The Role of the CA3 Hippocampal Subregion in Spatial Memory: A Process Oriented Behavioral Assessment

Abstract

Computational models, behavioral data, and electrophysiological data suggest that the CA3 subregion of the hippocampus may support multiple mnemonic processes critical to the formation and subsequent retrieval of spatial memories. Multiple researchers have proposed that the CA3 subregion contains an autoassociative network in which synaptic connections between CA3 neurons that represent different components of a memory are strengthened via recurrent collateral connections. As a result, it has been suggested that the CA3 autoassociative network may support multiple processes including the formation of spatial arbitrary associations, temporary maintenance of spatial working memory, and spatial pattern completion. In addition, the CA3 subregion has been suggested to be involved in spatial pattern separation. The separation of patterns is hypothesized to be accomplished based on the low probability that any two CA3 neurons will receive mossy-fiber input synapses from a similar subset of dentate gyrus cells. The separation of patterns also may be enhanced by competitive inhibition within CA3 and dentate gyrus. This review will focus on the mnemonic processes supported by CA3 neurons and how these processes may facilitate the encoding and retrieval of spatial information. Although there is growing evidence indicating that the hippocampus plays a role in the processing of nonspatial information as well, the scope of the present review will focus on the role of the CA3 subregion in spatial memory.

Introduction

Since the early 1960s, the hippocampus has been one of the most studied structures in the mammalian brain. Based on many years of research, the hippocampus has been determined to be highly involved in learning and memory. Although a great deal of progress has been made in understanding how the hippocampus processes information and what types of information the hippocampus may process, there is still debate as to the precise function of this structure. As discussed in a recent publication by Manns and Eichenbaum (2005), early descriptions of the hippocampal formation suggested that information was serially processed through the hippocampal subregions via a trisynaptic loop (Lorente De No´, 1933; Ramon y Cajal, 1995). Information was suggested to enter via entorhinal cortex projections to dentate gyrus with serial projections from the dentate gyrus to the CA3 subregion, which projects serially to the CA1 subregion. The CA1 subregion has projections to the subiculum that in turn has projections back to entorhinal cortex to complete the trisynaptic loop. Based on this idea, a lesion or damage to any anatomical component of the trisynaptic loop would cause the serial processing loop to fail resulting in hippocampal dysfunction. Early studies tested rats with lesions in dentate gyrus, CA3, or CA1 on a working memory version of the radial eight-arm maze. The results demonstrated that a lesion to the dentate gyrus (Emerich and Walsh, 1989; McLamb et al., 1988; Tilson et al., 1987, Walsh et al., 1986), CA1 (Davis et al., 1987; Davis et al., 1986; Davis and Volpe, 1989), or CA3 (Handelmann and Olton, 1981; Jarrard, 1983) all resulted in deficits similar to complete hippocampal lesions. Rats with lesions of dentate gyrus (Nanry et al., 1989; Sutherland et al., 1983), CA1 (Auer et al., 1989; Block, 1999; Nunn et al., 1994; Olsen et al., 1994; Whishaw et al., 1994), or CA3 (Sutherland et al., 1983) also showed deficits comparable to complete hippocampal lesions on the Morris water maze. Therefore, early behavioral studies involving selective lesions to hippocampal subregions suggested that a lesion to any subregion results in a deficit similar to a complete hippocampal lesion. These data could be considered support for the existence of a trisynaptic loop. However, more recent anatomical studies discussed below have demonstrated that the hippocampal anatomical connections are not serial but rather there are projections from entorhinal cortex to each hippocampal subregion (Amaral and Witter, 1995; Witter, 1993). Based on the hippocampal architecture and connectivity, recent models and behavioral studies have demonstrated that the various subregions of the hippocampus may support specific processing functions (Bennett et al., 1994; Gilbert and Kesner, 2003, 2006; Gilbert et al., 2001; Gold and Kesner, 2005; Granger et al., 1996; Guzowski et al., 2004; Hasselmo and McClelland, 1999; Hasselmo et al., 2002; Jensen and Lisman, 1996; Kesner and Hopkins, 2006; Kesner and Rolls, 2001; Kesner et al., 2000; Kesner et al. 2004; Kesner et al., 2005; Lee and Kesner, 2002, 2003, 2004a, 2004b; Lee et al., 2004; Lee et al., 2005b; Leutgeb, et al., 2004; Lisman, 1999; Marr, 1971; McClelland and Goddard, 1996; McNaughton and Morris, 1987; McNaughton and Nadel, 1989; Mizumori et al., 1999; O'Reilly and McClelland, 1994; Rogers et al., 2006; Rolls, 1989, 1996; Rolls and Kesner, 2006; Samsonovich and McNaughton, 1997; Samura and Hattori, 2005; Shapiro and Olton, 1994; Tanila, 1999; Treves, 2004; Treves and Rolls, 1992, 1994; Vazdarjanova and Guzowski, 2004; Wallenstein and Hasselmo, 1997; Wiebe et al., 1997). In addition, recent research has shown that the mnemonic processes supported by each hippocampal subregion can be functionally dissociated using behavioral and electrophysiological measures (Gilbert and Kesner, 2003, 2006; Gilbert et al., 2001; Hunsaker et al., 2006; Jerman et al., 2006; Kesner and Hopkins, 2006; Kesner et al., 2000; Kesner et al., 2004; Lee and Kesner, 2004a, 2004b; Lee et al., 2005a; Lee et al., 2005b; Leutgeb et al., 2004; Rolls and Kesner, 2006; Vazdarjanova and Guzowski, 2004). This review will focus on the mnemonic processes supported by CA3 neurons and how these processes may facilitate the encoding of spatial information. Although there is growing evidence indicating that the hippocampus plays a role in the processing of nonspatial information as well, the scope of the present review will focus on the role of the CA3 subregion in spatial memory.

Basic Hippocampal Anatomy

For a complete review of hippocampal anatomy, please see Amaral and Witter (1995). The main input into the hippocampal system is from entorhinal cortex, which receives inputs from multiple cortical regions and all sensory modalities. The cortical inputs that terminate on the superficial layers (I, II, and III) of the entorhinal cortex comprise the primary inputs to the hippocampus (Witter, 1993). In the rat, the cortical inputs to the superficial layers of entorhinal cortex originate in the olfactory domain of the telencephalon, perirhinal cortex, and pre- and parasubiculum. The entorhinal cortex then projects directly to the dentate gyrus, CA3, and CA1 subregions (Amaral and Witter, 1995; Witter, 1993). Cells in layer II of entorhinal cortex project primarily to the dentate gyrus and also to CA3/2. The projections that terminate in the CA1 region originate in layer III of the entorhinal cortex. The primary projection of the entorhinal cortex is to the DG. The connections between entorhinal cortex, DG, and CA3 are generally reported to be feed-forward (Ishizuka et al., 1990; Witter, 1993). The DG granular neurons project to CA3 pyramidal neurons via mossy fiber projections. The neurons that comprise CA3, in turn, project to CA1 neurons via the Schaffer collaterals. Recurrent collateral connections exist within both the DG and CA3 that serve to interconnect neurons within these respective regions. The DG recurrent pathway includes a layer of excitatory interneurons, the hilus, which interconnects granule cells and a layer of inhibitory interneurons that provide recurrent inhibition. The CA3 subregion has extensive interconnections among the principal cells via a recurrent collateral fiber system (Amaral and Witter, 1995). The primary output from the hippocampus to neocortex originates in CA1 and projects to subiculum, entorhinal cortex, and parahippocampal structures (Witter, 1993). In addition to the projections originating in CA1, projections out of Ammon's horn originate in CA3. Swanson and Cowan (1977) report that the septal region of CA3 projects to dorsal subiculum, parasubiculum, and the cingulate. Many researchers have reported that CA3 projects to the lateral and medial septal nuclei (Amaral and Witter, 1995; Gaykema et al., 1991; Risold and Swanson, 1997). The lateral septum then has projections to the medial septum (Jakab and Leranth, 1995), which in turn projects to subiculum and eventually entorhinal cortex (Amaral and Witter, 1995; Jakab and Leranth, 1995).

One of the most prominent features of the CA3 subregion cytoarchitecture is that there are extensive interconnections among the principal cells via a recurrent collateral fiber system forming an autoassociative network (Amaral and Witter, 1995). Multiple researchers have suggested that the hippocampus, and specifically the CA3 subregion of the hippocampus, contains an autoassociative network that may support a number of mnemonic processes including the formation of arbitrary associations, pattern completion, and working memory (Alvarez and Squire, 1994; Bennett et al., 1994; Bunsey and Eichenbaum, 1993; Granger et al., 1996; Gluck and Myers, 1997; Hasselmo et al., 1996; Jensen and Lisman, 1996; Kesner et al., 2000; Marr, 1971; McClelland and Goddard, 1996; McNaughton and Nadel, 1989; Redish et al., 2001; Rolls, 1996; Rolls and Kesner, 2006; Treves and Rolls, 1992, 1994; Wallenstein and Hasselmo, 1997; Wiebe et al., 1997). The term autoassociative means that synaptic connections between neurons that represent different components of a memory are strengthened. The CA3 subregion of the hippocampus is considered an autoassociative network because of the aforementioned recurrent excitatory connections and synaptic modification among CA3 neurons6 (Rolls and Kesner, 2006). For detailed descriptions of autoassociative networks see Hertz et al. (1991), Rolls and Treves (1998), and Rolls and Deco (2002). CA3 also receives converging inputs from multiple input pathways; for example, perforant path inputs from the entorhinal cortex, mossy fiber inputs from the dentate gyrus, and its own outputs fed back as inputs via the recurrent collaterals (Amaral and Witter, 1995). Most of the synaptic connections embedded in those different pathways in CA3 are modifiable in their strength (Marr, 1971;Treves and Rolls, 1994). The aforementioned anatomical and physiological characteristics inspired many theoretical and computational models to assign specific processes to CA3 (Hasselmo and McClelland, 1999; Hasselmo et al., 2002; Jensen and Lisman, 1996; Kesner et al., 2000; Kesner et al., 2005; Lisman, 1999; Marr, 1971; McNaughton and Nadel, 1989; O'Reilly and McClelland, 1994; O’Reilly and Rudy, 2000; Rolls, 1996; Rolls and Kesner, 2006; Samsonovich and McNaughton, 1997; Shapiro and Olton, 1994; Treves and Rolls, 1992; Treves and Rolls, 1994; Wiebe et al., 1997).

Spatial Arbitrary Associations

The CA3 autoassociative network has been suggested to be responsible for the formation of arbitrary associations or paired associate learning (Bennett et al., 1994; Gilbert and Kesner, 2003; Hasselmo and McClelland, 1999; Kesner et al., 2000; Kesner et al., 2005; McNaughton and Nadel, 1989; Rolls 1996; Rolls and Kesner, 2006; Wallenstein and Hasselmo, 1997; Wiebe et al., 1997). For example, Rolls and Kesner (2006) suggest that information from parietal cortex regarding the location of an object may be associated with information from temporal cortex regarding the identity of the object see also (Rolls, 1996). Therefore, CA3 could enable the organism to remember a particular object and its location. To test this hypothesis, Gilbert and Kesner (2003) trained rats with CA3, CA1, or dentate gyrus lesions on a successive discrimination go/no-go task to examine object-place paired associate learning. This task has been shown to be sensitive to complete hippocampal lesions (Gilbert and Kesner, 2002). In this task, two paired-associates were reinforced that consisted of one particular object (A) in one particular location (1) and a different object (B) in a different location (2). Mispairs that were not reinforced included object (A) in location (2) and object (B) in location (1). Rats should learn that if an object was presented in its paired location then the rat should displace the object to receive a reward (Go). However, the rat should withhold displacing the object if it was not in its paired location (No-Go). In a second task, rats were trained on a successive discrimination go/no-go task to examine odor-place paired-associate learning. In this task, the same procedure was used, except that rats needed to learn that when an odor was presented in its paired location the rat should dig in sand mixed with the odor to receive a reward. The results indicate that relative to controls, rats with CA3 lesions were impaired in learning both the object-place and odor-place paired associations (Gilbert and Kesner, 2003). However, rats with lesions to the dentate gyrus or CA1 hippocampal subregions matched the performance of controls. Thus, the results suggest that the CA3 subregion of the dorsal hippocampus contains a mechanism to support the formation of object-place and odor-place arbitrary associations. A recent study also demonstrated that CA3 lesions disrupt learning of an object-trace-place task where a short delay was implemented between the presentation of the object paired-associate and the spatial location (Hunsaker et al., 2006). Wallenstein et al. (1998) developed a computational model demonstrating that the CA3 recurrent collateral network may be capable of associating temporally discontiguous events.

In support of the CA3 subregion in paired-associate memory, a recent study has shown CA3 N-methyl-D-asparate receptor involvement in associative memory recall (Nakazawa et al., 2002). As electrophysiological evidence for a role of CA3 in forming object-place associations, Rolls et al. (1989) have recorded cells in the monkey hippocampus (CA3 and CA1) that combine information regarding an object and its position in space. A recent study using a viral vector to temporally control a focal deletion of the NR1 gene demonstrated that the learning of novel associations between specific contexts was dependent on CA3 NMDA receptors (Rajji et al., 2006). In this task, rats were presented with two different odors each mixed with sand and placed in small digging cups. The rats were then simultaneously presented with each odor in two different contexts. In one context, the rat was required to dig in the cup containing one of the odors and in the other context the rat was required to dig in the other odor cup. Therefore, to receive a reward, the rat must learn to associate one odor with one context and the other odor with the other context. Deletions of the NR1 gene in the CA3 subregion of dorsal hippocampus were shown to disrupt learning of the odor-context associations but did not affect expression of previously learned associations (Rajji et al., 2006). A previously published study by Gilbert and Kesner (2004) also showed that hippocampal lesions, including the CA3 subregion, do not completely abolish retrieval of previously learned object-place associations but rather result in a 8 transient deficit. Therefore, the CA3 subregion of the hippocampus may be necessary for the formation of novel arbitrary associations but may play only a limited role in the retrieval of previously learned associations.

Although the CA3 subregion may support the formation of novel arbitrary associations involving a spatial component, rats with large lesions to the dorsal and ventral hippocampus, including the CA3 subregion, have been shown to learn a nonspatial object-odor paired associate task as readily as controls (Gilbert and Kesner, 2002). In addition, humans, nonhuman primates, pigeons, and rodents with hippocampal lesions have been reported to display normal learning of nonspatial paired associations involving odor-odor (Bunsey and Eichenbaum, 1995,1996; Li et al., 1999), object-object (Bingman et al., 1999; Cho and Kesner, 1995; Murray et al., 1993; Petrides, 1990), visual-response (Petrides, 1997; Winocur, 1992; Wise and Murray, 1999), and auditory-visual (Jarrard and Davidson, 1990) associations. However, some researchers have reported impaired visual-visual (Ridely et al., 1995; Sutherland et al., 1989) and visual-auditory (Honey et al., 1998) paired-associate learning in animals with hippocampal damage. Therefore, the majority of the existing data suggest that the rodent hippocampus, and particularly the CA3 subregion, is involved in the formation of arbitrary associations when a stimulus must be associated with a spatial location. However, the CA3 subregion of the hippocampus may not be necessary for the formation of all arbitrary associations.

Spatial Working Memory

Many researchers have suggested that the CA3 autoassociative network also may be involved in temporarily maintaining working memory representations (Gilbert and Kesner, 2006; Granger et al., 1996; Jensen and Lisman, 1996; Kesner et al., 2000; Kesner et al., 2004; Kesner and Rolls, 2001; Lee and Kesner, 2002, 2003; Lee et al., 2005b; McClelland and Goddard, 1996, O’Reilly and McClelland, 1994; Rolls, 1996; Rolls and Kesner, 2006; Shapiro and Olton, 1994; Treves and Rolls, 1992; Wiebe et al., 1997). Rolls and Kesner (2006) suggest that the excitatory recurrent collateral network can implement a short-term memory by maintaining the firing of neurons in the recurrent connections (see also Rolls, 1996). It is suggested that a stable short-term attractor can maintain a spatial location across a considerable delay or until a different input pushes the attractor to maintain a new spatial location (Rolls and Kesner, 2006). Wiebe and colleagues (1997) suggest that the dense recurrent connectivity within the CA3 subregion could serve as a “holding memory” that would be capable of maintaining short-term memory representations.

Using a classic paradigm originally developed by Olton et al. (1979), researchers tested rats with CA3 lesions on a task designed to measure both spatial working and reference memory using the radial eight-arm maze (Handelmann and Olton, 1981; Jarrard, 1983). On this task, four arms on the maze always contained a food reward, whereas the remaining four arms never contained food. On each trial, the animal was placed in the center of the maze and was allowed to traverse the arms until all of the food was collected. Repeated entries into reward arms were scored as working memory errors, whereas entries into non-reward arms were considered reference memory errors. The data indicated that rats with CA3 lesions committed significantly more spatial working memory errors compared to controls (Handelmann and Olton, 1981; Jarrard, 1983). A more recent study by Lee and Kesner (2003) examined the effects of CA3 subregional lesions on a delayed nonmatching-to-place task using a radial 8-arm maze. During the sample phase of the task, each rat was allowed to visit a randomly chosen arm and then return to the center platform. Following a 10 s delay, the rat was presented with two adjacent arms. One arm was the same arm visited on the sample phase and the second was a new arm. The rat was required to select the new arm to receive a food reward. CA3 lesions disrupted acquisition of the task with short 10 s delays; however, CA3 was not necessary for task performance with intermediate (5 min) or long (24 hr) delays (Lee and Kesner, 2003). Another recent study conducted by Gilbert and Kesner (2006) tested rats with CA3 lesions on a delayed-match-to-sample for spatial location task. On each trial, an object covered a baited food well in one of fifteen spatial locations along a row of food-wells perpendicular to the start box. Once the rat exited the start box, displaced the object to receive a food reward and then returned to the start box, the same food well was then quickly re-baited, an identical object was positioned to cover the food well and another identical object was positioned in a different location along the row of food wells covering a different unbaited food well. On the ensuing choice phase, the animal was allowed to choose between the two objects. The object that covered the same food well as the object in the sample phase was the correct choice and the second foil object was the incorrect choice. Five spatial separations (15–105 cm) were randomly used to separate the correct object from the foil object during the choice phase. Following surgery, rats with CA3 lesions were significantly impaired relative to controls across all spatial separations, suggesting that CA3 lesions impaired working memory. Chronic glucocorticoid exposure or chronic stress, which results in increased glucocorticoid levels, has been shown to result in significant atrophy of the CA3 pyramidal neurons within the hippocampus (Bodnoff et al., 1995; Kant, 1993; Shors and Dryver, 1992; Stein-Behrens et al., 1994; Watanabe et al., 1992). Similar to neurotoxin lesion studies, rats treated with chronic glucocorticoids or exposed to chronic stress also show impairments in spatial working memory (Bardgett et al., 1994; Luine et al., 1994; Shors and Dryver, 1992; Szuran et al., 1994).

Electrophysiological data also suggest that the CA3 subregions may be involved in working memory. Researchers have recorded units in CA3 that showed a preferential increase or decrease in firing rate during the delay period of a delayed-response task (Cahusac et al., 1989; Colombo et al., 1998; Watanabe and Niki, 1985). Similar findings also have been reported in the CA3 and CA1 regions in rats on a delayed-match-to-position task (Hampson et al., 1993). Using behavioral and electrophysiological techniques, Lee and Kesner (2002) showed that NMDA receptors in the CA3 subregion are involved in the acquisition of spatial working memory. Therefore, multiple studies have provided behavioral and electrophysiological data suggesting that the CA3 hippocampal subregion may maintain short-term working memory spatial representations.

Spatial Pattern Completion

The CA3 autoassociative network has been suggested to support pattern completion. Marr (1971) suggested that the recurrent collateral system in CA3 might be capable of completing stored patterns of information during retrieval based on partial or incomplete input to the hippocampus, termed the collateral effect. More recent models have been generated to describe how the CA3 autoassociative network may support pattern completion (Guzowski et al., 2004; Kesner, et al., 2005; Lee et al., 2004; McNaughton and Nadel, 1989; Mizumori et al., 1989; O’Reilly and McClelland, 1994; O’Reilly and Rudy, 2000; Rolls, 1996; Rolls and Kesner, 2006; Samura and Hattori, 2005; Treves, 2004). Rolls and Kesner (2006) argue that a fundamental property of the CA3 recurrent collateral network is that recall can be symmetric, meaning that the entire memory may be retrieved from any part. Therefore, if an incomplete retrieval cue is given, the autoassociative network is capable of retrieving the original pattern via the recurrent collaterals, resulting in pattern completion.

Until recently, few behavioral or electrophysiological studies had been conducted to examine the role of the CA3 subregion in spatial pattern completion. Gold and Kesner (2005) trained rats on a delayed-matching-to-sample for spatial location task involving systematic manipulations of four available spatial cues to examine spatial pattern completion. During the sample phase, the rats were allowed to displace an object to receive a food reward in one of five spatial locations on a cheeseboard apparatus. On the ensuing choice phase, the rats were required to return to the same spatial location without the aid of the visual object. Therefore, the rat must rely on the spatial cues to locate the same location on the maze. On some trials, all four spatial cues were available. However, on other trials, one, two, three, or all four cues were removed. The removal of the cues used to encode the sample phase location was hypothesized to increase the amount of pattern completion necessary to retrieve the location of the sample phase object on the choice phase. The results indicated that control rats displayed pattern completion across all manipulations of the spatial cues. However, rats with CA3 lesions showed impaired pattern completion as evidenced by a linear increase in errors as the number of available cues was reduced on the choice phase (Gold and Kesner, 2005). A somewhat similar study was conducted by Nakazawa and colleagues (2002) who investigated pattern completion in genetically engineered mice that lacked the NMDA receptor gene in the CA3 subregion. The mutant mice were able to perform a spatial reference memory task in the Morris water maze when all spatial cues were present; however, retrieval of the platform location was impaired when a portion of the extramaze cues were removed.

Lee et al. (2005b) also examined the role of the CA3 subregion in delayed memory for a temporal sequence of spatial locations using a newly devised apparatus termed the “Tulum Maze.” On this task, each rat was required to remember four different spatial sections of the maze sequentially presented during the sample phase. Each section was cued by a unique object that was specifically associated with a location within the section during the sample phase. Then following a brief 15 s delay, the animal was required to revisit the location within a particular section of the maze in the absence of the cued object. Lesions of the CA3 subregion impaired memory for the first three spatial sections visited on the sample phase but not the most recently visited section. Lee and colleagues (2005b) suggest that the task requires spatial pattern completion because the animal must remember a target location without the cued object that marked the location during encoding.

Electrophysiological studies also have provided evidence for the role of the CA3 subregion in spatial pattern completion. Studies have reported that some cells in CA3 respond when a monkey’s view of a particular part of space is briefly obscured by a curtain or darkness (Robertson et al., 1998; Rolls, 1996). This pattern of firing may reflect a completion of the scene in the absence of the visual input. Similar findings also have been reported in rats (Samsonovitch and McNaughton, 1997). In a study conducted by Lee et al. (2004), firing activity of CA3 and CA1 place cells was recorded concurrently in freely moving rats. The animals were placed on a circular track surrounded by both local and distal cues. During the training sessions, the cue configurations were held constant; however, on testing sessions the cue configurations were rotated in opposite directions (counterclockwise and clockwise, respectively) producing varying degrees of mismatch. When the animals were exposed to stable environments, a backward shift was observed across laps in CA1, but not CA3. When cue configurations were altered, both CA3 and CA1 displayed shifting responses; however, the amount of time required to produce a shift varied across sessions. Backward shifts in CA3 place field locations were observed when there was a high degree of mismatch between cue configurations during the first exposure; whereas, CA1 place fields exhibited a delay in shifting response. On the second day of testing the opposite results were observed. CA1 place field locations displayed a backward shift; whereas, CA3 place fields did not demonstrate a significant shift in location. These results suggest that CA3 may flexibly support the rapid acquisition of novel spatial representations. Further, CA3 representations were more coherent than CA1 as degrees of mismatch among environments increased, thus supporting the role of CA3 in pattern completion. In another recent study, Vazdarjanova and Guzowski (2004) allowed rats to explore two environments with each exploration period separated by 30 min. The time course of activations of neuron ensembles in both CA3 and CA1 were measured using a new immediate-early gene-based brain-imaging method (Arc/H1a catFISH). Rats exposed to the same environment during each exploration period showed a high degree of overlap among CA3 and CA1 ensembles. When cue configurations (either local or distal) were altered in similar environments, CA3 ensembles exhibited a greater overlap than those observed in CA1, supporting a pattern completion processes in CA3 (Vazdarjanova and Guzowski, 2004). Therefore, the results of the aforementioned studies provide behavioral and electrophysiological data suggesting that the CA3 hippocampal subregion may support spatial pattern completion.

Spatial Pattern Separation

Based on characteristics of the mossy fiber system, researchers suggest that pattern separation also may be a function of the dentate gyrus and its mossy fiber projections to CA3 (Bischofberger et al., 2006; Gilbert and Kesner, 2006; Gilbert et al., 1998; Gilbert et al., 2001; Kesner and Hopkins, 2006; Kesner et al., 2004; O’Reilly and McClelland, 1994; Rolls and Kesner, 2006; Rolls, 1989, 1996; Shapiro and Olton, 1994; Treves and Rolls, 1992). Pattern separation is described as a mechanism for separating partially overlapping patterns of activation so that one pattern may be retrieved as separate from other patterns. Pattern separation is facilitated by sparse connections in the mossy-fiber system, which connects dentate gyrus granular cells to CA3 pyramidal neurons (Rolls, 1989, 1996). The separation of patterns is accomplished based on the low probability that any two CA3 neurons will receive mossy-fiber input synapses from a similar subset of dentate gyrus cells. The mossy-fiber inputs to CA3 from dentate gyrus are suggested to be essential during learning and may influence which CA3 neurons will fire based on the distributed activity in the dentate gyrus (Rolls and Kesner, 2006). The cells of the dentate gyrus are suggested to act as a competitive learning network with Hebb-like modifiability to reduce redundancy and produce sparse, orthogonal outputs. O'Reilly and McClelland (1994) suggest that the separation of patterns is enhanced by lateral inhibition within CA3. This assumption is described as a competitive "k-Winner-take-all" (kWTA) activation function. The lateral inhibition and the sparse connections are proposed to function contemporaneously to assure that an approximately constant (k) number of neurons in CA3 that receive the excitatory input become active. These "k" neurons are described as the neurons that have an excitatory level that exceeds the inhibitory input level of the inhibitory interneurons. The model suggests that the entorhinal cortex to dentate gyrus perforant pathway may dominate during encoding and hence, enhance pattern separation. In contrast, the direct one-stage pathway from entorhinal cortex to CA3 may be the preferential processing pathway during retrieval (O’Reilly and McClelland, 1994). In support of the role of the dentate gyrus in encoding, a recent study conducted in mice showed that the mossy-fibers projections to CA3 are essential for the encoding of spatial information but are not necessary for retrieval (Lassalle et al., 2000). Lee and Kesner (2004a,b) also have shown that the mossy fiber input to CA3 is critically involved in encoding of spatial information in rats, but may not be involved in retrieval. The study also showed that the perforant path input to CA3 from entorhinal cortex may support retrieval of spatial information, but is not necessary for encoding (Lee & Kesner, 2004a,b).

Rolls (1996) notes multiple characteristics of the mossy-fiber projection system that may promote pattern separation in the mossy fiber system and hence, efficient information storage in CA3. First, mossy-fiber synapses are very large and terminate close to the soma of the CA3 pyramidal neurons in the pyramidal layer. Therefore, the mossy-fiber synapses will be relatively powerful in activating the CA3 cell. The projections from layer II of entorhinal cortex to CA3 terminate in the lacunosum moleculare layer of the CA3 pyramidal cells, which is much further from the soma. Second, the firing activity of granule cells within the dentate gyrus is sparse (Jung and McNaughton, 1993) and coupled with the small number of connections to CA3 cells may produce a sparse signal that may be transformed into an even sparser signal in CA3. It also has been demonstrated that the place fields of dentate gyrus cells (Mizumori et al., 1990) and specifically granular cells (Jung and McNaughton, 1992) are small and highly reliable which may support the role of dentate gyrus in pattern separation. In addition, the mossy-fibers, demonstrate nonassociative plasticity (Brown et al., 1989), which may enhance the signal-to-noise ratio such that the mossy-fiber cell would produce a nonlinearly amplified current in the CA3 cell. Due to this type of plasticity in the dentate, a particular stimulus such as spatial location, across subsequent presentations would be likely to activate the same population of CA3 neurons, which may result in economical storage because one CA3 cell could be used in different memories (Rolls, 1996).

Behavioral studies have suggested that the dentate gyrus may support spatial pattern separation (Gilbert et al., 2001); however, until recently no behavioral studies have examined the role of the CA3 subregion in spatial pattern separation. A recent study (Gilbert and Kesner, 2006) tested rats with CA3 lesions on a delayed-match-to-sample pattern separation task developed by Gilbert and colleagues (Gilbert et al., 1998, 2001). On each trial, an object covered a baited food well in one of fifteen spatial locations along a row of food-wells perpendicular to the start box. Once the rat exited the start box, displaced the object to receive a food reward and returned to the start box, the same food well was then quickly re-baited, an identical object was positioned to cover the food well and another identical object was positioned in a different location along the row of food wells covering a different unbaited food well. On the ensuing choice phase, the animal was allowed to choose between the two objects. The object that covered the same food well as the object in the sample phase was the correct choice and the 15 second foil object was the incorrect choice. Five spatial separations (15–105 cm) were randomly used to separate the correct object from the foil object during the choice phase. Following surgery rats with CA3 lesions were significantly impaired relative to controls across all spatial separations suggesting that CA3 lesions may impair spatial pattern separation as well as spatial working memory (Gilbert and Kesner, 2006). A study by Tanila (1999) offered cellular recording evidence for the existence of a pattern separation mechanism for generating separate representations in CA3 based on overlapping inputs. In this study, neuronal ensembles were recorded while rats explored two visually identical rectangular environments connected by a hidden door and received electrical stimulation of the lateral hypothalamus as reinforcement for visiting particular locations in the environment. During the first recording session, the rat was allowed to explore the first environment. Then the door to the second environment was opened allowing the rat to enter the second environment for a 5 min recording session. The rat was then placed in a bucket for a 5 min period before another recording session in the second environment followed by a recording session in the first environment. Finally, the rat was allowed to re-explore the second environment following a 90-degree rotation of the environment relative to its original position in the room. It was shown that CA3 place cells were able to maintain distinct representations of two visually identical environments and selectively reactivate either one of the representation patterns depending on the experience of the rat. These data provide further evidence for the role of dentate gyrus and CA3 in spatial pattern separation. In addition, McNaughton et al. (1989) have shown that following colchicine-induced lesions of the dentate gyrus, there is a significant decrease in reliability in CA3 and CA1firing. Therefore, if the CA3 and CA1cells display less reliability following dentate gyrus lesions, the cells may not be forming accurate representations of space due to decreased efficiency in pattern separation.

In a more recent study, Leutgeb et al. (2004) conducted cellular recordings in CA3 and CA1 while rodents were exposed to various contextual conditions. Animals were placed in geometric shaped enclosures with varying degrees of similarity and were allowed to explore for food. The enclosures were distributed into three different rooms. The amount of overlap in CA1 varied as a function of the degree of geometric similarity. As similarity increased, the amount of overlap in activation patterns increased. Conversely, as contexts became less similar, the amount of overlap decreased. In contrast, CA3 place cells were able to maintain separate patterns of activation regardless of the degree of geometric similarity. The firing characteristics in CA3 resembled those in CA1; however, the firing was sparse in CA3, suggesting that CA3 may support pattern separation or the orthogonalization of similar activity patterns into distinct representations.

Based on the aforementioned electrophysiological data, it appears that the CA3 subregion of the hippocampus may support both pattern separation (Leutgeb et al., 2004) and pattern completion as mentioned in the previous section of this review (Lee et al., 2004). The results of these two studies may be viewed as contradictory and lead to the question of how the CA3 subregion could support two disparate functions (see Guzowski et al., 2004 for a discussion). A resolution to this apparent disparity has been proposed by Vazdarjona and Guzowski (2004). The time course of activations of neuron ensembles in both CA3 and CA1 were measured using a new immediate-early gene-based brain-imaging method (Arc/H1a catFISH). According to Vazdarjona and Guzowski (2004), CA3 appears to support different processing functions, pattern completion and pattern separation, depending on manipulations of the environment. In the task, rats were allowed to explore two environments with each exploration period separated by approximately 30 min. Rats exposed the same environment during each exploration period showed a high degree of overlap among CA3 and CA1 ensembles. When cue configurations (either local or distal) were altered in similar environments, CA3 ensembles exhibited a greater overlap than those observed in CA1, suggesting that CA3 may support pattern completion as discussed previously. However, when exposed to alterations to both local and distal cues in two completely different environments, CA3 ensembles exhibited a lower overlap than those observed in CA1, suggesting that CA3 also may support a pattern separation. Coupled with the findings from Leutgeb and colleagues (2004) and Lee and colleagues (2004), these studies suggest that CA3 may flexibly shift between pattern completion and pattern separation in a competitive and discontinuous manner depending on environmental conditions and task demands (Guzowski et al., 2004). Since attractor networks, such as CA3, respond nonlinearly to input patterns (Rolls and Treves, 1998), pattern completion may occur with minimal changes to sensory inputs; however, when changes to sensory inputs are increased, pattern separation may occur (Guzowski et al., 2004). This may be accomplished by an inherent threshold mechanism existing in the CA3 subregion (Kesner et al., 2004). When input patterns are dissimilar enough, CA3 may support pattern separation. However, CA3 may instead support pattern completion when input patterns are not sufficient enough to cross the threshold. O’Reilly and McClelland (1994) proposed a computational model of hippocampal function demonstrating how the hippocampus may be capable of supporting both processes and how the structure may minimize the trade-off between pattern completion and pattern separation. The competition between pattern separation and pattern completion has been modeled as a sigmoidal function (McClelland and Goddard, 1996; O’Reilly and McClelland, 1994). Guzowski et al. (2004) suggest that one possible role of place cell representations is to learn different adaptive responses to the same stimuli based on different contextual information. As reviewed by Knierim (2003), place cells can remap in a familiar environment with sufficient alterations in the environment or changes in the required behavior of the animal. Guzowski et al. (2004) suggest that this remapping may reflect pattern separation and may allow the animal to generate independent representations of behavioral contingencies in a particular context to reduce the chance of engaging in an inappropriate behavioral response. However, if every small change in an environment resulted in remapping and the creation of a new representation, pattern separation may be disruptive to CA3 function (Guzowski et al., 2004). Therefore, pattern separation may be tempered by pattern completion to reconstruct the original representation from a degraded input while ignoring small irrelevant alterations in the context (Guzowski et al., 2004). This multiple processing capability of CA3 may provide an efficient means for storing and subsequently retrieving memory representations.

Summary

Based on the hippocampal architecture and connectivity, behavioral and electrophysiological studies have demonstrated that the various subregions of the hippocampus may support specific processing functions. In addition, recent research has shown that the mnemonic processes supported by each hippocampal subregion can be functionally dissociated. Computational models, behavioral data, and electrophysiological data suggest that the CA3 subregion of the hippocampus may support multiple mnemonic processes critical to the formation of accurate spatial memories. It has been suggested that the CA3 recurrent collateral network may support the formation of spatial arbitrary associations, temporary maintenance of spatial working memory, and spatial pattern completion. In addition, based on the low probability that any two CA3 neurons will receive mossy-fiber input synapses from a similar subset of dentate gyrus cells and competitive inhibition, the CA3 subregion has been suggested to be involved in spatial pattern separation. Thus, the CA3 subregion may facilitate the accurate encoding and subsequent retrieval of spatial memories by supporting mnemonic processes to 1) associate a stimulus with a spatial location, 2) separate partially overlapping patterns of activation so that one spatial location can be retrieved as separate from other spatial locations, and 3) retrieve a spatial memory based on partial or incomplete input.