the X, Y and Z axis, respectively. Kolmogorov-Smirnov test: p = 2.7 × 10−4, 7.9 × 10−9, 2.5 × 10−13, 0.027, 1.5 × 10−8 and 8.5 × 10−5 in order. I, J The density plot showing absolute values of the place field’s azimuth along the horizontal plane (left panel) and its pitch along the plumbline (right panel) in FC (I) and HPC (J). Kolmogorov-Smirnov test: p = 0.53, 6.8 × 10−113, 0.15 and 1.5 × 10−62 in order. Characterizations of place fields of PCs. A Box plot of the number of place fields in each 3D PC in FC (blue dots, n = 104) and HPC (yellow dots, n = 51). B Box plot of the volume of each place field in PCs in FC (blue dots, n = 164) and HPC (yellow dots, n = 90). C Box plot of the total volume of all place fields in each 3D PC in FC (blue dots, n = 104) and HPC (yellow dots, n = 51). A–C The box plot shows values at min, 25%, 50%, 75% and max. Wilcoxon rank sum two-sided test: p = 0.44, 0.50 and 0.45 in order. D Histogram of the volume of place fields of PCs in FC (left, blue bars) and HPC (right, yellow bars). Gray lines were the exponential fit. Inset showed the scatter plot of the same data with log-scaled Y axis. Straight lines were the linear regression (p = 5.2 × 10−8 and 0.0016). E, G Examples of colored place fields of some PCs in FC (E) and HPC (G). F, H Upper panels, the width of each place field (gray line) of PCs in FC (F) and HPC (G) projected onto X, Y and Z axis, respectively. Place fields in each panel were sorted by the centroid location (colored dot) projected onto the axis, respectively. Lower panels, the density plot of the centroid location in the X, Y and Z axis, respectively. Kolmogorov-Smirnov test: p = 2.7 × 10−4, 7.9 × 10−9, 2.5 × 10−13, 0.027, 1.5 × 10−8 and 8.5 × 10−5 in order. I, J The density plot showing absolute values of the place field’s azimuth along the horizontal plane (left panel) and its pitch along the plumbline (right panel) in FC (I) and HPC (J). Kolmogorov-Smirnov test: p = 0.53, 6.8 × 10−113, 0.15 and 1.5 × 10−62 in order. To further delineate the spatial distribution of place fields, we projected the place field onto the X, Y and Z axes of the 3D space, respectively, and found that place fields of PCs in FC and HPC showed nonuniform distributions along all three axs (Fig. 3E–H). Place fields in X and Y axes were enriched in the center, whereas those in the Z axis (gravity axis) were enriched in the ground floor and the ceiling. Such uneven distribution of place fields may have to do with the uneven distribution of head directions or time spent in the 3D space during the free foraging task, but their distributions were not well aligned with the place field distribution on X, Y and Z axis (Fig. S2D and S2E). Thus, the uneven representation of 3D space may be a combination of intrinsic properties and place fields and animal’s behavioral experiences. Finally, we characterized the orientation and the geometry of each place field by determining its principal axis and elongation and sphericity indices (Fig. S13A). We found that orientations of place fields in both FC and HPC were biased towards 0° in pitch direction, but were evenly distributed in the azimuth direction (Fig. 3I, J). This was reminiscent of the distribution of animal’s head direction. Place fields of FC and HPC neurons have a similar level of elongation and sphericity (Fig. S13B and S13C). Bootstrap analysis of elongation indices showed that a small number of place fields (about 15%) of FC and HPC neurons were significantly elongated with similar indices and horizontal orientation (Fig. S13D–H), and were primarily located close to the floor and ceiling (Fig. S13I–K). Thus, the 3D space is primarily represented by non-elongated (spheric-like) place fields with a supplementary contribution from elongated place fields that are biased to the horizontal direction. Together, these results suggest that PCs in FC and HPC share similar place field properties with sphere and small-size geometry and uneven spatial distribution. Enhanced functional connections among PCs in FC-HPC network The FC and HPC neurons and their interactions are essential for spatial cognition and memory23–28. We thus examined the functional connections between FC and HPC neurons, as well as within FC neurons and within HPC neurons, during the session in which FC and HPC neurons were simultaneously recorded (N = 11 sessions). As illustrated in Fig. 4A (see Methods), the functional connection was identified by a dynamical differential covariance method34, and was exemplified by two pairs of functionally connected neurons with either direction (Fig. 4B). The functional connection was quantified by the network density as depicted in Fig. 4C (density values of 1 and 0 represent the highest and lowest connection rates). We found that the density of FC-FC and HPC-FC functional connections were similar and were higher than that of HPC-HPC functional connections (Fig. 4D). Interestingly, the functional connections among PCs were significantly higher than those between PCs and non-PCs, as well as among non-PCs (Fig. 4E, F).Fig. 4Functional connections among neurons in the FC-HPC network.A Simulated firing rates of two cells with functional connection from X to Y: the activity of cell X (x) is correlated with the first-order derivative (y’) of cell Y’s activity, which can be understood as y’=ax+b, where a is the correlation coefficient and b represents noise. B Example recording epochs for two pairs of functionally connected cells (from green to purple), respectively: left panel, a FC neuron projected to a HPC neuron; right panel, a HPC neuron projected to a FC neuron. C Schematic illustration of network density, which is described as existed connection numbers divided by the maximal potential connection numbers (max = 16 for n = 4 nodes). Left example, density is 2/16. Right example, density is 13/16. D Box plot of the density of functional connections between FC neurons (intra-group, n = 10), between HPC neurons (intra-group, n = 6), and between FC and HPC neurons (inter-group, n = 10). E Box plot of the density of functional connections between PCs (intra-group, n = 10), between non-PCs (intra-group, n = 11), and between PCs and non-PCs (inter-group, n = 20). D, E The box plot shows values at min, 25%, 50%, 75% and max. Wilcoxon rank sum two-sided test: p = 0.025, 0.57, 0.039, 0.024, 0.037 and 0.27 in order. F Schematic summary of the functional connection (lines) among PCs and non-PCs in FC and HPC. Functional connections among neurons in the FC-HPC network. A Simulated firing rates of two cells with functional connection from X to Y: the activity of cell X (x) is correlated with the first-order derivative (y’) of cell Y’s activity, which can be understood as y’=ax+b, where a is the correlation coefficient and b represents noise. B Example recording epochs for two pairs of functionally connected cells (from green to purple), respectively: left panel, a FC neuron projected to a HPC neuron; right panel, a HPC neuron projected to a FC neuron. C Schematic illustration of network density, which is described as existed connection numbers divided by the maximal potential connection numbers (max = 16 for n = 4 nodes). Left example, density is 2/16. Right example, density is 13/16. D Box plot of the density of functional connections between FC neurons (intra-group, n = 10), between HPC neurons (intra-group, n = 6), and between FC and HPC neurons (inter-group, n = 10). E Box plot of the density of functional connections between PCs (intra-group, n = 10), between non-PCs (intra-group, n = 11), and between PCs and non-PCs (inter-group, n = 20). D, E The box plot shows values at min, 25%, 50%, 75% and max. Wilcoxon rank sum two-sided test: p = 0.025, 0.57, 0.039, 0.024, 0.037 and 0.27 in order. F Schematic summary of the functional connection (lines) among PCs and non-PCs in FC and HPC. To understand whether the functional connectivity between paired PCs correlate with the spatial overlapping between place fields, we measured the overlapping probability and ratio between two place fields as well as reach frequency that animals traversed from one field to another. Analysis revealed similar results for paired and non-paired PCs (Fig. S14A–F), arguing against the correlation between the functional connectivity and spatial overlap between place fields. Next, we inquired whether the functional connectivity correlated with the conjunctive encoding capacity. Analysis showed that functional cells with higher connection rates processed more types of encoding for position, speed, and head direction (Fig. S14G). Taken together, these results suggest that functional connections among PCs are crucial for the 3D position encoding as well as other conjunctive encodings. Recapitulated PCs in an RNN-AutoEncoder network Next, we sought to recapitulate PCs using a simulated neural network system with various initial settings, in which a recurrent neural network (RNN) model and an AutoEncoder model simulate FC and HPC neurons, respectively (see Methods, Fig. S15A–C). We found that many nodes of model-FC and model-HPC exhibited spatially modulated activity and were identified as PCs (Fig. S15D and S15E, see more examples in Fig. S16). These simulated PCs in FC and HPC showed very small, but significant differences in spatial information, sparsity, and stability (Fig. S17A). The size of place fields in FC and HPC neurons were similar (Fig. S17B) and followed an exponential distribution (Fig. S17C). We found that place fields of simulated PCs and their preferred head directions showed uneven distributions in the 3D space (Fig. S15F-K). These results demonstrated that the RNN-AutoEncoder model recapitulated the PCs in FC and HPC with similar place field properties as experimentally measured. Finally, we took advantage of the full control of animal’s locomotion in modeling and provided completely even distribution of animal’s positions in the 3D space to make sure that behavioral experiences exert homogenous impact on the neural representation of the 3D space (Fig. S15L). Interestingly, the uneven representation of 3D space still emerged in the RNN-AutoEncoder model (Fig. S15M and S15N), implying an intrinsic uneven neutral representation of the spatial environment in the neural network of FC and HPC neurons. Discussion By combining wireless recording, 3D-position tracking, and quantitative behavioral analysis of freely moving marmosets, we have recorded single-unit activity in frontal cortex and hippocampus during a 3D-space free forging task. Our study revealed the existence and similar properties of PCs widely distributed in the neural network of frontal cortex and hippocampus. These PC’s place fields exhibited some interesting features: (1) a spatial distribution that was scattered yet enriched in the center and border; (2) a log-scaling distribution of the place field size; (3) a generally sphere-like geometry that was independent of the locomotion patterns. An RNN-AutoEncoder model could recapitulate these features and suggests a possible contribution from recurrent neural networks, in line with experimentally measured functional connections in FC and HPC. Our study supports the hypothesis that the allocentric encoding of 3D positions is widely represented in a distributed network of frontal cortex and hippocampus in primates. Distributed 3D position encoding in the primate brain We have identified widely distributed PCs in frontal cortex by recording from multiple sites in medial and dorsal lateral prefrontal areas (A8, A9, A10) and the premotor area (A6). Similar encoding of allocentric 3D position has also been found in prefrontal and premotor areas of macaques21. Such notion is also consistent with prior studies on 2D position encoding: (1) Rodent studies showed allocentric position encoding in piriform cortex19 and the somatosensory cortex20. It remains to be examined in future whether such allocentric position encoding exist in other brain areas outside of hippocampal formation and frontal cortex in primates. (2) Human studies identified place cells in the frontal cortex and hippocampus as well as parahippocampal region and amygdala22,23. Interestingly, the stronger theta modulation in CA2/3 neurons than in CA1 neurons indicates a regional difference within the hippocampus impinging on a distributed system for 3D position encoding in marmosets. A 3D positioning system centered on the frontal cortex The notion of distributed encoding of allocentric positions in marmoset brains is well supported by our data. Moreover, all the properties of PCs encoding 2D and 3D positions in frontal cortex are comparable to those in the hippocampus, suggesting that frontal cortical neurons exhibited similar characteristics of spatial position encoding as hippocampal neurons in the marmoset. Our finding that the proportion of PCs in frontal cortex is about twice as that in hippocampus suggest a larger contribution to the 3D positioning system by frontal cortex. This is consistent with a recent human study that PCs in frontal cortex were about twice as those in hippocampus22. In contrast, other function cells such as speed, direction, and facing position-modulated cells were comparable in frontal cortex and hippocampus. We speculate that the center of classical 3D positioning system shifts from hippocampus to frontal cortex in the marmoset. The primate hippocampus, together with frontal cortex, may also represent the 3D space by encoding facing positions during visually guided navigation7,8,11, compensating for the generally lower spatial information of PCs in primates and birds35 than in rodents. The elongated geometry of place fields in rats but not in primates and bats indicate the species-specific behavioral repertoire in the 3D space. Our study is limited by the inability to separate the facing position and spatial view without eye tracking. Nevertheless, previous studies showed that eyes and head of marmosets were well aligned and marmosets’ eye-in-head saccades (<5°) were much smaller than those of macaques8,36, and facing position encoding in macaques was much more powerful than spatial view encoding7. In summary, our study suggests that the primate brain maintains a classical self-positioning system in a distributed neuronal network of hippocampus and frontal cortex with a focus shifted to the latter. Self-positioning system in primates We noted that the PCs in both frontal cortex and hippocampus are mostly mixed with conjunctive coding for facing position and spatial view, which is allocentric positioning for facing (visual) locations7,8,21. These two types of allocentric positioning could represent two completely different functional needs at the same time. For instance, when an animal walks through or stands at one location, its brain could encode both its own position and the location it is facing/looking at in the environment at the same time. The rodent PCs might be also mixed with conjunctive coding to visual positions, because rodent PCs could sample animal’s surrounding locations beyond its traveled path37. As the nonlinear mixed selectivity of conjunctive coding is an advantageous coding scheme38,39, we argue that the self-positioning system is still notable and critical for primates based on following evidences: (1) A prominent number (this study, 24%; Piza et al. 32%) of hippocampal neurons were able to encode 3D positions8. Similar number of 1D/2D PCs have been identified in various primate species including macaques, squirrel monkeys, and marmosets6,7,40–42. (2) Among the PCs in our study, 27 % (7/26) in HPC and 61 % (11/18) in FC were not modulated by the facing location (Fig. S11), suggesting that neurons for self-position system and visual position system are not entirely overlapped. (3) When the facing direction and location were very limited for animals running in the linear track, there were still notable number of PCs identified in the marmoset hippocampus6. In line with this, when the facing location was fixed (eye viewing) for macaques, grid cells in the entorhinal cortex could contribute to the position system for the attention location43,44. (4) The distributed neural encoding in the primate brain, particularly in the prefrontal cortex of marmosets (this study) and macaques21, may benefit from the evolutionarily expanded and diversified prefrontal cortex and hence contribute more to the self-positioning system than that in hippocampus. Therefore, the primate hippocampus and frontal cortex may enable an optimized and comprehensive 3D positioning system by combining classical self-positioning system and visual positioning system, which may complement each other in spatial navigation by encoding both self and visual positions. It could well be that multiple positioning systems for self, facing and attention are implemented widely in the primate cortex and hippocampus, in which neurons with mixed encoding play critical functional roles. Biased distribution of place fields and spatial navigation in the 3D space The place fields in the 3D space were enriched in the positions close to the floor and ceiling, which was recapitulated in our modeling with uniform behavioral trajectories. This uneven distribution of place fields suggests critical roles of the floor and ceiling in shaping the cognitive map for the 3D space. In the free foraging task, animals can walk on the floor and climb along the wall to reach the ceiling, whereas animals sometime could also reach the ceiling by directly jumping from the floor, which would be a shortcut in the cognitive map underscored by the place cells and biased distributions of their place fields in the 3D space. It remains to be studied in the future how the marmoset brain uses these 3D position encoding to identify its self-location and accomplish the goal-directed spatial navigation. The emergence of PCs encoding 3D positions in the prefrontal and premotor cortex (A6, A8, A9 and A10) may stem from their local recurrent connections and brain-wide long-range connections, as these brain areas are generally principal clusters with strong functional connectivity in the default mode network45,46. Although the functional interaction within PCs was stronger than that within non-PCs, it remains to be examined in the future how the long-range connectivity between hippocampus and frontal areas help to generate PCs and support the self-positioning system for primates. Taken together, our study has delineated a distributed encoding scheme in the network of frontal cortex and hippocampus that serves as a critical 3D positioning system in the primate brain, in which the prefrontal cortex is emerging as a key node. Methods Animals Four marmosets were individually housed (55 ~ 65% humidity, 26 ~ 28 °C) under a 12 h light/dark cycle with food and water ad libitum in the animal facility of Institute of Neuroscience (CAS Center for Excellence in Brain Science and Intelligence Technology). All experimental procedures were approved by the Institutional Animal Care and Use Committee (IACUC, CEBSIT-2021038) of the Institute of Neuroscience. Two male and two female adult common marmosets were used (weight 320 ~ 420 g, age 2 ~ 3 years) for behavioral task, electrodes implantation, and electrophysiology recording. Surgeries Reference points The animal was anesthetized with 1 ~ 2% isoflurane (RWD Life Science, Shenzhen). An incision (~4 cm) was made on the scalp to expose the skull, including the coronal suture, sagittal suture and anterior fontanel (Bregma). Using the Bregma as the center of the circle (reference point O), a curve with a circular gauge spreading 4 mm was drew and its intersection with the sagittal suture was defined as reference point P. Two drops of sterile gel (Yifu Medical Devices Inc.) were applied to point O and P (1 mm in diameter) by a syringe, respectively. These gels remained stable in the form of spherical or hemispherical and served as markers during the MRI scanning afterwards. MRI-guided location of brain areas Animals wearing an anesthetic breathing mask underwent a 9.4 T NMR scanning (Bruker BioSpec 94/30) and were allowed for at least two weeks of recovery before electrode implantation. The scanning images were reconstructed into 3D images using software Mango (Version 4.1, UTHSC) and then rotated such that the line of two ear canals and the line of point O and P were both leveled. The posterior hippocampus and frontal cortex were identified in the MRI images and their coordinates were calculated in relation to point O and P, and the brain surface. These coordinates were used to locate brain areas during electrode implantations with point O and P identified in the same way. Electrode implantation Animals were fasted for 12 h and free from drinking for 2 h prior to the electrode implantation. Animals were injected with atropine (0.16 mg, Shanghai Pharma Inc.), and then anesthetized by 1 ~ 2% isoflurane. The scalp was removed to expose the point O and P and the ear-canals and O-P lines were leveled again. A customized titanium chamber to house the electrodes and headstage was printed at Xiamen Everything 3D Printing Inc, and was fixed onto the skull by 4-6 titanium 1.2D screws (RWD Life Science Inc.) and dental cements. The bottom part of the titanium chamber (ca. 37 mm × 27 mm × 12 mm) was oval and customized to the size and geometry of individual marmoset skulls to ensure a tight fit. Based on the measured coordinates of target brain areas, windows (~2 mm × 2 mm) for electrode implantation were marked on the skull surface. After opening the skull, the dura was carefully cauterized with an electrocoagulator or cut out by ophthalmic shears to prevent bleeding. The electrodes were inserted into the target brain areas slowly and constantly. Once the electrodes were in place, a sterile gel was applied to the brain surface for isolation. The electrodes and chambers were fixed by dental cements. Animals were provided fluid foods after the surgery and then recovered for ~6 weeks. Behaviors Setup and free foraging behavior The 3D maze was composed by scaffolds of aluminum alloy (European standard 3030L-2.2) and walls with wire mesh for climbing. The top and bottom of the maze were protected by two 150 cm × 150 cm acrylic classes. Animals were habituated to the experiment room and recording procedures before formal electrophysiological recording. The daily habituation time gradually extended from 10 to 60 min in a period of 3 weeks. On the day of free foraging task with electrophysiological recording, animals were encouraged to forage actively by a portion of the food, or sometimes a glove with urine and feces. The recording session was at least 25 min long. The animals were returned to the feeding room together with remaining food. The maze was cleaned with rags soaked in 70% ethanol after the experiment of each animal. The traveling (active) time was scored if the animal was moving (>5 cm/s) for more than 2 seconds. The active rate was defined as the ratio of traveling time to session time. Behavioral tracking Behavioral videos were captured at 25 Hz by 4 cameras around the maze. We used the 3D version of DeepLabCut30 to generate training datasets for the animal’s trajectories, and triangulate and reconstruct the animal’s 3D position by DANNCE31. The cameras were rectified for the external parameters according to the rectifying method in DANNCE: (1) Filming a 12 × 5 checkerboard (each square 12 mm × 12 mm) for 20 min for each camera; (2) Adjusting synchronized and intrinsic-rectified cameras such that each corner in the maze can be captured by at least 3 cameras. (3) Rectifying the extrinsic parameters by filming four ‘L’ shape markers and labeling the four makers orderly in each camera frame. (4) Finalizing the extrinsic parameters for the camera by calculating the relative position between the camera and the ‘L’ shape markers. The camera positions and maze were kept constant throughout experiments. The 3D positions of animal body and head markers were triangulated from the 2D video frames captured during the experiment. To track the animal’s body position and sponge markers, we labeled 300–600 frames in the video of each camera and trained the DeepLabCut model with default parameters using a total of 10400 representative epochs from all sessions. Velocity measurement was based on body positions. Head direction measurement was based on the facing direction of a rigged triangle of three sponge markers. The azimuth and pitch of head directions were computed by the antitrigonometric function (MatLab atan2) accordingly. Electrophysiological recording Wireless 64-channel recording Two 32-channel tetrodes (Kedou Brain Computer Technology Inc.) were implanted into the brain above FC at depth of 2 mm and above posterior HPC at depth of 11 mm, respectively. After the recovery period as described above, the tetrodes were gradually moved to reach FC and posterior HPC. The tetrodes were moved ca. 50 μm deeper after each recording session. A custom helmet (ca. 37 mm × 27 mm × 22 mm, ~4 g) for recording sessions was attached to the chamber by screws to construct a closed safeguard of the electrodes and headstage. The helmet was replaced regularly to avoid abrasion. During the recording session, the transmitter (CerePlex Exilis, Balckrock Neurotech Inc.) was connected to the two tetrodes and wirelessly transmitted the raw data to a Cereplex 64-channel data acquisition system via eight antennas positioned around the 3D maze. During the recording in a 2D arena, four antennas were used and positioned at the four corners of the arena. Antennas orientations and working frequency were adjusted and fine-tuned for the 3D maze and the 2D arena, respectively, and were maintained throughout the entire experimental period. The tetrodes were adjusted by the driving screw 0.25 turn (50 µm) at the end of experiment in each recording session, or sometimes 0.5-1 turn (100-200 µm) if number of single units was small. Wireless 256-channel recording Two 128-channel untraflexible electrodes with a linear array of recording sites32 were inserted into FC (depth 6 mm) and HPC (depth 14 mm) with the guidance of a tungsten wire, respectively. The electrodes were connected to a headstage (HH256, SpikeGadgets) by a custom designed cable (Kedou, China). The synchronizing antenna was placed in the same room to wirelessly control the HH256 for initiating and terminating recording. The raw data was streamed to a microSD chip which was protected in the helmet and connected with a battery (~18 g) in the backpack. At the end of the recording, animal brains were perfused, scanned by MRI, and sectioned in a vibratome. The electrodes tracks were examined and confirmed to be in target brain areas. Spike sorting The raw data from the Cereplex 64-channel data acquisition system (ns6 format) and SpikeGadgets (bin format) were converted to plx format by Off-Line Spike Sorter (OFSS, Plexon) for storage and analysis. The spike sorting was done in OFSS with following steps: (1) Apply a 4-pole 250-Hz low-cut Bessel filter. (2) Manually set a threshold (usually >36 μV) for each channel to detect waveforms. (3) Sort with tetrodes based on waveform features, including Peak-Valley, Energy, and principal components. (4) Define individual single units as a group of waveforms which exhibited a distinct cloud in the feature space. (5) Evaluate the sorting quality by J3 (the ratio of between-cluster to within-cluster scatter) and the Davies–Bouldin validity index (DB, the ratio of the sum of within-cluster scatter to between-cluster separation). Good sorting would show a large J3 and small DB. (6) Invalidate a single unit if it contains fewer than 100 spikes in a session, or if its amplitude is lower than 36 μV, or if its recording variability index is higher than 0.35. The recording variability index was calculated as (a-b)/(a + b), where a stands for the mean firing rate of this unit in the 1st half session, and b stands for mean firing rate in the 2nd half session. High recording variability index indicates bad stability. (7) Remove duplicated units. If two unites share more than 50% of spike times with each other when estimated at 1 ms temporal resolution, they were considered as duplicate units and the one with fewer spikes would be deleted. To identify putative pyramidal cells from well-isolated units (amplitude > 36uV; mean firing rate > 0.05 spike/s; recording variability <0.35), we plotted histograms of the peak-to-valley latency and half-height width, where broad-spike cells and narrow-spike cells could be separated using a peak-to-valley latency criterion of 0.36 ms. Definitions of place cells and other function cells Spatial encoding The 3D maze was split into 400 voxel bins (bin size 75 mm × 75 mm × 75 mm). The 3D heatmap for spikes was smoothed using a 5 × 5 × 5-voxel Gaussian filter (smooth3, Matlab R2021b function in italic font hereafter). Unsampled bins were not included in the subsequent analysis. Behavioral epochs with instantaneous speed lower than 5 cm/s were excluded from analysis. The stability index was calculated by the correlation between the 1st half session heatmap and 2nd half session heatmap. The spatial information (SI) and sparsity index for each cell were calculated as Spatial information = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sum \left({p}_{i}{r}_{i}/r\right)\log 2\left({r}_{i}/r\right)$$\end{document}∑piri/rlog2ri/r and Sparsity index = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r}^{2}/\left(\sum {r}_{i}^{2}\right)$$\end{document}r2/∑ri2, where the ri is the firing rate of the neuron in the i-th bin, pi is the probability of the animal being in the i-th bin, r is the overall mean firing rate. Cells with significant spatial information (SI, bootstrap p < 0.05) and mean firing rate above 0.05 Hz were identified as place cell as previously described9,33. Theta modulation index Power spectra derived from spike-train temporal autocorrelograms were calculated at 10-ms resolution. The theta modulation index was calculated by dividing the maximum power spectrum value within the theta frequency band (5-13 Hz) by the mean spectral power across the full range (1–50 Hz). Speed encoding The instantaneous firing rates were smoothed by a 750ms-sliding-window Gaussian filter. The behavioral epochs with instantaneous speeds in a range of 5–150 cm/s were split into 15 bins. For visualization, the mean firing rate of each bin was smoothed by smoothdata using a 5-bin-width Gaussian filter. The speed score was determined by the Pearson correlation between instantaneous speeds and firing rates. Direction encoding The azimuth directions (−180 ~ 180°) were divided into 18 bins. The azimuth tuning curve was smoothed using a 3-bin-width circular Gaussian filter. The azimuth direction modulation index was expressed as the Rayleigh vector length (RVL): Rayleigh vector length = || ∑rie i2πj/N | | / ∑ri, where ri is the firing rate of the neuron in the i-th bin, N is 18 as the total bin number, j is the imaginary number. The pitch directions (−75 ~ 75°) were divided into 10 bins. The pitch tuning curve was smoothed using a 3-bin-width Gaussian filter. The pitch direction modulation index was defined as the number of bins in the top 25% of the tuning curve. Facing position encoding The facing position on the walls or floors was determined by projecting the facing direction onto the walls and floors of the 3D maze. The activity of each cell was binned (7.5 cm × 7.5 cm) and shown in a flattened heatmap. Bootstrap analysis (100 random circular shuffles of single-unit spike data) was used to determine the significance of functional cells if the measured value exceeded 95% of the null distribution of spatial information, speed score, and RLV, or if the measured value was lower than 95% of the null distribution of pitch direction modulation index. All the analyses related to direction cells and facing position cells were performed using data obtained from sessions with three fluorescent sponges on the head (n = 30 sessions). Definition of the place field and its geometry The place field is characterized by neural activities in the 3D space. For each place cell, the connected voxels (>18) within the 3D heatmap exhibiting firing rates exceeding 40% of the peak value were designated as components of place fields. We have also used other criteria (connected voxels >21, 23, or 26; 35% or 45% of the peak value) and obtained similar results. Finally, continuous components were joint together as place fields by bwlabeln. The geometry of place fields was visualized by trisurf. The centroid, volume, surface area, equivalent radius, and principal axes of place fields were calculated by regionprops3. The longest terminal-to-terminal line was defined as its principal axis, and its orientation was defined as the orientation of the place field. The azimuth direction and pitch direction of the principal axis were determined by atan2. To determine the geometry of a place field, spikes were shuffled within the radius range centered at the centroid of each original place field for 100 times to get the null distribution of the elongation index and sphericity. The place field with an elongation index higher than the 95% of the null distribution is defined as an elongated (non-isotropic) place field. The place field with a sphericity index higher than 95% of the null distribution is defined as a spherical place field, otherwise a non-spherical place field. Functional connectivity analysis The adapted dynamical differential covariance (DDC) method was used to reveal directional network structure in multi-scale neural systems34. The directional functional connectivity of A and B cells was defined based on their projection index, the Pearson correlation between A’s firing rate time series and the differential of B’s firing rate time series, where the firing rate was obtained by smoothing the single-unit spikes with a 750-ms sliding-window Gaussian filter. Each pair of cells in a group could form two connections, and a single cell could form a self-projection. For a given group of cells, the network density represents the mean probability of a cell connecting to other cells. In a directional network consisting of N nodes, its density was defined as the number of existing connections divided by maximal possible connection numbers N2. A density close to 1 indicates that all nodes are connected, while a density of 0 signifies no connections exist in the network. Sessions were included in the functional connectivity analysis if at least 6 FC cells and 6 HPC cells were simultaneously recorded. Computational modeling for PCs in FC and HPC The FC was modeled by a recurrent neural network (RNN) which had intensely interconnected nodes at hidden layers47 and may contribute to place encoding via path integration48–50. The HPC was modeled by a memory compression theory-based AutoEncoder49,51,52 with no interconnections at hidden layer for simplification. There are recurrent connections among CA3 neurons, but they are much weaker than those among cortical neurons53. Model construction The FC was modeled by an upgraded variant of RNN, Gated Recurrent Units (GRU)54,55. The model-FC in this study is constructed as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha=\sigma \left({W}_{{rec}}^{\alpha }r+{W}_{{in}}^{\alpha }u+{b}^{\alpha }\right)$$\end{document}α=σWrecαr+Winαu+bα\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta=\sigma \left({W}_{{rec}}^{\beta }r+{W}_{{in}}^{\beta }u+{b}^{\beta }\right)$$\end{document}β=σWrecβr+Winβu+bβ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau \frac{{dx}}{{dt}}=-\alpha \odot x+\alpha \odot \left({W}_{{\!rec}}(\beta \odot r)+{W}_{{\!in}}u+b+\xi \sqrt{2\tau {k}_{{rec}}^{2}}\right)$$\end{document}τdxdt=−α⊙x+α⊙Wrec(β⊙r)+Winu+b+ξ2τkrec2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r={[x]}^{+}$$\end{document}r=[x]+Where the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β represents update gate reset gate. The matrices \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{rec}}^{\alpha }$$\end{document}Wrecα, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{rec}}^{\beta }$$\end{document}Wrecβ, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{rec}}$$\end{document}Wrec are N-by-N matrices holding recurrent connection weights, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{in}}^{\alpha }$$\end{document}Winα, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{in}}^{\beta }$$\end{document}Winβ, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{in}}$$\end{document}Win are N-by-Nin matrices holding connection weights from input units to recurrent units. The vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}^{\alpha }$$\end{document}bα, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}^{\beta }$$\end{document}bβ, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$b$$\end{document}b are bias. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u$$\end{document}u is input vector. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ()$$\end{document}σ() is the sigmoid function and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${[]}^{+}$$\end{document}[]+ is threshold-linear activation function guarantees that this is a nonlinear dynamic system. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\odot$$\end{document}⊙ denotes the Hadamard product. the vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\xi$$\end{document}ξ is independent Gaussian white noise scaled by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${k}_{{rec}}^{2}$$\end{document}krec2, which presents noise intrinsic to the RNN. These leaky threshold-linear units in GRUs are modulated by the time constant \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tau$$\end{document}τ, with an update gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α and reset gate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta$$\end{document}β. The HPC was modeled by an AutoEncoder (model-HPC)56. The input and reconstruct layers of model-HPC were constructed as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x=\sigma \left(r{W}_{{\!in}}+{b}_{{in}}\right)+n$$\end{document}x=σrWin+bin+n\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r^{\prime}=x{W}_{{\!out}}+{b}_{{out}}$$\end{document}r′=xWout+boutWhere \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x$$\end{document}x is the 1 × N encoded vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r$$\end{document}r is the 1 × M activity vector of Model-FC and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{in}}$$\end{document}Win is the M × N connection matric from input layer to reconstruct layer, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}_{{in}}$$\end{document}bin is the bias. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r^{\prime}$$\end{document}r′ is the reconstructed 1 × M Model-FC activity vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${W}_{{out}}$$\end{document}Wout is the connection from the reconstruct layer to the input layer, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${b}_{{out}}$$\end{document}bout is the bias. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n$$\end{document}n represents the independence scaled noises added to avoid local optimum solution. Where the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma ()$$\end{document}σ() denotes the sigmoid function. The cortical-hippocampal interaction in this study was modeled by the forward AutoEncoder network52,57. All model constructing and training in this study were implemented by Python 3.9 package PyTorch 2.6. The hidden layer of the model-FC was 1 layer of 128 nodes, and the encoder vector layer of model-HPC was 1 layer of 32 nodes. Model training The input of the model-FC included sensory and motor information. Sensory information can be understood as environmental cues. The 3D maze was split into 75 bins (0.3×0.3×0.25m3 for each cubic). Thus, the one-hot code has 75 bits, for instance, 00010…0 means animal was at the 4th cubic. We used animal’s behavioral trajectories as datasets to train and test the model by transforming the real trajectory into time series of 75 bit one-hot codes. Speed (1×1 vector) and direction (1×3 vector) comprised motor information. Random noises (1×2 vector), another consonant matrix, mapped a 1×81 vector to a 1×256 vector as the final input to RNN. The mapping matric was constant during model training (no back propagation) and was randomly initiated as an 81×256 matric ranges from −1 to 1. The output variables included position (1×3 vector), speed (1×1 vector), direction (1×3 vector) and noise (1×2 vector). Their relative weights were 5.3, 0.4, 2.7 and 0.02, respectively. The model-FC may predict future positions of a few steps ahead based on current speed and direction information. The weight for current time step, 50 ms future step and 100 ms future step were 0.8, 2.2 and 0.6 respectively. The loss function during training was mean squared error (MSE) without regularization method, and was minimized by Adam stochastic gradient descent (SGD). All the parameters of the model were initialized to 0. The learning rate was initially set to 0.0001, then dynamically adjusted from 0.005 to 0.0001 as training progressed from 20 to 800 epochs, before being fixed at 0.0001 after reaching the 800-epoch threshold. For each experiment, the model was trained for 800 to 1200 epochs after initialization, despite the loss function having already converged before 200 epochs. Analysis of modeling results After training, all the trainable parameters including connection matrices and bias were frozen. Model did same free foraging task using 400–3000 traveling passages to cover the 3D space. The neuronal activities at 128 model-FC nodes (neurons) and 32 model-HPC nodes (neurons) were extracted for analysis. Only nodes with sufficient activity (variation>0.06) were included in analysis. The tunning curve was offset to zero. Function cells were identified same as described above. For each place cell, the voxels within the 3D heatmap exhibiting firing rates exceeding 70% of the peak value were designated as components of place fields. Quantification and statistical analysis Data are summarized as mean values with standard error of the mean (SEM). The box plot shows values at min, 25%, 50%, 75% and max. The cell number is indicated with n. Statistical analysis was performed in MATLAB (R2021b, MathWorks). If the normal distribution of a dataset was rejected by Kolmogorov-Smirnov test, non-parametric statistical tests were used. Two-sided Wilcoxon rank sum test and unpaired t-test were used to determine the statistical significance. Multiple comparisons correction was not applied unless stated otherwise. Statistical parameters and significance are reported in the text or in the figure legends. The significance threshold was placed at p = 0.05 (n.s., p > 0.05; *, p < 0.05; **, p < 0.01; ***, p < 0.001). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file