Proposed relationship between grid cells and place cells. Each place cell captures the conjunction of what happened and where it happened. (A) When revisiting a position, the memory associated with that position is retrieved, providing feedback to the non-spatial what cells (e.g., a 60 Hz electronic noise), which reside in medial entorhinal cortex (mEC). Because the non-spatial attribute is common to all positions (e.g., the electronic noise occurs everywhere), and because various mechanisms ensure that cells avoid firing constantly at their maximum possible rate during the entire recording session, grid cells fire preferentially in the locations of each memory. (B) If memories are formed whenever the current situation is sufficiently different from prior situations, this results in a hexagonally arranged grid of place cells that tile the two-dimensional surface because the only thing that varies in a typical navigation experiment is X/Y location; each memory is formed whenever prior memories are sufficiently far away (the circles represent a fixed dissimilarity between memories).

Assumed connectivity between place cells in hippocampus and border cells, head direction cells, and a grid cell in medial entorhinal cortex. When a memory is formed, a place cell (e.g., p1) is recruited and the weights between the entorhinal inputs and the place cells are set equal to the inputs (e.g., W1j for weights to p1), with j ranging over the 13 entorhinal inputs to hippocampus. These weights are bidirectional, with feedback supporting memory recall. Because feedback modulates the response of the grid cell, this produces a higher firing rate at positions where the non-spatial attribute is remembered. Code for the model can be found at: https://github.com/dhuber1968/GridCellMemoryModel.

Comparison between memory retrieval with two orthogonal dimensions (X/Y) versus three non-orthogonal dimensions (E/F/G) that are 60 degrees apart. Each dimension is represented by a circular basis set with three equally spaced sine waves with a period of 2. When a place cell is learned, the weights connecting each sine wave input and the place cells are set equal to the input values. (A) In the case of orthogonal X/Y dimensions, this results in a pattern of 6 weight values (w1 – w6) across the two dimensions, as shown by the intersection of the red lines emanating from the position where the memory was formed (the red dot) and the three sine waves for each dimension. After memory formation, the current position (green dot) reactivates the memory based on the 6 current position response values (r1 – r6), summing the multiplication of the response values and the weight values. (B) The graph shows the result of randomly sampling 1,000 different memory positions and retrieval positions, plotting retrieval strength as a function of Euclidean distance for each pair of positions. Retrieval strength is variable with Euclidean distance because the sum across the two orthogonal dimensions is a city-block metric (e.g., the same Euclidean distance can map onto multiple city-block distances). (C) To capture Euclidean distances, three non-orthogonal dimensions (E/F/G) were used. (D) This produces a retrieval function that is approximately monotonic with Euclidean distance.

Memory encoding, memory consolidation, and an example sequence. (A) When first entering a novel environment, the animal creates a memory of the non-spatial attributes of that environment (e.g., the 60 Hz sound of electronics) at each location where the attribute is found. The gray curved arrow shows the random path taken by the simulated animal and the blue dots show the positions where memories are created. The activation threshold, θa (blue dashed circle), dictates whether previously created memories are retrieved, or if none are retrieved, a new memory is formed. This produces a minimum distance between memories. (B) The representations of memories are altered by an online consolidation process that produces unbiased memory representations that tile the environment (i.e., a cognitive map). In this process, the most strongly active memory (the yellow dot retrieved memory) is slightly altered in relation to competing memories that are also activated by the current position (the red and green dots). Other memories (gray dots) remain inactive because they are too dissimilar to the current position (outside the blue dashed circle centered on current position). After initial memory retrieval, the retrieved inputs are used to activate the competing memories and this strength of activation of competitors is compared to a consolidation threshold, θc (yellow dashed circle, centered on retrieved memory), which is smaller than the activation threshold, such that consolidation pushes memories to become maximally dissimilar (pattern separation). Competing memories that are more active than the consolidation threshold (red dot) push the weights of the retrieved memory away from the competing memory (red arrow). Competing memories that are less active than the consolidation threshold (green dot) pull the weights of the retrieved memory towards the competing memory (green arrow). For memories arranged in two real-world dimensions, this typically results in activation of three surrounding memories and consolidation makes the triangle formed by these memories an equilateral triangle (gray arrow). (C) An example path with 1,000 simulated steps is shown, with the blue circles indicating the initial positions of memories and the yellow circles indicating memory positions after consolidation. The red dots show positions where the simulated grid cell fired. The firing threshold was set such that the cell fires 5% of the time, resulting in 50 positions where the grid cell fired. A movie showing the simulation in C, including memories that flash yellow, red, and green as outlined in B, can be found at: https://youtu.be/Ts66gBxGdWs.

Position of square recording enclosure relative to circular border cell dimensions. The 9 outer graphs show simulated spike rate maps for the border cells under the assumption that each border cell fires 5% of the time based on the summation of its bottom-up and top-down inputs (Equation 5). The enclosure was assumed to be half as wide as the full period of the circular border cells, minimizing confusion between opposite borders. The dashed lines indicate wraparound seams such that each dashed line is connected to the opposite dashed line to create a cylinder for that dimension, resulting in a hexagonally connected 3-torus for the entire space across the three non-orthogonal dimensions. The gray curved lines within each of the 9 border cell graphs show the path of a simulated animal across 10,000 steps. The red dots (500 per graph) show positions where the simulated border cell fired. The graphs for the G and F dimensions are rotated to align the graphs with the corresponding allocentric directions. The preferred positions for each of the 9 cells are indicated by the entire line length of the red, green, or blue arrows that point to the corresponding firing map. The letter labels inside each graph indicate the simulated cell using the same labeling scheme as in Figure 2.

Results for simulated grid cells representing a non-spatial attribute common to a set of place cell memories when not including head direction in the place cell memories (allocentric memories). These simulations are an exploration of how the model behaves with different parameter values. In each case, the first four simulations are shown, regardless of outcome. Results are shown when adopting one of three different consolidation thresholds, θc, which produce different spacings between memories. The corresponding activation thresholds, θa, were .86, .9, and .92 to make sure that memories were created and activated with a somewhat closer spacing than that dictated by the consolidation threshold. Each pair of novel and familiar firing maps is the same simulation, with the novel firing map showing the first 10,000 simulated steps in a novel environment and the familiar firing map showing 10,000 simulated steps after 100,000 prior steps (e.g., after the equivalent of 10 sessions of prior experience).

Simulation results when using the same parameters and settings as in Figure 6 for a circular enclosure.

Results from a simulation that included head direction (.8 consolidation threshold). These results are for a familiar box (10 sessions of prior experience). The plotted firing maps (red dots for spikes) and spatial autocorrelation (blue for low correlation and yellow for high correlation) maps are labeled according to the specific cell labels used in Figure 2. The border cell results for this particular simulation are shown in Figure 5. (A) The grid fields for the non-spatial attribute, k, are vertically aligned. (B) In contrast the non-spatial attribute, the grid fields for the head direction cells are horizontally aligned. Cells h1 and h3 have the same grid orientation and spacing as each other, as would cell h2, except that for h2, its corresponding place cell memories are mostly outside the box. As in real grid cell modules, the grid fields for each of the three head direction cells (h1 to h3) are shifted relative to each other. (C) The head direction grid spacing is the square root of 3 larger than the non-spatial attribute grid spacing, reflecting the superposition of face-centered cubic packing layers. (D) For this grid spacing, there were 6 HD layers, such that every third layer produced a nearly identical spatial phase of HD-sensitive place cell memories, except that the preferred HD was 180 degrees opposite. This provides an allocentric arrangement of the place cell memories whereby the non-spatial attribute at each location was remembered from two 180 degree opposite viewpoints (head directions). The colored dots in the three-dimensional plot are based on the final positions of the 14 memories at the end of the simulation, as calculated from the weight matrix for each place cell, with the color of the dots showing the preferred head direction of the corresponding memory.

Simulations depicting the learning of non-spatial grid fields (k) and head direction conjunctive grid field (h), using the same parameter values as in Figure 8. (A) Average grid scores across 100 simulations, with prior experience ranging from novel (no prior experience) to 10 sessions of prior experience, show that the grid field of the non-spatial cell, k, was immediately apparent whereas the head direction grid cells required prior experience before grid fields stabilized. Because one of the three head direction grid cells tended to have a single central grid field, with potentially outside the box additional grid fields (see cell h2 in Figure 8), the maximum grid score (hmax) from the three head direction cells was used. Error bars are plus and minus one standard error of the mean. (B) A plot showing the number of place cell memories created by the end of the simulation reveals that the number of memories grew in a similar manner to the stabilization of the head direction grid fields. (C) Representative firing map and spatial autocorrelation results for a novel environment and a familiar one reveal that that the non-spatial cell’s grid fields were regular but less precise in a novel environment, reflecting the shifting nature of the place fields. (D) Plots showing the final positions of all place cell memories for the representative simulations, reveal that the main effect of prior experience was formation of outer memories that conform to the shape of the box; as the animal learned head direction representations for the borders of the box, the grid cell module based on head direction stabilized. The color of the dots represents the head direction associated with each memory using the same color scale as Figure 8. A movie showing memory formation and consolidation in this situation that includes head direction be found at: https://youtu.be/_yNb-x4an_A.

Using the same parameters as Figures 8-10, a comparison of place cell firing maps for interior versus exterior place cells demonstrates how head direction sensitivity depends on location and how interior place cells give rise to head direction conjunctive grid cells. (A) The final positions (preferred positions and preferred head directions) of the 48 place cell memories and head direction cell firing maps. (B) Firing maps of the place cells. The place cells were divided into 29 “exterior” place cells that were active less than 2.5% of the time (these are accumulated onto the same graph on the left) versus 19 “interior” place cells (one graph per cell on the right). The 8 interior place cells in the top two rows of were the cause of the head direction conjunctive grid cell firing map shown immediately above in panel. The color bar that relates head directions to colors includes the labels for the head direction cells (h1, h2, and h3) and the adjacent arrows depict the three head directions. The 11 interior place cells in the bottom two rows selected for head directions that were between the preferred head directions of the head direction cells (i.e., the other 3 hexagonal layers from the 6-layer hexagonal close packing). The color of each spike for the place cells shows the head direction of the simulated animal at the time of the spike. For the place field center of the interior place cells, head direction sensitivity was weak (i.e., a greater range of colors in the center of each place field), whereas the firing maps for the exterior place cells were highly view dependent (consistent color for each cluster of spikes).

Simulation results using the wide-spacing activation/consolidation threshold parameter values from Figure 6/7 but when including head direction in the hippocampal place cell memories. (A) The non-spatial attribute cell (k) and head direction cells (h1 – h3) produced similar grid field patterns to those seen in Figures 9/10. (B) The right-hand graph shows the final preferred positions of hippocampal place cells, with color indicating preferred head direction. The wider spacing was satisfied through 3 rather than 6 layers along the head direction dimension (only 3 different colors). Six interior place cells (p1 – p6) are selected for additional analyses. The lower-left graph plots x-position by head direction for the firing map of cell p2, collapsing over y-position: the cell was active for a full 2/3 of head directions at its place field center. The color of each spike indicates the head direction of the spike according to the color map. (C) Head direction sensitivity of the place cells (shown in six firing map plots for each of six place cells) for the six interior place cells. The black dots show approximate place field centers, and the black arrows show approximate head direction sensitivity in different locations. For cell p2 head direction sensitive increases with distance from the center. For the other 5 cells, a different pattern emerges along the borders of the box. Along the borders, some head directions are not empirically observed, as shown by the inset graph for cell p4, which plots x-position by head direction. The absence of data for a subset of head directions gives the spurious appearance of head direction sensitivity that is in line with the borders in one direction or the other, or possibly both directions.

Place cell responses using same parameters as Appendix Figure 1 but with an enclosure that is 5% as tall as it is wide, simulating behavior in a highly familiar narrow passage, such as an arm of a radial arm maze or an elevated track. (A) The positions of all 15 consolidated memories are shown, with the color indicating preferred head direction according to the color map. (B) The firing maps show the place field for each place cell with spike color indicating head direction. Even though these same parameter values produced place cells that were relatively insensitive to head direction in the open field enclosure of Appendix Figure 1, in this case each place cell appeared to only prefer one direction or the other (blue or red) along the narrow passage.

Simulation results for a familiar circular enclosure followed by a change to a square enclosure placed in the same context, containing the same non-spatial attribute. (A) The memory locations (dots colored by head direction) and firing maps for the familiar circular enclosure show a 6-layer face-centered cubic packing for the head direction grid module (h1-h3). The border cells (e1-e3, f1-f3, and g1-g4) captured one side or the other except for the between-border cells, which show a patchy response owing to memory feedback. (B) Upon entry into a novel square enclosure that contained the same non-spatial attribute as the circular enclosure and was the same width/height as the circular enclosure, the head direction grid module was immediately apparent (in contrast to the results reported in Figure 9). This occurred because the memories created in the circular environment were recalled in the square environment and the geometry was similar enough that very few of the retrieved memories changed their locations. The “tails” emanating from each place cell memory location show the shift from the prior location in the circular enclosure to the new location in the square enclosure. The color of the tail indicates the prior head direction of the memory if head direction of the memory changed with consolidation (e.g., the blue memory on the upper border has a pink tail). As shown, there was a subtle change in the positions of the exterior place cells to accommodate the novel square shape. Two exterior place cells migrated to interior positions (long red and blue tails) and, correspondingly, two interior place cells migrated to exterior positions (cyan and pink tails). The length of the tails is sometimes misleading owing to the circular nature of the border cells (in truth, the two very long red and blue tails migrated in a wraparound fashion, which was a much smaller change). Finally, one new memory was formed (yellow memory encircled in black).

Simulation results for a familiar rectangular enclosure followed by a change to a square enclosure placed in the same context, containing the same non-spatial attribute. (A) As with the narrow passage in Appendix Figure 2, the familiar rectangular enclosure was too narrow to produce a highly regular hexagonal grid. (B) In switching to the much larger square enclosure, most place cells changed their locations (global remapping) and 5 new memories were formed. Despite the considerable disruption in the place fields, the memories from the rectangular enclosure allowed for the rapid establishment of the hexagonal grid, including head direction conjunctive grid cells (see Appendix Figure 3 caption for additional information).