Summary
Hippocampal place cells are spatially tuned neurons that serve as elements of a “cognitive map” in the mammalian brain1. To detect the animal’s location, place cells are thought to rely upon two interacting mechanisms: sensing the animal’s position relative to familiar landmarks2,3 and measuring the distance and direction that the animal has travelled from previously occupied locations4–7. The latter mechanism, known as path integration, requires a finely tuned gain factor that relates the animal’s self-movement to the updating of position on the internal cognitive map, with external landmarks necessary to correct positional error that eventually accumulates8,9. Path-integration-based models of hippocampal place cells and entorhinal grid cells treat the path integration gain as a constant9–14, but behavioral evidence in humans suggests that the gain is modifiable15. Here we show physiological evidence from hippocampal place cells that the path integration gain is indeed a highly plastic variable that can be altered by persistent conflict between self-motion cues and feedback from external landmarks. In a novel, augmented reality system, visual landmarks were moved in proportion to the animal’s movement on a circular track, creating continuous conflict with path integration. Sustained exposure to this cue conflict resulted in predictable and prolonged recalibration of the path integration gain, as estimated from the place cells after the landmarks were extinguished. We propose that this rapid plasticity keeps the positional update in register with the animal’s movement in the external world over behavioral timescales (mean 50 laps over 35 minutes). These results also demonstrate that visual landmarks not only provide a signal to correct cumulative error in the path integration system, as has been previously shown4,8,16–19, but also rapidly fine-tune the integration computation itself.
Path integration is an evolutionarily conserved strategy for self-localization that allows an organism to maintain an internal representation of its current location by integrating over time a movement vector representing distance and direction travelled 4–7. Place cells and entorhinal grid cells have been implicated as key components of a path integration system in the mammalian brain20–22. Thus, we recorded place cells from area CA1 (Extended Data Fig. 1) in 5 rats as they ran laps on a 1.5 m diameter circular track while foraging for liquid reward dispensed at randomized spatial intervals. The track was enclosed within a planetarium-style dome where an array of three visual landmarks was projected onto the interior surface to create an augmented reality environment (Fig. 1a,b). In contemporary virtual reality systems3, 23–25, head- or body-fixed rats fictively locomote on a stationary air-cushioned ball or treadmill. Notwithstanding the flexibility of these systems to manipulate the visual experience of the animal, we built the dome apparatus to instead more completely preserve natural self-motion cues, such as vestibular, proprioceptive, and motor efference copy. This system enabled us to test the a priori hypothesis that manipulating the animal’s perceived movement speed relative to the landmarks results in a predictable recalibration of the path integration gain.
To create the visual illusion that the animal was running faster or slower, the array of landmarks was rotated coherently as a function of the animal’s movement speed. Movement of the landmarks was controlled by an experimental gain, G, which set the ratio between the rat’s travel distance with respect to the landmarks (landmark reference frame) and its travel distance along the stationary circular track (laboratory reference frame) (Fig. 1c). Recording sessions always began with G = 1 (Epoch 1), a control condition where visual landmarks remained stationary alongside the track, so that the rat traveled the same distance in both the landmark and laboratory frames (Fig. 1d). The gain was then gradually changed over the course of multiple laps (Epoch 2) to become less than or greater than one. For G < 1, the landmarks moved at a speed proportional to (but slower than) the rat in the same direction; hence, the rat ran a shorter distance in the landmark frame than in the laboratory frame. For G > 1, the landmarks moved in the opposite direction; hence, the rat ran a greater distance in the landmark frame than in the laboratory frame. In Epoch 3, G was held at a steady-state target value (Gfinal). In some sessions, the landmarks were then extinguished (Epoch 4) to assess whether the effects of gain adjustment persisted in the absence of the landmarks.
Under gain-adjusted conditions, CA1 units (mean 7.2 ± 5.8 S.D. units/session) tended to fire in normal, spatially specific place fields when the firing was plotted in the reference frame of the visual landmarks, but not when plotted in the reference frame of the lab (Fig. 1e). The strength and continuity of visual cue control over the place fields is highlighted by special cases of G (Fig. 2). As G was ramped down to 0, the place fields became increasingly large in the laboratory frame (Fig. 2a; Extended Data Video 1). As G approached 0, individual units maintained normal place fields only in the landmark frame (Fig. 2b), which resulted in spiking activity that spanned multiple laps in the laboratory frame. At G = 0, the animal’s position became locked to the landmark frame, as the landmarks moved in precise register with the rat. Consequently, a unit that was active at that moment would typically remain active throughout Epoch 3, (e.g. yellow unit, Fig. 2a); in contrast, a unit that was inactive at that moment would typically remain silent throughout Epoch 3 (e.g. red unit, Fig 2a). When G was clamped at integer ratios such as 3/1 (Fig. 2c) or 1/2 (Fig. 2e), the units maintained the typical pattern of one field/lap in the landmark frame, while firing at the expected periodicity such as 3 times per lap (Fig. 2d) or every other lap (Fig. 2f) in the lab frame. Remapping events sometimes caused different populations of place cells to be active at different times. For example, place cells active during the initial part of the session sometimes went silent (loss of field; Fig. 2e, yellow unit), and place cells silent during the initial part of the session sometimes began firing at a preferred location (gain of field; Fig. 2e, red unit). The remapped cells exhibited normal place fields only in the landmark frame. These examples illustrate that the landmark array exercised robust control over the place fields, outweighing any subtle, local cues on the apparatus as well as nonvisual path integration cues, such as vestibular or proprioceptive cues.
To quantify the degree of landmark control over the population of recorded place cells, we developed a novel decoding algorithm that was robust to the remapping events described above. We measured each unit’s time-varying spatial frequency (i.e., the frequency of repetition of its spatially periodic firing pattern) from which we then computed a hippocampal gain, Hi, for every individual unit, i. The median value of Hi over all simultaneously recorded active units during a given set of laps was taken as a population estimate of the hippocampal gain, H, for those laps. Just as G quantifies the ratio between the rat’s travel distance in the landmark frame versus laboratory frame, H quantifies the ratio between the rat’s travel distance in the internal hippocampal “cognitive map” frame1 versus the laboratory frame. Hence, if the hippocampal cognitive map is anchored to the landmark frame, then the experimental and hippocampal gains should be identical (G = H). An ensemble coherence score for each unit was computed as the mean value over the session of | 1 - Hi / H |, measuring the deviation of Hi from H (Methods). The distribution of coherence scores (Fig. 2g) shows that Hi was within 2% of H for 80% (399/500) of individual units, and deviations >5% were rare. Even when individual cells remapped, they still exhibited spatial periodicity at gain factors Hi which were close to H (see red and yellow units in Fig. 2c). Hence, the population of place cells acted as a rigidly coordinated ensemble from which a precise estimate of H could reliably be computed, despite occasional remapping by some place cells.
We quantified the degree of cue control in each session by computing the mean ratio H/G for Epochs 1-3 of a session; a ratio close to 1 indicates that the cognitive map was anchored to the landmark frame. The majority of sessions (83.33%, 60/72) exhibited H/G near 1, but the rest showed substantially larger H/G (> 1.1) indicating loss of landmark control (Fig. 2h; Extended Data Fig. 2). For sessions with H/G < 1.1, the spatial information per spike in the landmark frame far exceeded that in the laboratory frame (Fig. 2i; paired t-test, n = 5 rats, t4 = 6.213, p = 0.0034). We restricted further quantitative analyses to these sessions, which we defined as demonstrating ‘landmark control’. These results indicate that the augmented reality dome was successful in producing the desired illusion by strongly controlling the spatial firing patterns of the hippocampal cells in the majority of sessions (Extended Data Figs. 3, 4).
Despite strong cue control in the majority of sessions, place fields tended to drift by a small amount against the landmark frame on each successive lap (Fig. 3; Extended Data Fig. 5; also visible in Figs. 2a,c,e and 4a,b) leading to total drifts of up to ~80° over the course of a session. When G was gradually decreased, place cells tended to fire earlier in the landmark frame with each successive lap, whereas when G was gradually increased, place cells tended to fire later with each successive lap. The accumulated drift over each session was linearly correlated with Gfinal during Epoch 3 for that session (n = 55 sessions, r53 = 0.61, p = 7.2 × 10-7; Fig. 3c). The direction of this systematic bias was consistent with a continuous conflict between the dynamic landmark reference frame and a path-integration-based estimate of position (although we cannot rule out the possible contribution of subtle uncontrolled external cues on the track or in the laboratory). That is, when path integration presumably undershot the landmark-defined location systematically (G < 1), the place fields shifted slightly backwards in the landmark frame; conversely, when path integration overshot the landmarks (G > 1), the place fields shifted forward. The shift may reflect a conflict resolution that is weighted heavily, but not completely, in the direction predicted by the landmark reference frame.
Given the apparent influence of the path-integration circuit revealed by systematic place field drift even in sessions with strong landmark control, we tested whether anchoring of the cognitive map to the gain-altered landmark frame induced a recalibration of the path integrator that persisted in the absence of landmarks. Such recalibration would be evidenced by a predictable change in the hippocampal gain H when visual landmarks were extinguished (Fig. 1d, Epoch 4). If the path integrator circuit were unaltered, one would expect that the place fields would revert to the laboratory frame (H ≈ 1). Alternatively, if the path integrator gain were recalibrated perfectly, one would expect instead that the place fields would continue to fire as if the landmarks were still present and rotating at the final experimental gain (i.e., H ≈ Gfinal). We found that the hippocampal representation during Epoch 4 was intermediate between these extremes (Fig. 4a,b; Extended Data Video 2): there was a clear, linear relationship between Gfinal and the hippocampal gain H estimated during the first 12 laps after the landmarks were turned off (n = 38 sessions, r36 = 0.94, p = 7.9 × 10-19, Fig. 4c). Moreover, this linear relationship was maintained when H was estimated during the next 12 laps (n = 18 sessions, r16 = 0.87, p = 3.37 × 10-6, Fig. 4d). The values of H for the first and second 12 laps were highly correlated (n = 18 sessions, r16 = 0.972, p = 1.72 x 10-11, Fig. 4e) with a slope near 1. Thus, H was stable over at least 18 laps (i.e., the middle of the second estimation window). Despite this overall stability, there were still fluctuations in H in the absence of landmarks (Fig. 4f, Extended Data. Fig. 6). We tested whether changes in behavior could account for the hippocampal gain recalibration by computing several behavioral measures for each epoch, such as running speed and number of pauses/lap (see Extended Data, Behavioral Analysis). Multiple regression analysis showed that Gfinal strongly predicted H, whereas the behavioral variables had negligible influences on H (Extended Data Table 1).
Using a novel augmented reality dome apparatus, we show here that the path integration system employs a modifiable gain factor that can be recalibrated to a new value that can remain stable for at least several minutes in the absence of salient landmarks. This sustained recalibration can be detected from the spiking activity of hippocampal place cells. Recalibration of this nature has been described extensively in other systems. The cerebellum plays a key role in recalibration of feedforward motor commands during reaching tasks in artificial force fields and during walking on split-belt treadmills26. Similarly, the gain of the vestibulo-ocular reflex adapts to changes in the magnitude of retinal slip caused by magnifying glasses, an effect that persists even after the glasses are removed27. As with our own results, this recalibration is not perfect in these motor adaptation tasks; i.e., the gain measured after the training trials are biased towards, but not precisely the same as, the experimental gain implemented during the training trials. To our knowledge, such gain recalibration has not been demonstrated physiologically in cognitive phenomena such as spatial representation and path integration (but see 15). The lack of complete recalibration may be due to an insufficient number of training laps during Epoch 3, or may reflect inherent limits on the plasticity of the path integrator gain variable.
It is widely accepted that visual landmarks provide a signal to correct error that accumulates during path integration28. The results in this paper demonstrate physiological evidence for a role of vision in the path integration computation itself by providing an error signal analogous to retinal slip in the VOR27. Specifically, this error signal fine-tunes the gain of the path integrator15, minimizing the accumulation of error in the first place. Although it perhaps would not have been surprising to observe gain recalibration over developmental time scales, the rapid recalibration reported in this paper indicates that the path integration gain is constantly and actively fine-tuned even at behavioral time scales. This fine-tuning may be required to (a) maintain accuracy of the path integration signal under different behavioral conditions (e.g., locomotion in the absence of salient landmarks; locomotion on different surfaces that provide varying degrees of slip and cause alterations in the self-motion inputs to the path integrator); (b) synchronize the different types of self-motion signals (e.g., vestibular, optic flow, motor copy, or proprioception) thought to underlie path integration; and (c) coordinate the discrete set of different path integration gains thought to underlie the expansion of grid scales along the dorsal-ventral axis of the medial entorhinal cortex12,29,30. The recalibration might be implemented by changes to the head direction31 or speed32,33 signals that provide input to a path integration circuit. Alternatively, these representations may be unaltered and the gain changes are implemented by changing the synaptic weights between the inputs and putative attractor networks that perform the path integration9–11, 13. The augmented reality system described here will allow the investigation of mechanisms underlying the interaction between external sensory input and the internal neural dynamics at the core of the path integration system.
Author Contributions
J.J.K., N.J.C., and H.T.B. conceived the study. All authors designed the study. J.J.K. and N.J.C. advised on all aspects of the experiments and analysis. F.S. made key contributions to the analysis and interpretation of the data and provided supervision over data acquisition and analysis. R.P.J. and M.S.M. designed and constructed the augmented reality apparatus, performed neurophysiological experiments, and analyzed the data. R.P.J., M.S.M., N.J.C., and J.J.K. wrote the paper and F.S. and H.T.B. provided critical feedback.
Extended Data Video 1: Extreme control of place fields by landmarks. (left) Reproduction of Fig 2a, augmented with moving time marker (vertical dashed line). (right) Overhead videos (8x speed) of the rat running in the dome apparatus as viewed with respect to two distinct frames of reference, synchronized to the time marker in the left plot. Videos show the last ~ 6.5 laps (~ 8 min in real time) of Epoch 2 (G ramps to 0). The circular object in the center is the hemispherical mirror (not visible to the rat) used to project images to the inside surface of the dome. Reflections of the three landmarks as well as the annular ring can be seen in the mirror (a small lens artifact also appears on the mirror but was not visible to the animal). Spikes from the same three putative pyramidal cells (red, blue, yellow) are shown at the angular position of the rat. For clarity, spikes only persist for about one lap in their respective frame. (right, top) Original video recorded with respect to the laboratory frame. The yellow place cell is active for ~ 4 laps (over 4 minutes). (right, bottom) Modified video, counter-rotated by the landmark manipulation angle. This results in the reflection of landmarks on the mirror appearing stationary (with a small jitter due to video timestamp resolution). The yellow place cell forms a single field subtending ~180° to 0°.
Extended Data Video 2: Recalibration of place fields by landmarks. Same format as Extended Data Video 1. (left) Reproduction of Fig 4a. (right) Videos show approximately the last 4 laps of Epoch 3 (landmarks on) and the first 8 laps of Epoch 4 (landmarks off). (right, top) The place cells do not fire in consistent locations in the laboratory frame. (right, bottom) In the landmark frame, place cells fire in consistent locations during Epoch 3 and then drift slowly in the absence of landmarks during Epoch 4, because the hippocampal gain, H, is close to (but not identical to) the final experimental gain, Gfinal. (During Epoch 4, landmarks are off but the video is counter-rotated as if the gain were Gfinal.)
Acknowledgments
We thank Bill Nash and Bill Quinlan for technical assistance with the construction of the augmented reality apparatus; Marissa Ferreyros, Macauley Breault, Nick Lukish, Jeremy Johnson, and Douglas GoodSmith for technical assistance in running experiments; Balazs Vagvolgyi for software development for validation via camera tracking, Geeta Rao, Vyash Puliyadi, Cheng Wang, Heekyung Lee, Robert Nickl, and Jonathan Bohren for helpful discussions and technical advice; and Adrian Haith for discussions and comments on the manuscript. This research was supported by NIH grants R01 MH079511 (HTB, JJK), R21 NS095075 (NJC, JJK), and R01 NS102537 (NJC, JJK, FS), a Johns Hopkins University Discovery Award (NJC, JJK), a Johns Hopkins Science of Learning Institute Award (JJK, NJC), a Johns Hopkins Kavli Neuroscience Discovery Institute Postdoctoral Distinguished Fellowship (MSM) and a Johns Hopkins Mechanical Engineering Departmental Fellowship (RPJ).