Fifty years of neuropsychological memory research. The score: methods 5–theory 2
DOI:10.1093/acprof:oso/9780199228768.003.0015
Abstract and Keywords
This chapter is concerned with changes and developments in neuropsychological memory research during the past fifty years. While researchers in the 1950s generally thought of human memory as a unitary system, present-day researchers believe that there are different, hierarchically arranged, memory systems. Despite some major developments, there are still questions about whether memory for different kinds of information is mediated by distinct parts of the brain and whether the distinct brain systems not only mediate memory for different information, but do so in qualitatively different ways.
Compared with 50 years ago, how differently do we now think about the neural bases of human memory? Obviously, we know much about human memory today that was unknown in 1957 when Scoville and Milner first described HM, whose bilateral hippocampal damage permanently prevented him from acquiring new information, but has our basic theoretical thinking changed? There has certainly been one major change, but others have been more subtle. In the 1950s, researchers thought of memory as a unitary system, whereas today most researchers believe that there are different, hierarchically arranged, memory systems (e.g. conceptual priming 〈 priming 〈 non-declarative memory). Unfortunately, what is meant by ‘memory system’ is not as clear as it might be (see Foster and Jelicic 1999). To me, a memory system is an interconnected set of brain structures that work together to represent, store, and retrieve specific kinds of information. However, although there is little question that memory for different kinds of information is mediated by distinct (partially non-overlapping) parts of the brain, a more interesting, and far less fully resolved, issue is whether the distinct brain systems not only mediate memory for different information, but do so in qualitatively different ways.
Qualitative differences between two or more memory systems seem more likely if the cellular architecture (e.g. local circuitry) of the two systems is distinct and the synapse-altering intraneuronal processes that underlie consolidation and maintenance of long-term memory in distinct systems follow different learning rules. The involvement of different local circuitry and learning rules would suggest that the processes underlying information representation, storage, and retrieval are distinct. Since the early 1980s, parallel distributed processing (PDP)-style neural network modelling (see Rumelhart et al. 1986) has provided a potential means of using the constraints (p.194) imposed by specific regional architectures and learning rules to check exactly how varieties of neural processing, specific to certain brain systems, mediate different kinds of memory. Such modelling also makes it possible to check how different memory systems excite, inhibit, or modulate one another. But, we are still at the early stages of these developments because the effectiveness of this approach depends on the degree to which anatomical, physiological, and other knowledge constrains the models we specify—and those constraints are still fairly minimal.
It is still not known how strong the constraints need to be before computational modelling becomes heuristically worthwhile. As knowledge of relevant constraints improves, model-based predictions will become more interesting. We already know enough about the psychological and brain processes that are necessary and sufficient for each memory system so that rather vague information processing and storage theories can be formulated. To progress to the further stage of building realistic computational models, we need to know still more about several issues. First, the precise functional role of each regional component of each memory system must be identified so that we know what role each region plays in the initial representation of incoming information (encoding), its consolidation and storage, and its retrieval. Second, how the individual regions of each memory system work together in a concerted fashion to allow memory also needs to be identified. Third, we need to know more about the firing patterns of each region's constituent neurones during memory tasks, as well as about their patterns of anatomical connectivity locally and with other regions, and how their synaptic connectivity changes are achieved (which will require isolating the intraneuronal biochemical processes that consolidate and maintain the synaptic alterations underlying long-term memory). Considerable advances towards all of these goals over the past 50 years have been enabled by powerful new techniques.
In my view, apart from the change to thinking explicitly about multiple hierarchically arranged memory systems, the major changes in the past half century have been methodological advances. For example, the idea that memory storage ultimately depends on changes in synaptic connectivity has been around since Hebb (1949). This framework has not changed greatly, but new techniques have filled in many of its details since World War II. So, while we are beginning to get a good preliminary understanding about how neuronal interactions trigger the complex and drawn-out biochemical cascades that lead to changes in neurones' synaptic structure and efficiency, we are still nowhere near formulating a theoretical account of precisely how these changes between neurones are related to the storage of information by different memory systems.
(p.195) New techniques
When the modern era of research on memory neuropsychology began with Scoville and Milner's (1957) paper on the iconic amnesic patient, HM, knowledge of the medial temporal lobe (MTL) location of HM's lesion was based on Scoville's surgical notes. With non-surgical patients, precise lesion locations could be obtained for only the few patients whose brains underwent post-mortem analysis. For most research workers, including myself, this situation continued into the 1990s when structural magnetic resonance imaging (MRI) gradually improved to the point at which good volumetric measures of some key memory structures, such as the hippocampus, became feasible. As I was to discover in the late 1970s, although computed tomography (CT) had been invented earlier in 1972, CT images lacked the spatial resolution and contrast to identify the small diencephalic lesions that characterize patients with Korsakoff syndrome. When two patients with Korsakoff syndrome who had undergone CT earlier came to post-mortem analysis in the 1980s, there was little correspondence between what was found with the two methods (Mayes et al. 1988).
Today, research on amnesic patients depends on sophisticated structural MRI methods, which (like CT) have improved immensely since the mid-1980s. Although still not uncontroversial, automated procedures, such as voxel-based morphometry, make it possible to determine whether a patient's brain damage is relatively focal or more diffuse, spreading to several brain regions. Not only can the volumes of grey and white matter regions be measured, but tractography makes it possible to assess how intact fibre tracts are in patients (and how much they vary in intact brains). As human lesions are adventitious, it has proved helpful to explore cognition by using transcranial magnetic stimulation (TMS) to induce targeted transient disruptions or facilitations of relatively local neural activity. Unfortunately, this has not yet proved very helpful with memory research, because the regions involved often lie deep (as with the midline diencephalon, medial temporal lobes, and basal forebrain structures), and TMS or similar techniques have not yet proved able to affect deep structures selectively.
Despite these major methodological advances which have transformed research on human amnesic patients, research with animal models remains extremely important. This is because there are many manipulations, such as controlled lesions and single-unit recordings, that cannot be done in healthy humans but can be attempted in species with similar neuroanatomy, neurophysiology, and memory. There have also been major advances in methodology. For example, it is now possible to record from many neurones simultaneously (p.196) while an animal is engaged in a memory task. Further, before the development of tractography in 2000, identification of fibre tracts was dependent largely on animal research with model species.
Although single-unit recordings are made only occasionally in human patients, cruder forms of human functional imaging are now possible. Initially, with positron emission tomography (PET) and then with functional MRI (fMRI), in the late 1980s and early 1990s respectively, it became possible to monitor indirect markers of brain activity in humans whilst performing memory tasks. In the 1990s, fMRI of memory, mainly in healthy humans, became more dominant than explorations of memory in amnesic patients. However, a convergent operations methodology is clearly essential because interpretations of each approach's findings are helped greatly by the findings of the other approach. For example, although fMRI may locate all regions comprising a memory system, it does not prove that each such region plays a key causal role in producing the relevant kind of memory. That depends on using lesion or TMS studies. Conversely, many believe that brain damage can lead to brain reorganization, so that memory functions are mediated in ways not seen in people with intact brains. This belief can be checked by fMRI studies of amnesic patients.
How has our knowledge about fact and episode memory's neural bases changed?
When I began research on amnesia in the 1970s, this was the main way of exploring the neural bases of memory, and my research history reflects that of the field. As amnesics have impairments of episodic and semantic memory, I will now focus on what has changed in our knowledge of these kinds of memory.
Although Korsakoff first described the amnesic syndrome in 1889, Scoville and Milner's (1957) description of HM rapidly established several ideas that had previously been only implicit. First, the central role of the MTLs in acquiring new and retrieving long-established episodic and semantic memory became a major concern. Second, continued testing of HM, and later of other amnesics, strongly suggested that the critical damage did not noticeably disrupt other cognitive processes, tapped by tests of intelligence, language, attention, perception, and executive functions. I was drawn into memory research by the idea that memory disorders might be extremely selective, because I felt that lesion studies should have considerable potential as means of exploring memory's neural bases. Third, it also became increasingly apparent that, although amnesic recall and recognition of pre- and post-morbidly acquired episodic and semantic memories was disrupted, other kinds of memory, (p.197) such as that for motor skills, were not impaired (as had been noted anecdo-tally decades before by Claparede). In time, dissociations like this helped lead to the notion of multiple memory systems.
In relation to the first idea, it has long been clear that a connected system of structures plays a key role in episodic and semantic memory. Thus, it has become progressively clearer that damage to the MTLs, midline diencephalon, basal forebrain, or the fibre tracts that connect these regions, such as the fornix, can cause amnesia. This progress has depended not only on improved structural imaging of human amnesics (and on a handful of post-mortem studies), but also on animal studies. However, progress has not been smooth. Unfortunately, for years, MTL lesions in mammals did not produce the severe memory deficits that cases like HM suggested should be present. It was only when animal testing began to use memory tests that were more analogous to the ones on which human amnesics were impaired that the effects of MTL and other lesions in animals began to resemble more closely what was seen in human amnesics. With the development of analogues of object recognition memory testing, it became clear that large MTL lesions in monkeys and other animals consistently caused severe recognition memory deficits, similar to those seen in human amnesics (Mishkin 1978).
A central feature of human amnesia is impaired recall of episodic and semantic memories, but the development of acceptable animal analogues of recall has been a much slower process (although see Eacott et al. 2005). This has meant that progress has been very slow because accidental human lesions are rarely sufficiently focal, and, even if they are, current structural MRI is pushed to its limits in order to identify the damaged structure (e.g. nuclei and tracts within the thalamus). Furthermore, although the localization of animal lesions is much less of a problem, the absence, until recently, of recall tests has led to confusion. For example, monkey lesion studies have suggested that fornix and mammillary body lesions cause only transient recognition deficits (Zola-Morgan et al. 1989); this was interpreted as implying that damage to these structures played no part in amnesia. However, human studies are beginning to make clear that lesions to the fornix or mammillary bodies (to which the fornix projects) may minimally disrupt recognition, but still disrupt recall severely (see Mayes 2001). With colleagues, including John Aggleton and Daniela Montaldi, I have recently been able to confirm this pattern of memory impairment in a large group of patients with relatively selective fornix damage and accompanying mammillary body atrophy Tsivilis et al. (2008).
The second idea, that information can be normally processed and represented at input but that long-term memory for it still be disrupted following specific kinds of brain damage, is puzzling because it seems plausible that information is stored in the same neurones that represent it at input. If this is true, (p.198) selective memory disorders should be found only when a region that modulates activity in other regions is damaged. Such damage might make storage-site function suboptimal, thereby affecting storage ability without noticeably influencing processing efficiency. But the MTLs are not viewed as modulating structures, although the basal forebrain may well modulate MTL and association neocortex activity. It is interesting that several researchers have recently argued that perirhinal cortex damage disrupts high-level visual object perception (Buckley et al.2001; Bussey et al. 2002) and hippocampal damage disrupts high-level kinds of scene perception (Lee et al.2005). These studies suggest that the key lesions have removed vital representation and storage sites for the same information. Nevertheless, even if confirmed, these perceptual deficits are subtle, and not incompatible with the preserved ability of a mixed group of amnesics to retrieve several kinds of information (e.g. object colour, object position, and object semantics) when tested immediately after presentation (Mayes et al. 1993). Furthermore, the putative relationship of perirhinal/hippocampal perceptual deficits to accompanying memory deficits remains to be established and, if the relationship turns out to be weak, the basic puzzle of how selective brain damage causes memory deficits without also disrupting processing at input and retrieval remains unexplained.
Despite these recent developments, theoretical accounts of amnesia have tended to focus on the idea that MTL, midline diencephalic, or basal forebrain lesions disrupt episodic and semantic memory not because they affect initial representation of information, or even some aspect of retrieval, but because they affect consolidation and possibly long-term maintenance of episodic and semantic memories. These possibilities were apparent from the early work on HM around 50 years ago, but there has not been a rapid movement towards a generally accepted storage theory of amnesia and, relatedly, of the neural bases of episodic and semantic memory, for several reasons.
One reason for slow progress is the difficulty of proving that a structure is implicated in storage of specific kinds of information and that damage to that structure disrupts storage so that some information is lost from storage. If removing a structure disrupts memory, this might be because the neural machinery essential for retrieval (and perhaps representation at input) is no longer functional, because the memories have been lost from storage, or for both of these reasons. Simple lesion effects cannot establish which of these possibilities applies. The use of a technique in which a lesion is followed by a stem-cell transplant procedure, which allows the lost tissue to regrow relatively normally, can, in principle, provide a partial answer. If memory is lost following the lesion, but returns once the transplant has grown, the damaged structure (p.199) presumably forms a critical part of the retrieval machinery, but is not a storage site for the memory in question. Conversely, if the memory does not return once the transplant has grown, the site was presumably involved in storage of the lost memory (because the acquired synaptic connections will have been lost), although it may also separately form a critical part of the retrieval machinery. One hippocampal lesion study with monkeys that used this procedure found no evidence that the hippocampus was involved in storage (Virley et al. 1999). However, the lesion was confined to the CA1 field of the hippocampus, so a larger lesion may have produced a different result. In addition, animal research on long-term potentiation (LTP) over the past 30 years does suggest (although not yet conclusively) that the hippocampus is implicated in storage of the kinds of memory that hippocampal lesions disrupt. LTP has long been regarded as a model of the kind of synaptic change that mediates memory storage. It occurs prominently in the hippocampus, and evidence continues to grow that LTP-like changes in fields such as CA1 accompany the kinds of learning supported by the hippocampus (e.g. Whitlock et al. 2006). Nevertheless, there are still some researchers unconvinced that the hippocampus plays any direct role in the storage of episodic and semantic memory (e.g. Gray and McNaughton 2000).
Related to the idea of hierarchically arranged multiple memory systems, perhaps the most important reason why progress has been slow is that the organic amnesia syndrome probably comprises several distinct components, each with its distinctive lesion and pattern of functional breakdown, but these components have proved exceedingly hard to dissociate convincingly. For example, the extent to which anterograde and retrograde amnesia dissociate (with each occurring in relative isolation in different patients) and the factors that determine this are still unresolved. There has also been a long running dispute (at least 30 years) about the factors that determine whether retrograde amnesia relatively spares older premorbid memories or whether it affects them equally, regardless of how far back in the past they were formed. Indeed, there is a major dispute between two positions that assume the two kinds of amnesia have a common neural and functional cause, at least when amnesia results from MTL lesions. The standard view holds that MTL lesions disrupt the initial consolidation of episodes and facts, but that they also disrupt the interaction between the MTL and the association neocortex that would gradually develop stable neocortical storage, which makes the MTLs unnecessary for episodic and semantic memory (e.g. Squire and Alvarez 1995). This view predicts temporally graded retrograde amnesia for both episodic and semantic memories. The main rival hypothesis also postulates the same mechanism for semantic memory and therefore also predicts that older premorbid semantic (p.200)memories are preserved in amnesia. It has, after all, been known for over a hundred years that very well established semantic memories are not affected in organic amnesia, only in dementias, such as the more recently characterized semantic dementia where neocortical damage is more extensive. However, this view differs in postulating that certain aspects of episodic memory remain dependent on the MTLs, gradually becoming more redundantly represented in this region so that only very large MTL lesions cause temporarily ungraded retrograde amnesia for episodic memories (Nadel and Moscovitch 1997).
Resolution of this dispute depends on determining precisely which MTL lesions cause retrograde amnesia, what kinds of memory are affected, and which factors control the duration and severity of the memory loss. It has proved surprisingly difficult to specify both the neural and the psychological factors involved. Good in vivo assessment of the size, location, and functional effects of lesions became available only in the 1990s, so that, assuming that lesion extent and location are the major determinants of the features of the retrograde amnesia, it is not surprising that progress has been slow. Further, animal studies began only around 1990 (e.g. Zola-Morgan and Squire 1990) and, because it was hard to execute studies of retrograde amnesia in animal model systems, progress has again been very slow despite the greater control of lesion extent and location, and of factors such as timing of learning. Useful animal models of episodic and semantic memory recall are only now being developed.
Another reason for slow theoretical progress is that it has proved very difficult to define the boundaries of which kinds of memory are preserved and which are impaired in organic amnesics. This has been particularly difficult in the case of priming, a form of memory that is usually regarded as non-declarative and contrasted with episodic and semantic memory. Priming is revealed by a change in the way that studied information is subsequently processed; such information may be identified more quickly or more accurately. For example, a sentence may be easier to hear in white noise if it has been heard previously than if it has not. Considerable evidence suggests that this memory-related increased processing efficiency occurs whether or not the participant is aware of previously encountering the prime. In other words, effective priming does not require any conscious memory for the remembered information. If, as many people believe, priming is preserved in amnesia, because the same kinds of information that can act as primes (e.g. words, sentences, faces, and objects) can also be recalled and recognized, this suggests that organic amnesia may not disrupt episodic and semantic storage, but rather the process that gives rise to feelings of memory. It is, therefore, crucial to establish (a) whether not merely some, but all, forms of priming are (p.201) preserved after MTL, midline diencephalic, or basal forebrain lesions, and (b) whether priming and recognition/recall really are for the same information. Over 25 years of research have not fully resolved these two issues. For example, it is still unclear whether certain kinds of priming for novel items and associations are impaired in amnesia (e.g. Gooding et al. 2000). It has to be admitted that the area is a minefield because it abounds with tricky control problems. Thus, even in normal subjects, priming effects tend to be weak and performance is variable so that statistical power is low. Conversely there is always a risk that normal subjects may use recall or recognition in an intentional or unintentional way to boost their ‘priming’ scores inappropriately.
Relevance of more recent developments
As we have seen, techniques introduced or elaborated in recent decades now allow much more precise structural and functional measures of the brain to be made in life in humans, and also in animals. As a result, a wider range of evidence from animals as well as humans can be brought to bear on the central theoretical issues related to amnesia and the neural bases of episodic and semantic memory. So, although there is not complete agreement about whether the MTL is a storage site (and, even if it is, only for a while, according to the standard view), it is now very clear from a variety of sources that processed sensory, motor, and semantic information converges on the MTLs where some kind of binding occurs (e.g. Teyler and Discenna 1986). This reflects a fundamental property of episodic and semantic memories: they are associative combinations of more basic components. It is in the MTLs that associative binding occurs. It still needs to be elaborated how this binding relates to the connected regions where lesions also cause amnesia. Thus, although the basal forebrain modulates activity in the MTLs and in several neocortical regions, its precise role in memory is not established. Similarly, the role of midline diencephalic structures remains unclear. The problem is partly that these structures almost certainly have different roles, but they work very closely together with the MTLs as a system in order to mediate episodic and semantic memory.
Identification of this system has been facilitated by functional neuroimaging since the early 1990s, first with PET and later with fMRI. This established that a wide range of neocortical structures, including parts of the parietal, temporal, and frontal lobes, are activated during encoding and retrieval of episodic and factual information. It has actually proved easier to find activations in these regions initially than in the MTLs, midline diencephalon, and basal forebrain. One effect of these neuroimaging findings is to encourage researchers to (p.202) check whether parietal and various frontal lesions cause the disruptions that a simple interpretation of the findings would suggest. Although the existence of some kind of frontal role in episodic and semantic memory (involving the elaboration of encoding, retrieval, and possibly other processes) has been proposed for some time, the effects of lesions in those parietal regions activated in memory neuroimaging studies is only now beginning to be explored (e.g. Simons et al. 2008).
Improved structural and functional techniques are now allowing the exploration of smaller subregions of the brain. This can be illustrated by work on the MTLs over the past decade. Realization that the individual regions of the MTLs, although highly interconnected, may play slightly different roles has been gradually emerging. In the 1980s, human studies did not differentiate the mnemonic role of the hippocampus, perirhinal cortex, entorhinal cortex, and parahippocampal cortex, although it was clear that larger MTL lesions produced more severe amnesia and monkey studies showed that rhinal cortex lesions produced very severe recognition impairments (see Zola-Morgan and Squire 1993). In humans, large MTL lesions led to very severe impairments in the ability rapidly to acquire episodic and semantic memories, whether these memories were tapped by recall or by recognition. However, studies of patients with MRI-confirmed lesions relatively confined to the hippocampus later showed that some of these patients had anterograde amnesias in which item recall and the form of cued recall, known as recollection, were badly impaired, but recognition was relatively intact (e.g. Mayes et al. 2002). Furthermore, familiarity, in which a feeling of memory for previously presented items that is unaccompanied by any recall, was intact in these patients (e.g. Holdstock et al. 2002). Consistent with this, Aggleton and Brown (1999) have argued that the hippocampus is not involved in familiarity (which is mediated by the perirhinal cortex and other structures), whereas it is critical for recollection (which, like familiarity, plays a role in recognition) and recall.
Interestingly, there are other patients with relatively selective hippocampal lesions who show much more severe item recognition deficits and impaired familiarity, as well as recollection (for a review see Mayes et al. 2007). The reason for the discrepancy is currently unexplained. Nevertheless, evidence is accumulating for a division of function between the hippocampus and perirhinal cortex. First, we have found that familiarity feelings for studied objects modulate the activity of the perirhinal cortex, but not that of the hippocampus, as a function of their strengths (Montaldi et al. 2006). Similar fMRI findings have been reported by other groups. Second, if familiarity and recollection really are mediated by distinct parts of the MTL and other brain (p.203) regions, then perirhinal (and perhaps other lesions) should disrupt familiarity, but not recollection. Unfortunately, relatively selective perirhinal cortex lesions are virtually unknown in humans, and, even if this cortex does not mediate recollection directly, it was uncertain what effect lesions would have because the structure projects important inputs to the hippocampus. However, a recent study has shown that a very rare patient, who had undergone resection of part of the left perirhinal cortex, but not the hippocampus, in the treatment of intractable epilepsy, showed intact recollection, but impaired familiarity when familiarity was measured in three different ways (Bowles et al. 2007).
These fMRI and lesion studies support the view that different parts of the MTLs mediate qualitatively different kinds of memory. This view is also supported by a neural network model, which proposes that the hippocampus and perirhinal cortex represent information very differently in memory (Norman and O'Reilly 2003). These two MTL regions have different neural architectures, and the model argues that the hippocampus rapidly and automatically represents even similar inputs in a distinctive way (it pattern-separates), which suits it for pattern completion or recalling memories from partial cues (recollection). In contrast, the perirhinal cortex represents its inputs by identifying common features between even fairly different inputs, which, with rapid learning, suits it best for familiarity memory, but very poorly for recollection.
Since 2000 the emerging picture is that the MTLs are convergence zones for the kinds of information that constitute episodic and semantic memories. This information converges on different MTL structures, some of which also have different neural architectures. Not only is it likely that different MTL regions bind different kinds of information, but it is also likely that, in some cases, they do this in qualitatively different ways. The details are still unresolved, and several views are current. The one that I favour (Mayes et al. 2007) is that after one or two learning exposures the perirhinal cortex binds components into items (e.g. objects, faces, or words), but also binds similar items (e.g. two faces) together without creating new items so as to create a memory representation that mainly supports familiarity. The hippocampus binds any set of components together to make flexible representations that support recollection and recall. In particular, it binds associations between different kinds of information (e.g. faces and voices) that do not converge in other parts of the MTL. As the parahippocampal cortex has a similar structure to the perirhinal cortex, but receives different inputs, one might predict that it will show rapid learning by binding familiarity-supporting representations for whatever informational components converge within it. It has been suggested that these components might fall under the heading of ‘context’ (Eichenbaum et al. 2007).
(p.204) It should not be forgotten that the MTLs mediate episodic and semantic memory by working not only with basal forebrain and midline diencep-halon, but also with parts of the frontal, parietal, temporal, and occipital lobes. We are still a long way from having a fully developed theory of how this happens, and similar problems face all the memory systems that have been postulated.
Conclusion
I have argued that the major changes in research on the neuropsychology of human memory have been developments of research techniques, particularly those that enable exploration in life of brain structure and function. Application of these techniques has led to an extraordinary expansion of detailed knowledge of brain changes related to memory. For example, we now have a much more detailed knowledge about the biochemical, physiological, and structural changes that underlie storage at the neuronal level. If anything, we are less sure than we were decades ago about which psychological functions are unaffected By MTL lesions. Thus, on the basis of new detailed knowledge, some researchers believe that MTL lesions disrupt certain forms of working memory, high-level perceptual processing, and priming. So these functions may not be completely preserved. But, apart from ideas about multiple memory systems, general theoretical thinking about the kind of processing machinery that is needed to support memory has revealed only relatively minor elaborations on ideas that were around 50 years ago. However, this is beginning to change as the constraints that can be applied to neural network models of memory increase. The rich armamentarium of techniques that continues to grow in sophistication will deepen our understanding of the processing machinery at an increasing rate as our knowledge of structure, connections, biochemistry, and the learning rules followed by memory-related brain structures grows. With any luck, the next 50 years should prove to be a very exciting time in research on the neural bases of memory.
No comments:
Post a Comment