Skip to main page content
Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
, 112 (1), 75-116

The Temporal Context Model in Spatial Navigation and Relational Learning: Toward a Common Explanation of Medial Temporal Lobe Function Across Domains


The Temporal Context Model in Spatial Navigation and Relational Learning: Toward a Common Explanation of Medial Temporal Lobe Function Across Domains

Marc W Howard et al. Psychol Rev.


The medial temporal lobe (MTL) has been studied extensively at all levels of analysis, yet its function remains unclear. Theory regarding the cognitive function of the MTL has centered along 3 themes. Different authors have emphasized the role of the MTL in episodic recall, spatial navigation, or relational memory. Starting with the temporal context model (M. W. Howard & M. J. Kahana, 2002a), a distributed memory model that has been applied to benchmark data from episodic recall tasks, the authors propose that the entorhinal cortex supports a gradually changing representation of temporal context and the hippocampus proper enables retrieval of these contextual states. Simulation studies show this hypothesis explains the firing of place cells in the entorhinal cortex and the behavioral effects of hippocampal lesion in relational memory tasks. These results constitute a first step toward a unified computational theory of MTL function that integrates neurophysiological, neuropsychological, and cognitive findings.


Figure 1
Figure 1. TCM describes the recency effect in immediate, delayed and continuous-distractor free recall
Experimental and predicted values of the probability of first recall, a sensitive measure of the recency effect across delay schedules. a. In immediate free recall, the recall test follows immediately after the presentation of the last item. b. In the delayed condition, sixteen seconds of a distractor task intervened between presentation of the last list item and the recall test. Accordingly, the recency effect, the advantage for recall of the last items in the list, was greatly reduced. c. In continuous-distractor free recall, sixteen seconds of distractor intervened between the last item of the list and the recall test, but also in between each item of the list, effectively “stretching out” the list while preserving the relative temporal spacing of the list. Under these circumstances, the recency effect was much larger than that observed in delayed recall. Because information that enters ti decays gradually, TCM, when coupled with a competitive retrieval rule, can describe the persistence of the recency effect across time scales. Model results are from Howard and Kahana (2002a). The experimental data is taken from Howard and Kahana (1999).
Figure 2
Figure 2. TCM provides a natural explanation of asymmetric association in free recall.
a. In TCM there are two sources of associative effects. One source relies on the ability to retrieve contextual elements consistently from presentation to presentation of an item. The cue strength derived from these “old” item-to-context associations provides an asymmetric cue that only helps recall items forward in the list. The other source is the ability of an item to retrieve contextual elements that were already present when the item is presented. The cue strength derived from these “new” item-to-context associations provides a symmetric cue that helps both forward and backward recalls. The combination of these two cues leads to the characteristic shape of the CRP. After Howard and Kahana (2002a). b. The combination of an asymmetric retrieval cue and a symmetric retrieval cue is an asymmetric retrieval cue. This results in good quantitative fits to observed CRP curves. The left panel shows data from a delayed free recall study of younger adults along with predicted data from TCM. The right panel shows analogous curves from older adults. The decrease in associative tendencies for older adults was modeled as a result of including a noise term in Eq. 9. This data was originally presented in Kahana, et al (2002). The modeling of the older adults’ data is explained in greater detail in Howard et al (in revision).
Figure 3
Figure 3. A linking hypothesis between TCM and the MTL
a. “Items” are patterns of activity in semantic memory (SM), which is presumed to reside in cortical association areas. These areas project to parahippocampal (PH) regions, including at least EC, which support a state of context ti which serves as the cue for episodic recall. Presentation of an item in semantic memory calls up a set of elements tiIN in PH. The state of context also includes patterns activated by previous item presentations (the red and green patterns). The set of elements activated by the item causes a set of elements in the hippocampus (H) to be activated, perhaps biased by the other contextual elements active in EC and/or the prior state of activation in H. Hebbian association (indicated by the thin solid lines) takes place between the state of context in PH and the state in semantic memory to allow contextual states to cue the item in semantic memory. b. Repetition of the item in semantic memory reactivates the stimulus-specific elements in PH. Because the stimulus-specific elements remained active in PH following the initial presentation of the stimulus, their reactivation serves as a cue for items that followed the initial presentation. c. The proposed function of the hippocampus is to allow retrieval of contextual states upon re-presentation of an item. In this case, when the item is re-presented in semantic memory, it again activates the set of stimulus-selective elements in PH, as in b. However, H functions to reinstate the entire contextual state that obtained when the stimulus was originally presented. Because this state includes elements derived from items presented prior to the original item presentation, this “retrieved context” functions as a symmetric cue for recall of other stimuli.
Figure 4
Figure 4. Retrospective encoding requires an imperfect spatial representation
a. Simple schematic model of the paths taken by the animal in the W-maze. The animal repeatedly traveled the path 1-2-3-4-. . . -10-11-12-1-2-3. . . . Initially the animal traveled from the center arm to the left arm, a center-left trip (steps labeled 1–3), followed by a left-center trip (4–6), followed by a center-right trip (7–9) and a right-center trip (10–12). A different representation on step 6 compared to step 12 is evidence for retrospective coding. b. A simplified version of the TCM context evolution equation was presented with velocity vectors corresponding to the series of movements to generate a positional representation p. We defined retrospective encoding as 1 − (p6·p12). This reflects the degree to which p6 and p12 are different from each other. Retrospective coding is plotted as a function of ρ in a general integration scheme, where pi = ρpi−1 + vi. When ρ 0, p is just the most recent movement and the model provides a “pure head direction” representation. When ρ = 1, p reflects the sequence of all prior movements and the model provides a perfect place representation. At both of these extremes, the model fails to show evidence for retrospective coding. In contrast, for intermediate values of ρ, the model shows retrospective coding, as seen in EC and the hippocampus (Frank, et al, 2000). Although this is an imperfect representation of Euclidean space, it is in some sense superior to a perfect representation, in that it discriminates different episodes that happen in the same location (Wood et al, 2000).
Figure 5
Figure 5. Cells in layer V EC integrate their inputs
Recordings were made from slices bathed in a solution including low concentrations of the cholinergic agonist muscarine. a. Cells from layer V, when presented with a depolarizing input, began firing at a stable rate (epoch 1). As subsequent depolarizing inputs were presented, the cell adopted a new, higher, stable firing rate. c. Analogously, when the cell is at a high firing rate, hyperpolarizing inputs cause the cell to adopt a lower, stable firing rate. b. Shows power spectra for the epochs labeled in a (left) and b (right). From Egorov, et al 2002.
Figure 6
Figure 6. The cellular simulation implements key properties of Eq. 6
A network of 300 integrator cells was prepared. These figures show the firing rate of cell 1 as a function of time. a. At time 0, an input of β was provided sequentially to each of the other cells in the network, one at each time step. The solid line was calculated with β = 0 2. The dashed line was calculated with β = 0 1. The firing rate of cell 1 decays exponentially, with a time constant that depends on the value of β. b. At time steps 0–50 and > 100–200, input was provided sequentially to each cell in the network, one at each time step. At the other times (time steps 50–100 and >200), no input was provided. As in a, cell 1’s firing rate decays exponentially when the other cells are being driven (note the logarithmic scale which makes exponential decay appear linear). However, the decay of cell 1’s firing rate stopped when no input was provided.
Figure 7
Figure 7
A weighted sum over recent movements predicts place-specific coding as a consequence of the kinematic constraints of the enclosure. The set of paths that lead to a position on the Western wall of the enclosure is different from the set of paths that lead to a point on the Eastern wall of the enclosure.
Figure 8
Figure 8. The cellular simulation shows place fields that are topologically similar in similar environments
Four representative cells from the simulation of motion in the open field. Paths were generated using positions and head directions from an experimental session presented in Lever et al 2002. The vast majority of cells showed apparent location-specific firing in both the circular and square enclosures. Like cells in EC described by Quirk et al 1992, the simulated cells showed large irregular place fields with a definite spatial correlate. Further, the simulated cells, like entorhinal cells observed by Quirk et al 1992, showed a high correlation between the place fields observed in the circular environment and in the square environment. The location of the place field is determined by the preferred direction of the input to the cell and the movements taken within the environment. a. Simulated cell 170 (preferred direction East). This cell fires preferentially in the East of the circular and square environments. b. Simulated cell 48 (W) fired preferentially in the Western edge in both environments. c. Simulated cell 20 (SW) fired in the southwest of both environments. d. Simulated cell 75 (NW) fired in the northwest of both environments.
Figure 9
Figure 9. Trajectory coding in the cellular simulation
Average firing rate as a function of position for different trips in the W-maze. The trip each map corresponded to is indicated at the top of each column. Positions the animal did not visit on a particular trip are colored grey. a–d. Simulated cell 64 (preferred direction east) fired preferentially on left-center and center-right trips. The scale bar for this cell is shown to the right of d. e–h. Simulated cell 155 (preferred direction Northwest) fired preferentially on right-center and center-left trips. The scale bar for this cell is shown to the right of h.
Figure 10
Figure 10. The cellular simulation showed retrospective and prospective cells
Firing rate is shown as a function of distance from the end of the center arm for different trips in the W-maze. The label “CP” indicates the location of the choice point defined by Frank et al 2000. Panels a and c show retrospective cells. Panels b andd show prospective cells. a. Simulated cell 30 (preferred direction E) showed a peak in firing just after the choice point for left-center trips. This cell showed differential firing after the choice point, but comparable firing toward the end of the center arm. b. Simulated cell 130 (NNW) showed elevated firing along most of the length of the center arm on both journeys to the right arm and journeys to the left arm. On center-left trips the cell showed elevated firing that began shortly before the choice point. In contrast, the cell showed depressed firing just before the choice point on center-right trips. The curves diverge about 10 cm before the choice point. c. Simulated cell 50 (ESE) showed a peak firing rate shortly before the choice point for left-center paths. Elevated firing lasted most of the length of the center arm. d. Simulated cell 83 (NNE) showed a pattern comparable to cell 130, except the elevation in firing came on center-right trips.
Figure 11
Figure 11
Schematic of the transitive association experiment used by Bunsey and Eichenbaum (1996). In an initial learning phase, animals learned to choose between choice odors (B and Y) depending on which cue odor was presented. The effect of this training was to form an association within each of the pairs, indicated by the arrows. In a second learning phase, the choice odors from the first learning phase became cues used to discriminate between another pair of odors. In a third phase, the animals were tested for their generalization across learning phases. In this transfer phase, animals were given cues from the first stage and choices from the second stage. Animals were tested, in the absence of reward, for their preference of the choice that would result if they formed a transitive association across phases (arrow with question mark). Although animals with hippocampal damage learned as well as controls on each of the learning phases, they were impaired at the transfer stage.
Figure 12
Figure 12. Impairment of new item-to-context learning specifically affects development of transitive associations
Performance as a function of learning is shown for the different stages of the Bunsey and Eichenbaum (1996) study. In all three panels, the intact model, with γ = is shown with solid symbols; the lesioned model, with γ = o simulate hippocampal damage, is shown with open symbols. a. Probability of a correct response during the first phase of learning. Both the intact and lesioned models learned the AB and XY pairs. b. Probability of a correct response during the second phase of training. Both the intact and lesioned model learned BC and YZ. c. Performance on the probe trials, AC and XZ. Probe trials were performed at each stage of learning, but in a way that neither subsequent second phase trials nor subsequent probe trials were affected. While the intact model develops a transitive AC association, the lesioned model does not. This is consistent with the effects of hippocampal lesion observed by Bunsey and Eichenbaum (1996).
Figure 13
Figure 13. Rapid development of a “memory space” in TCM with an intact hippocampus
The network was presented with the double function pairs AB, BC etc in 1,000 different random orders, intermixed with a parallel series of XY, YZ, etc, pairs. Each panel shows the similarity matrix among tAINtBIN etc., at various stages of learning. For instance, the color of cell D, F indicates the value of tDINtFIN Black indicates a value of one, white indicates a value of zero. The top row of matrices is for the “intact” model. The bottom row is for the lesioned model. Parameter values are the same used in Figure 12. The intact and impaired model both start with an orthonormal representation of the tIN’s. This can be seen by the value of one for all the cells along the diagonal and zeros for all off-diagonal cells on the left-most panels. With learning, however, the intact model develops a similarity structure that comes to reflect the “distance” within the double function list. This can be seen by the development of non-zero off-diagonal cells whose magnitude falls off with distance from the diagonal. This is what enables the intact model to “generalize” associations to pairs that were never presented together. In contrast, the lesioned model always evokes the same tIN in response to each stimulus. This prevents the lesioned model from generalizing, although the orthonormal tIN representation may be associated to any presented stimulus to support a forward association.
Figure 14
Figure 14. The stimulus representation of arbitrary stimuli comes to reflect temporal context in the inferior temporal cortex of monkeys
Single units were recorded from area TE of the inferior temporal cortex of macaque monkeys while they performed a delayed match to sample (DMS) task using abstract visual stimuli. The y-axis shows the correlation coefficient calculated for pairs of stimuli. High values of the correlation coefficient mean that neurons tended to fire selectively in response to both stimuli. The correlation coefficient, then, provides a measure of the overlap between the patterns of neural activity corresponding to different stimuli. The cues constituting the DMS task were presented in a fixed order. The filled symbols show the correlation coefficients for stimuli as a function of their distance within a “list” of DMS trials that was presented many times. The open symbols are for an unfamiliar list. The correlation coefficient falls off with distance for the familiar list such that remote pairs are no more correlated than by chance, or for pairs from the new list. For stimuli that were presented many times, the representation of stimuli that were presented in similar temporal contexts becomes more similar. Graph based on data from Miyashita (1988).

Similar articles

See all similar articles

Cited by 89 articles

See all "Cited by" articles

Publication types