Separating hippocampal maps

The place elds of hippocampal cells in old animals sometimes change when an animal is removed from and then returned to an environment Barnes et al The ensemble correlation between two sequential visits to the same environment shows a strong bimodality for old animals near indicative of remapping and greater than indicative of a similar representation between experiences but a strong unimodality for young animals greater than indicative of a similar representation between experiences One explanation for this is the multi map hypothesis in which multiple maps are encoded in the hippocampus old animals may sometimes be returning to the wrong map A theory proposed by Samsonovich and McNaughton suggests that the Barnes et al experiment implies that the maps are pre wired in the CA region of hippocampus Here we o er an alternative explanation in which orthogonalization properties in the dentate gyrus DG region of hippocampus interact with errors in self localization reset of the path integrator on re entry into the environment to produce the bimodality To appear in Spatial Functions of the Hippocampal Formation and the Parietal Cortex edited by Neil Burgess Kathryn Je ery and John O Keefe Oxford University Press Author s current address NSMA University of Arizona Life Sciences North Bldg Rm PO Box Tucson AZ Electronic mail address adr nsma arizona edu Electronic mail address dst cs cmu edu Introduction Reference frames Hippocampal place cells also show correlations to non spatial aspects of the world including en vironment Kubie and Ranck Thompson and Best Muller and Kubie task within environment Markus et al and even stage within task Eichenbaum et al Eichenbaum and Cohen Cohen and Eichenbaum Hampson et al Gothard et al Some have argued that these experiments imply that place cells should be understood as being general context cells with space being only one of many parameters to which they are tuned Eichenbaum et al Wiener Eichenbaum However even when a cell shows di erent ring patterns under two conditions the cell still show place elds under both conditions For example a cell that shows a di erence between two tasks performed within the same environment as reported by Markus et al still only res within a constrained place eld in each task if it res at all To say that a cell is sensitive to a non spatial aspect such as task means that if the cell has a place eld under one condition it may or may not show a place eld under the other and if two cells both show place elds under both conditions then the spatial relationships between them may change drastically from one condition to the other Essentially a cell s place eld under one non spatial condition in fact whether it has a place eld at all is independent of its eld under other non spatial conditions One way to explain this is the multi map hypothesis multiple maps in the hippocampus O Keefe and Nadel active subsets Muller and Kubie reference frames Wan et al a Wan et al b Touretzky and Redish Redish and Touretzky a Redish or charts McNaughton et al Samsonovich and McNaughton Samsonovich All of these authors suggest that the hippocampus includes multiple maps and that each place cell takes part in one or more of those maps There are however minor di erences among these hypotheses In this paper we will concen trate on the two hypotheses which are the most detailed computationally charts and reference frames The multi chart hypothesis suggests that the maps are critically a hippocampal property arising from internal dynamics of the hippocampus In contrast the reference frame hypothe sis suggests that the maps arise from interactions of the hippocampus with extrinsic navigational structures such as a neural path integrator Barnes et al have shown an experiment in which the removal and return of an animal to an environment is su cient to produce a map transition In this paper we present an explanation for this result with simulations Our account says that errors in a path integrator reset process which occur on returning to the environment force a change in reference frame leading to the appearance of a map transition The reference frame theory Over the last few years we have synthesized a theory of rodent navigation bringing together ideas from the extensive work done on rodent navigation over the last century and showing how the interaction of several subsystems gives rise to a comprehensive computational theory of navigation Touretzky and Redish Redish and Touretzky a Redish and Touretzky c A complete description of the theory and its correspondence to the experimental literature is given in depth in Redish Here we will only present a short overview of the theory Then we will Path integration is a process that tracks an animal s position as it moves allowing it to later return to the starting point using only idiothetic cues Barlow Mittelstaedt and Mittelstaedt Gallistel compare its explanation for the Barnes et al experiment with that of the multi chart model McNaughton et al Samsonovich and McNaughton Samsonovich Extending O Keefe and Nadel the reference frame theory describes navigation as a consequence of four di erent functional systems taxon navigation direct approach avoidance of a landmark praxic navigation a sequence of motor actions driven from an internal sequencing mecha nism locale navigation map based navigation and route navigation chained stimulus response mechanisms It also describes locale navigation in detail as a consequence of an interaction among ve spatial representations local view spatial aspects of external landmarks head direction the orientation of the animal in space path integrator the vector home represented on a canonical map place code a representation of the animal s location in the current reference frame and goal memory allowing trajectory planning The anatomical instantiation suggested for these systems is shown in Figure This theory can explain results from a wide range of methodological paradigms including single and multi cell recording behavioral manipulations neuropharmacological manipulations and lesion studies it is consistent with anatomical data see Redish We have simulated most aspects of this theory demonstrating how it can replicate a variety of results including tracking of head direction by cells in postsubiculum and the anterior thalamic nuclei Taube et al Blair et al Taube et al see Redish et al Goodridge et al open eld navigation tasks Collett et al Saksida et al see Touretzky and Redish memory consolidation in the Morris water maze Morris Sutherland and Hoesing see Redish and Touretzky c changes in place elds as a consequence of interactions between a consistent entry point and cue card manipulations Sharp et al see Redish and Touretzky Redish and consequences of disorientation in rectangular arenas Cheng Margules and Gallistel Gallistel see Wan et al a Touretzky and Redish All of the simulations are also detailed in Redish Figure Anatomical realization of a comprehensive model of rodent navigation From Redish Amyg amygdala ATN anterior thalamic nuclei DG dentate gyrus CA CA hip pocampus proper ECs super cial entorhinal cortex ECd deep entorhinal cortex LMN lateral mammillary nuclei mPFC medial prefrontal cortex NAcb nucleus accumbens PaS parasubicu lum PeriRh Perirhinal cortex PoS postsubiculum Sub subiculum VTA ventral tegmental area HD Head direction subsystem pathways PI Path integration subsystem pathways Not all anatomical structures or connections are shown Functional pathways are meant to be indicative only structures not directly on a labeled pathway may also be involved in that subsystem The experiment of Barnes et al Barnes et al allowed an animal to walk around a gure eight maze for minutes They then removed the animal for one hour after which the animal was returned to the maze and allowed to walk around for another minutes During each minute experience Barnes et al recorded about three dozen place cells simultaneously When young animals returned to the environment they used the same set of place cells to encode location But when old animals returned to the environment they sometimes used a completely di erent set of cells The ensemble correlation between place elds in the two experiences was always high for young animals approximately indicative of a similar representation between experiences but was bimodal in old animals sometimes near indicative of a complete remapping other times near indicative of a similar representation Within a single experience in the environment place elds were very stable correlations between the rst and second halves of a single run were always high for both old and young animals Two competing explanations The explanation provided by the multi chart model Barnes et al see also McNaughton et al Samsonovich and McNaughton Sam sonovich explain the bimodal distribution in old animals as a problem in selecting the correct cognitive map Their theory includes a set of pre wired charts in the hippocampus such that the synaptic weight between two cells in hippocampus is inversely proportional to the minimum of their distances across all of the charts When combined with global inhibition this produces a local excitation global inhibition network structure This type of network has a coherent representation of a single location on a single chart as a stable state Samsonovich and McNaughton any other representation such as noise biased by extra activity at candidate locations suggested by sensory cues will be unstable and will settle into a stable state The major drawback of this theory is that it requires complex pre wired connections within the hippocampus Each place cell needs to be more strongly connected to cells with place elds nearby in some chart than to cells with place elds that are distant in all charts There is evidence that this connection structure exists after exploration Wilson and McNaughton but the theory requires that the connection structure be in place before exploration According to this theory on entering a novel environment one location on one chart map will win the competition among competing representations and become the preferred representation for the entry point As young animals explore the environment representations of the local view become bound to places on the currently active chart Then on a return visit to the environment entering at the same point as before the local view representation biases the dynamics in the hippocampus so that the same representation of location on the same chart is reinstantiated In the case of old animals de ciencies in LTP see Barnes for a review prevent the local view from becoming as tightly bound to the currently active chart during initial exploration Thus according to the multi chart model on returning to the environment old animals experience a much weaker bias to select the same location on the same chart as before Figure Hippocampal model used to simulate the Barnes et al experiment During storage solid lines drive place cell activity and dashed lines show correlational learning during recall dashed lines show synaptic transmission and drive place cell activity and path integrator reset See text for details The explanation provided by the reference frame model We propose that the phenomenon seen in older animals is not a consequence of pre wired chart selection within the CA population but rather an interaction between a non linearity of the path integrator and the orthogonalization properties of dentate gyrus The important points for this experiment drawn from the theory described in Section are The path integrator is extrinsic to the hippocampus O Keefe Wan et al b Touretzky and Redish Redish and Touretzky a During normal navigation place cells require both local view and path integrator input O Keefe Wan et al b Touretzky and Redish Redish and Touretzky a The dentate gyrus orthogonalizes the combined local view and path integrator inputs Marr McNaughton and Morris Rolls O Reilly and McClelland Rolls so that if either one changes a new set of place cells is selected Path integrator reset occurs on re entry into an environment Touretzky and Redish Redish and Touretzky a Redish see also Rawlins Rotenberg et al for similar hypotheses Organization of the model The components required for simulating this experiment are shown in Figure The model includes an extrinsic path integrator PI an extrinsic local view LV strong random connections from each to the dentate gyrus DG and strong random connections from the dentate gyrus to hippocampus HC We do not di erentiate between CA and CA in this model and so HC includes both recurrent connections as in CA and outputs to the path integrator as in CA Both the PI and HC models are composed of excitatory E and inhibitory I pools The path integrator in this model is assumed to consist of a two dimensional representation of location in which cells show place elds but the elds do not change from environment to environ ment Cells in entorhinal cortex and subiculum show these environment independent place elds Quirk et al Sharp Following these results we have suggested that the path integra tor consists of a loop between three extra hippocampal structures subiculum parasubiculum and super cial entorhinal cortex Redish and Touretzky a The path integrator representation can be updated by o set connections Zhang Samsonovich and McNaughton In addition we assume that the path integrator has a local excitation global inhibition network structure This means that the path integrator reset process can occur by assuming the path integrator is initialized with noise and then biased by input from the place cells which are in turn biased by the local view This attractor network structure has been extensively studied both in one dimension Wilson and Cowan Amari Ermentrout and Cowan Kishimoto and Amari Kohonen Kohonen Skaggs et al Redish et al Zhang Redish and two Kohonen Kohonen Droulez and Berthoz Munoz et al Arai et al McNaughton et al Zhang Redish and Touretzky c Redish Samsonovich and McNaughton Samsonovich The reference frame model requires that cells in the path integrator be most strongly connected to other cells with nearby place elds which is similar to the pre wired connections required in hippocampus in the multi chart model However the reference frame model only requires this connection structure to pre exist for a single map located outside the hippocampus which simpli es the model immensely The model also includes local excitation within reference frame connections within the hippocampus as does the multi chart model However in the reference frame model this complex connection structure is only assumed to exist after exploration As has been shown by Muller et al b see also Redish and Touretzky c Redish this connection structure can be learned by random exploration combined with correlational LTP i e Hebbian learning Entering a novel environment When an animal is placed in an environment we assume that it does not have preconceived path integrator coordinates The path integrator representation in the model is assumed to initially be random noise Because the animal has not explored the environment yet the learnable connections are assumed to have small uncorrelated random strengths The learnable connections shown by dashed lines in Figure are LV HC HC PI and recurrent connections in HC Because these connections are very weak they do not provide any bias to the settling of the path integrator Therefore the path integrator settles to a representation of random coordinates a hill of activation somewhere on the neural sheet that will serve as the origin or reference point for the new reference frame We call this settling process the self localization or PI reset process In contrast to the learnable connnections the pre wired connections are assumed to be sparse and have strong synaptic weights These connections are indicated by solid lines in Figure LV DG PI DG and DG HC Activity in the dentate gyrus is a consequence of both the LV and PI representations In order for a DG cell to re it must receive input from both LV and PI Early in the self localization process the PI representation is incoherent i e the component neurons show small random ring rates This means that early in the self localization process DG is e ectively silent due to the lack of a coherent representation in PI This allows the LV HC PI pathway to drive the self localization process In contrast during navigation the sparse strong connections passing through DG drive activity in the hippocampus Because both LV and PI ring elds are spatially localized a DG cell will show a high ring rate only in a small compact portion of the environment the place eld of the cell Because most of the possible LV PI combinations do not occur in an environment most DG cells are silent Each HC cell receives input from DG cells Activity in one DG cell is su cient to make the HC cell show a high ring rate HC cells can therefore have varying numbers of place elds depending on their speci c inputs from DG In practice we have found that most HC cells simulated with this model show at most one place eld within a reference frame but occasionally some cells do show two sub elds Cells with multiple sub elds have been reported in real animals e g O Keefe and Nadel Muller et al a Wilson and McNaughton Markus et al As the animal explores the environment LTP occurs along the learnable dashed connections We assume that this LTP is Hebbian and recti ed at so that synaptic strength can only increase We do not model LTD Returning to a familiar environment According to the reference frame theory when young animals return to the environment LTP has created associations between the LV and HC modules and between HC and PI There is a self localization or PI reset process each time the animal enters the environment On a return visit the learned dashed connections that were strengthened by LTP will provide biases to the HC and PI networks Therefore upon reentering the environment the local view representation instantiates a previous representation in hippocampus and this in turn via the HC PI pathway forces the path integrator to reset to the same representation of location as in the young animal s previous experience In old animals however LTP is de cient as reviewed by Barnes and thus there is little or no bias along the LV HC pathway to reset the hippocampus and hence via the HC PI pathway the path integrator to the same location Because each DG cell performs a logical and function of its path integrator and local view inputs if the path integrator representation is not reset correctly this produces a dramatic change in DG representation which will be seen in hippocampus as a low overlap of place codes across di erent visits to the environment E ect of nonlinearity The attractor network structure hypothesized to underlie the path inte grator has an important nonlinearity dependent on where excitatory bias is input into the network There are four important cases depending on the location and magnitude of the extra population input Redish and Touretzky b Redish see also Skaggs et al Elga et al Redish et al Zhang Samsonovich and McNaughton Samsonovich for discussions of speci c cases If an attractor network is in a stable state and receives input synapsing on excitatory cells that is peaked at the same position as is currently being represented then nothing will change The attractor network will still be in a stable state representing the same position The overall activity in the attractor network may increase slightly but the represented position will not change If the input is o set slightly then the attractor network will precess until the new represen tation is centered at the input position If the input is o set by a large amount but is small in magnitude it will not be strong enough to a ect the current representation and so the representation will not change If strong enough input is o set by a large amount the hill of activiation will jump i e the representation of the current position will disappear and activity will reappear at the o set location The e ect of this nonlinearity is that if the bias supplied by the HC PI connections is near the position that the path integrator is settling to it will draw in the representation to match it whether in strong LTP young or weak LTP old animals However if the bias is distant from the position to which the path integrator is settling the representation will jump only if it is su ciently strong i e in young animals but not in old animals Similarities and di erences The interpretations of this experiment o ered by the multi chart and reference frame models have some similarities and some crucial di erences Because place cells are active on initial entry into the environment Hill Austin et al Wilson and McNaughton Tanila et al there must be some pre wired connections producing place eld activity In the multi chart model the charts are pre wired in CA In the reference frame model pre wired connections labeled PI DG LV DG and DG HC produce place cells with stable elds on initial entry into the environment The di erence between the pre wired connections hypothesized by the multi chart model and those hypothesized by the reference frame model is that the latter are initially random Because the place cell instability observed by Barnes et al is bimodal in older animals there must be some sort of nonlinear process occurring during reentry In the chart model this nonlinearity exists in the competitive dynamics between charts in CA In the reference frame model it is found in the nonlinear settling behavior of the path integrator Because the place cell instability observed by Barnes et al only occurs on entry into the environment there must be something special about entry into the environment We explain this by hypothesizing that the path integrator is only reset on entry into the environment During normal navigation the path integrator is not reset it continues to be driven by internal dynamics more than external But on returning to an environment the path integrator is reset and external dynamics can have a strong in uence