There has been a great deal in the literature
about the hippocampus
, its anatomy
, its physiology
, and its role in memory
. But the general consensus seems to be that it has some kind of associative role in memory and learning
, particularly for declarative
and episodic memories
As with many aspects of memory, interest in the hippocampus was sparked by looking at human amnesic patients. Scoville and Milner (1957) were the first to report the much-studied amnesic H.M., who developed severe anterograde and temporally graded retrograde amnesia after the complete bilateral removal of his medial temporal lobes in order to prevent severe epilepsy. This lesion included the entire hippocampal formation, and Scoville and Milner postulated that the hippocampus was somehow the key factor in H.M.'s memory impairment.
However, the extent of H.M's lesion makes the inference of the cause if his amnesia difficult. Patient R.B. has managed to aid in implicating the hippocampus as a key area in human memory (Zola-Morgan, Squire, and Amaral, 1986). R.B. suffered an ischemic event during heart surgery, and the lack of blood flow to his brain caused bilateral damage confined to the CA1 field of his hippocampus, with very limited, and non-significant, pathology in other brain regions (the CA1 field of the hippocampus is the most sensitive area of the brain to a lack of blood flow). Thenceforth, R.B. suffered a moderately severe memory impairment. Victor and Agammolis (1990) have also reported a similar case of a patient with a memory impairment following a bilateral hippocampal lesion. Further implication of the hippocampus playing an important role in human memory has come from studying patients with memory impairments with MRI (magnetic resonance imaging), showing that they have a shrunken and atrophic hippocampus, about 57% of its normal size (Press, Amaral, and Squire, 1989).
However, human studies are limited in what they can tell us about the role of the hippocampal formation in memory. Animal studies allow a greater degree of experimenter control, and have been very illuminating. Monkeys have often been studied due to the similarity of their brains to the human brain. Obviously, monkeys cannot be asked directly about declarative and episodic memories the way humans can, but there are some tasks that monkeys with lesions of the hippocampal formation fail that are also failed by human amnesic patients. Zola-Morgan and colleagues have found this across many different tasks, such as retaining easy object discriminations, eight-pair concurrent discrimination learning, and delayed responses to stimuli. And, like amnesic humans, amnesic monkeys are still able to acquire skills and habits (Mishkin et al., 1984; Zola-Morgan and Squire, 1984). This is due to implicit learning being mediated by a separate system to declarative explicit memory. However, an important finding from experimental animal studies is that although partial hippocampal damage produces significant memory impairments, the more extensive the lesion, the larger the scale of the memory impairment. A similar effect is seen in humans.
The hippocampus is surrounded by several different structures - the amygdala, the parahippocampal cortex, the perirhinal cortex, and the entorhinal cortex. These are the areas that appear to all have roles in memory, since lesions to any of these areas cause impairments. It used to be thought that memory impairments such as those associated with classical human amnesia were caused by combined hippocampal and amygdaloid lesions. The amygdala was thought to have a contributory effect to memory processing, until it was realised that surgical amygdala lesions were also causing damage to the cortex underlying the amygdala - which includes the perirhinal cortex, which has been shown to have important memory functions (see Buckley and Gaffan, 2000). Some elegant lesion studies by Zola-Morgan and co-workers have shed light on the memory contributions of the various structures surrounding the hippocampus. Firstly, lesions confined to the amygdala alone were shown to have no effects at all on memory tasks (the amygdala is involved in memory, though, but emotional memory, such as in classical conditioning (Davis, 1986; LeDoux, 1987)). It was also found that monkeys with lesions of the hippocampus and surrounding cortex (the "H+" lesion) were not as impaired as similar lesions which in addition include the amygdala and surrounding cortex (the "H+A+" lesion) on memory tasks. However, monkeys with the H+ lesion with lesions of just the amygdala, or the "A" lesion - thus making the "H+A" lesion - where no more impaired than monkeys with just the H+ lesion. This nicely demonstrates that the amygdala has no major contribution to declarative memory.
Furthermore, it was demonstrated that the H++ lesion (the hippocampus, perirhinal cortex, and anterior entorhinal cortex) produced a similar impairment to the H+A+ lesion. This confirms that the perirhinal and entorhinal cortex does have a memory contribution. Further evidence for this comes from similar impairments to those caused by the H++ and H+A+ lesions being obtained with the PRPH lesion - which consisted of the perirhinal and hippocampal cortex, including the white matter which connects the neocortex to the hippocampus.
The conclusion of these lesion studies was that combined hippocampal and amygdaloid damage is not the cause of declarative memory impairment. Moreover, the cortex surrounding the hippocampus is not just a conduit to funnel information from the neocortex into the hippocampus. The PRPH lesion produced a greater memory impairment than lesions of just the hippocampus alone. These data seem to rather downplay the role of the hippocampus itself, but it does remain a crucial part of the memory system and has many important functions.
Aside from monkey studies, studies using rats have provided useful information about what exactly the hippocampus does. One might think that studying the rat hippocampus might not provide information which is that compatible with the human hippocampus, but across mammals the hippocampus is always very similar.
There have been several theories to emerge out of rat studies. One of the most prominent is to do with the hippocampus having a spatial function. O'Keefe and Nadel (1979) formulated the idea that the hippocampus represents an allocentric "cognitive map" of space. This is based on three classifications of cells recorded from in the rat hippocampus that respond to various spatial elements. "Displace cells", or "theta cells", were found to fire when a rat moves about, "place cells" were found to only fire when a rat was in a certain location, and "misplace cells" which fired when a rat finds something missing in the environment, or in the wrong location. They found that non-spatial cells were few in number - almost all the hippocampal cells recorded from were all responsive to spatial things. Rats with hippocampal lesions perform badly in maze tasks, such as the Morris water maze in which rats must learn to swim to the location of a platform just below the surface of a pool of cloudy water. Hippocampectomized rats are unable to learn the location of the platform, whereas control rats manage to swim almost directly to it. This appears to be a deficit involving spatial awareness and orientation, because if the hippocamectomized rats are placed in the pool from the same starting point each time, they do learn to swim to the platform as this no longer relies on using the spatial features of the environment to direct movement.
The idea of cognitive map theory is that the hippocampus is used to build up a spatial map of the environment through experience. This idea seems to correlate well with the finding that in birds that store food for later retrieval have larger hippocampuses that birds which do not exhibit this behaviour, yet are close relatives (Healy and Krebs, 1993). In addition, Maguire et al. (2000) have found that the hippocampuses of London taxi drivers are significantly larger than normal, and are larger as a function of how long they have been taxi drivers. There is also activation in their hippocampuses as they figure out routes. The spatial element to the hippocampus seems to be consistent with the fact that human amnesic patients do in fact lose their way when they are trying to navigate to places and cannot remember spatial layouts. However, amnesic humans also forget passages of prose, tactual impressions, odours, faces, and melodies, among many other things. Also, hippocampectomized rats are impaired in tasks which have no real spatial component, such as learning odour discriminations (Eichenbaum et al., 1988), timing tasks (Meck et al., 1989), and configural discriminations that involve unique combinations of visual stimuli (Rudy and Sutherland, 1989). This evidence seems to support the view that the spatial element is just one function of the hippocampus (Squire, 1979).
Hippocampal lesions in monkeys are impaired in performing a delayed non-match to sample task, and this impairment increases as the delay between the sample and matching stimuli is increased. This impairment is greater for parahippocampal lesions. Single cell recording during the delayed non-match to sample task has revealed mechanisms in the hippocampus and parahippocampal cortex that seems to be memory related. Some cells have been found to fire selectively to the sample stimuli and act to encode this stimulus. Some continue to fire in a stimulus-specific way during the delay period where the sample must be held in memory. These cells allow a representation of the sample to endure. There are a third sort of cells that show enhanced or suppressed responses to the familiar stimuli when they reappear, which indicates they are involved in the match/non-match judgements (see Eichenbaum, 2000). This is useful because the neocortical areas where the perceptual information is processed is only capable holding representations for a short time. The parahippocampal cortex is able to maintain this representation in the neocortex due to projections to neocortical areas.
The main function of the hippocampus seems to be in forming associations and conjunctions between normally unrelated events (Squire, Shimamura, and Amaral, 1989). A possible mechanism for doing this is long-term potentiation, or LTP (Bliss and Lomo, 1973), which can associate convergent inputs through a rapid and long lasting synaptic plasticity. So is the hippocampus a storage site for memories? The answer is "no". Amnesic patients with hippocampal damage suffer not only from anterograde amnesia, but also temporally graded retrograde amnesia. This condition is where the patient remembers nothing for about a year or two prior to the onset of their condition, yet, they can still remember without deficit events and facts from further back in time. Childhood memories are particularly robust. So, it would seem that the hippocampus cannot be the storage site of memories. Instead, the hippocampus is needed to form declarative memories. This idea is further supported by the fact that after recovery from transient global anterograde amnesia, a person will have normal memory function before and after the amnesic episode, but have a blank memory for the time frame during which the episode endured.
The evidence points towards the conclusion that the hippocampus and its related structures have a temporary role in memory. Memories initially depend on the hippocampus but become independent over time as associations are repeated between neocortical areas. Eventually these associations become strong enough between the neocortical areas so that there are direct connections and the hippocampal formation is no longer necessary to provide access to these associations (Eichenbaum, 2000).
So in conclusion, the hippocampus seems to be an associative device which is capable of associating inputs together from across all the sensory modalities, plus it is able to provide semantic, temporal, and spatial contexts for these associations. In these associations are all the ingredients required for declarative and episodic memories. But the importance of the hippocampus itself to declarative memory formation should not be overstated. The surrounding cortical areas, such as the perirhinal and entorhinal cortices, and the parahippocampal cortex are also very important to achieve declarative memory, and without these structures, declarative memories cannot be formed.
It has been tricky trying to pin down what the hippocampus does, but now that there seems to be a clearer answer to this question, the next step is identifying how it does it. The hippocampus appears to play a key role in forming episodic memories by associating together various sensory inputs. Rolls and Treves have proposed a neural network model of hippocampal function, based on neural principles, which performs this task. This sort of model has taught us much about how the hippocampus might function, and also, what each part of it might do.
Rolls and Treves' model is based around three main stages, each corresponding to the three main areas of the hippocampus - the dentate granule cells, area CA3, and area CA1. The first stage of processing in the model corresponds to the dentate granule cells. It is proposed that these dentate granule cells act in a way to provide a sparse and redundancy-free input to the CA3 region. The dentate gyrus takes input from entorhinal cortex, which itself has inputs coming from all over the neocortex - generally from the end of each cortical processing stream. Redundancy is removed by the dentate granule cells from its input by way of a competitive learning network based on Hebb-like processes. This means that only the most active neurons are allowed to succeed in providing input to the CA3 region due to the competitive feedback inhibition. An advantage of this competitive learning is that configural learning (such as that discussed by Sutherland and Rudy) can be performed. Overlapping inputs to the hippocampus can be separated, so the system can memorise A, B, and A+B in combination. If there were no configural learning, these three things would all be amalgamated into one memory. The sparseness of the representations that are produced as input to area CA3 allow more memories to be stored in the CA3 autoassociative network.
The CA3 network is, in many respects, the key component of Rolls and Treves' neural network model of memory. It is this autoassociative network that functions to actually store memories. The model has shown that memories can be stored as patterns of neural activity in an autoassociative network. The network is built around recurrent collateral connections where the output of every neuron in the network is fed back to itself and all the other neurons. Modifiable connection weights allow patterns of firing in the network to be associated together, and also each neuron can be associated with itself. The activity of one neuron can then be used to activate other neurons that have previously been associated with it. This is a key aspect of the model as it allows whole episodic memories to be recalled from just a fragment (this is called "completion"). The model network exhibits "one-shot" learning, thus reflecting the very rapid learning that episodic memories perform (we do not need to experience the same event repeatedly to remember it). Another important aspect of the autoassociative system is that is keeps memory representations separate.
Mossy fibre (a type of neural pathway that looks "mossy") input is important to the functioning of the CA3 network as it influences which neurons fire in response to the patterns of activity in the dentate granule cells. The perforant path (another type of hippocampal neural pathway) in the model would seem to initiate retrieval as it acts to relay the cues to the CA3 network that are used in completion. The roles of the mossy fibres and the perforant path are something that has been inferred from the construction of this model and predictions can be made based on their inactivation. If the mossy fibres are inactivated, then amnesia specific to events occurring during the window of inactivation should be produced, but there should be no problems recalling memories formed before inactivation and after the inactivation period has ended. However, if the perforant path is inactivated, then a deficit in the retrieval of memories still stored in the hippocampus should ensue. These predictions appear to be consistent with neurophysiological evidence, are computationally plausible, and have been obtained in computer simulations. These effects are a good example of what neural network models have taught us about hippocampal function.
The CA1 field has been shown to perform an important function through Rolls and Treves' work. The CA3 region associates parts of a particular episode, say A, B, C, D, and E. Each of these parts produces its own pattern of activity in CA3 which is linked together by autoassociation. It is suggested that CA1, which receives from CA3 the inputs representing A, B, C, D, and E simultaneously, acts to separate these parts of the memory through a competitive learning network. There may also be neurons in CA1 that act to represent clusters of these component parts to memories, such as A+B+C, or, B+D+E, to improve efficiency of recollection. There are modifiable connections between CA3 and CA1 that allow there to be no information loss between these two stages. However, information has already been lost in the CA3 network, so CA1 may well receive additional inputs from the perforant path, providing information in a more complete form to help in the accurate and full recollection of memory components. The CA1 region contains a much greater number of neurons than the CA3 region. CA3 is a "bottleneck" used to associate, while CA1 is a network used to redistribute information before it is fed into back projections to the neocortex. The greater number of CA1 neurons allows less information to be loaded onto each neuron, therefore creating a code more robust to further processing. The role of the CA1 cells is to activate entorhinal cells which activate back projections to neocortical areas. The function of these back projections is to reinstate the original patterns of cell firing the neocortex that served to form the memory in the first place. This can be used to initiate and guide action, and also help to incorporate memories into long term storage.
A controversial idea of Rolls and Treves' modeal is that the hippocampus functions as a short term store of memory (not short term memory itself, which is a different thing - i.e., "working memory"). It has limited capacity dictated by the number of synapses on the dendrites of each CA3 cell. The neural network model thus infers it must have an overwriting mechanism. Memories are eventually hippocampus-independent, as shown in human cases of hippocampal (and medial temporal lobe) damage, such as H.M. The temporal gradient of the retrograde amnesia in such cases can suggest a time limit of hippocampal storage. Rolls and Treves suggest that this must be a function of the number of memories formed. So, if an animal lives in a very constant and deprived environment then their hippocampal memory store will remain constant, as a very small number of new episodic memories will be formed. Also, the model provides an account for memory problems following the decrease in the number of synapses per CA3 cell that could come about from ageing, or even a degenerative disease (e.g., Alzheimer's disease). It is thought that memory problems in such scenarios would be the result of decreased hippocampal storage capacity.
An important thing that the neural network models have told us is how memory can be processed using various types of network arrangements. For instance, Rolls and Treves (1994) have run simulations of their model and found that the recurrent collateral connections in CA3 are very important, because if they are inactivated, poor memory recollection is the result. By basing models on the neurophysiology of the hippocampal system, neural network models have shown how the connections between neurons might work to represent memories, and mediate their storage and retrieval. They have also improved the understanding of what the various parts of the hippocampus does.
Squire, L.R. (1991) Memory and the hippocampus: a synthesis of findings with rats, monkeys, and humans. Psychol Rev 99:195-231
Eichenbaum, H. (2000) A cortical-hippocampal system for declarative memory. Nat Rev Neurosci; 1: 41-50.
Buckley, M.J. and Gaffan (2000) The hippocampus, perirhinal cortex and memory in the monkey. In (Ed JJ Bolhuis) Brain, Perception and Memory. OUP.
Sutherkland, R.J., Rudy, J.W. (1989) Configural association theory: the role of the hippocampal formation in learning, memory, and amnesia. Psychobiol 17:129-144
Rolls, E.T., and Treves, A. (1998) Neural Networks and Brain Function, OUP