Brain, Vol. 124, No. 4, 647-675,
April 2001
© 2001 Oxford University Press
Invited review |
Semantic dementia: relevance to connectionist models of long-term memory
1 Department of Psychology, University of Amsterdam, The Netherlands, 2 Medical Research Council—Cognition and Brain Sciences Unit, Cambridge and 3 Neurology Unit, Addenbrooke's Hospital, University of Cambridge, Cambridge, UK
Correspondence to:
Dr Kim S. Graham, Medical Research Council—Cognition and Brain Sciences Unit, 15 Chaucer Road, Cambridge CB2 2EF, UK E-mail: kim.graham{at}mrc-cbu.cam.ac.uk
 |
Abstract |
|---|
 Top Abstract 1. Introduction2. Disorders of long-term...3. Models of long-term...4. Can computational models...6. ConclusionReferences |
|---|
Semantic dementia is a recently documented syndrome associated with non-Alzheimer degenerative pathology of the polar and inferolateral temporal neocortex, with relative sparing (at least in the early stages) of the hippocampal complex. Patients typically show a progressive deterioration in their semantic knowledge about people, objects, facts and the meanings of words. Yet, at least clinically, they seem to possess relatively preserved day-to-day (episodic) memory. Neuropsychological investigations of semantic dementia provide, therefore, a unique opportunity to investigate the organization of human long-term memory and, more specifically, to determine the relationship between semantic memory and other cognitive systems, such as episodic memory. In this review, we summarize recent empirical findings from patients with semantic dementia and discuss whether the neuropsychological phenomena of the disease are consistent with current cognitive and computational models of human long-term memory and amnesia. Six specific issues are addressed: (i) the relative preservation of category-level (superordinate) compared with fine-graded (subordinate) semantic knowledge as the disease progresses; (ii) the better recall of recent autobiographical and semantic memories compared with those in the distant past; (iii) the preservation of new learning, as measured by recognition memory, early in the disease; (iv) the interaction between autobiographical experience and semantic knowledge in the current, but not the distant, time-period; (v) increased long-term forgetting of newly learned material; and (vi) impaired implicit memory. It is concluded that recent findings from semantic dementia offer strong support for the view that memory consolidation in humans is dependent upon interactions between the hippocampal complex and neocortex. Furthermore, these investigations have provided computational modellers of human memory with a novel set of neuropsychological data to be simulated and tested.
connectionist models; hippocampus; long-term memory; memory consolidation; retrograde amnesia; semantic dementia
BA = Brodmann area; rCBF = regional cerebral blood flow
|
1. Introduction |
|---|
|
TopAbstract1. Introduction2. Disorders of long-term...3. Models of long-term...4. Can computational models...6. ConclusionReferences |
|---|
In 1957, Scoville and Milner described eight patients who showed an extensive and persistent loss of memory after bilateral lesions of the medial temporal lobe. Three of these patients (H.M., D.C. and M.B.) were severely amnesic: they were unable to remember events from moment to moment (anterograde amnesia) and appeared to have a mild loss of old memories extending back in time for 2–3 years only (retrograde amnesia). Since Scoville and Milner's paper, there have been over 50 reported cases of patients with amnesia associated with hippocampal damage (see Anon., 1996).
Early studies of amnesic patients, like H.M., initially suggested that human long-term memory could be fractionated into at least two types: episodic and semantic (Tulving, 1972
). The term episodic memory refers to our store of personally based memories, the retrieval of which involves conscious recollection of the specific temporal–spatial setting of a previous experience, so called `mental time travel' (see Tulving, 1972, 1983, 1995; Tulving and Markowitsch, 1998). In contrast, the term semantic memory applies to our `knowledge of the world', including the meaning of vocabulary, concepts and facts: information which is retrieved without recalling when and where it was first learnt (Patterson and Hodges, 2000). Tulving (1972, 1983) proposed that these two types of memory were psychologically and neurologically distinct and that amnesia was the result of damage to the episodic memory system. Moreover, Scoville and Milner's data suggested a critical role of the medial temporal lobe, in particular the hippocampus, in this type of memory (Scoville and Milner, 1957).
Over the 25 years since Tulving's highly influential precis of human long-term memory (Tulving, 1983), it has become clear that a simple fractionation between episodic and semantic memory cannot explain all the data from patients with amnesia. For example, Tulving's view predicts that patients with amnesia should not show impairments to the acquisition or retrieval of semantic memory. In fact, H.M. himself shows poor post-morbid semantic learning, as measured by his knowledge of vocabulary that has entered the English language after 1955 (Gabrieli et al., 1988; see also the patient described by Verfaellie et al., 1995). For a fuller discussion of semantic learning in amnesia, see section 4.4.
In the domain of remote memory, amnesic patients have equal difficulty in retrieving recent episodic and semantic memories, but older events and knowledge can sometimes be retrieved (Ribot, 1882; Reed and Squire, 1998). The fact that episodic memories can be affected by time (with better preservation of distant events compared with more recent memories) is not predicted by a model in which amnesia is caused by a selective impairment to episodic memory. These neuropsychological findings strongly suggest that Tulving's original hypothesis (Tulving, 1972, 1983) regarding the organization of human long-term memory, in which episodic and semantic memory are neurologically and psychologically distinct, is clearly incorrect.
In order to accommodate these conflicting and complex findings from amnesia, some researchers have proposed a neuroanatomically based theory in which medial temporal lobe structures (i.e. the hippocampus, subiculum and entorhinal cortex) play a temporary, time-limited role in the acquisition of human long-term memories (Rempel-Clower et al., 1996; Graham and Hodges, 1997; Reed and Squire, 1998; although see Nadel and Moscovitch, 1997). More specifically, the hippocampal complex is necessary for the retrieval of recently experienced events, but is not involved in the retrieval of older episodic and semantic memories. In contrast, regions of the neocortex are thought to be the `permanent repository of memory' (Squire and Alvarez, 1995, p. 172). This theory provides a reasonable explanation for why patients with damage to the hippocampal complex show a temporally graded loss of memory, with recent memories affected more severely than older memories. In this article, we discuss this model of long-term memory consolidation in more detail, concentrating on data from the syndrome of semantic dementia and on current neuroanatomically informed computational models of human memory. These connectionist models allow us to test predictions from different cognitive theories of human long-term memory: computer simulation is used as a technique to determine the effects of different types of lesions on memory recall. For example, a number of researchers have shown that lesioning of the hippocampus in a connectionist model results in a similar temporal gradient in memory retrieval to that seen in amnesic patients with hippocampal damage (Alvarez and Squire, 1994; McClelland et al., 1995; Murre, 1996). Other experimental data related to retrograde amnesia, derived from studies with amnesic patients, have also been tested and modelled (Murre, 1996, 1997). This `two-pronged' approach to the study of human long-term memory allows neuropsychological theories to be tested in detail and results in the generation of new hypotheses, which can be investigated in brain-damaged subjects.
Patients with semantic dementia have a disorder that appears to be, both cognitively and neuroanatomically, the mirror image of that seen in amnesia (Snowden et al., 1989; Hodges et al., 1992, 1995; Hodges and Patterson, 1996; Graham et al., 1999b). This syndrome offers a challenge, therefore, to computational models of human memory; such models can only be regarded as valid models of long-term memory if they are able to simulate the neuropsychological phenomena observed in semantic dementia, as well as amnesia. The aims of this paper are, therefore (i) to introduce the neuropsychological profile of semantic dementia by describing data collected in a patient (A.M.) with the disease, (ii) to briefly discuss three neuroanatomically based computational models of long-term memory (Alvarez and Squire, 1994; McClelland et al., 1995; Murre, 1996, 1997), (iii) to suggest a number of predictions about semantic dementia that can be made from one of these models, TraceLink (Murre, 1996, 1997), (iv) to determine the validity of an interdisciplinary approach to semantic dementia by comparing Murre's predictions with empirical data from recent experimental studies of remote memory and new learning in patients with the disease, (v) to provide a theoretically and computationally driven explanation for the neuropsychological findings from semantic dementia, and (vi) to bridge the gap between neuropsychology and computational modelling by providing cognitive neuroscientists interested in long-term memory with a better understanding of how these two disciplines can inform each other and provide new directions for research.
|
2. Disorders of long-term memory: semantic dementia |
|---|
|
TopAbstract1. Introduction2. Disorders of long-term...3. Models of long-term...4. Can computational models...6. ConclusionReferences |
|---|
2.1 Background
In 1982, Mesulam described six patients who showed a progressive, yet selective, impairment to language, a phenomenon he referred to as `a slowly progressing aphasic disorder without the additional intellectual and behavioural disturbances of dementia' (Mesulam, 1982, p. 592). While these were not the first reported cases with progressive aphasia (Pick, 1892
; Warrington, 1975; Schwartz et al., 1979), Mesulam's paper highlighted the fact that patients with neurodegenerative disease could present with relatively focal cognitive deficits. Following this seminal report, more than 100 patients with progressive aphasia have been reported and it has become apparent that this broad term has been used to describe two very different syndromes: progressive non-fluent aphasia and progressive fluent aphasia (for a review of these two disorders see Hodges and Patterson, 1996).
While patients with non-fluent aphasia present with breakdown in the phonological and syntactic aspects of language (Croot et al., 1998), those with progressive fluent aphasia retain normal speech structure, but are unable to produce the names of previously familiar places, people and objects. Such patients also show deficits in word comprehension and fail to understand questions and follow conversation, television programmes, etc. Although anomia and impaired word comprehension are the most striking neuropsychological features, more detailed testing reveals a breakdown in both verbal and non-verbal semantic knowledge about people, objects, facts and words. Deficits are seen, therefore, on a number of different verbally based semantic tasks (Hodges et al., 1992; Hodges and Patterson, 1995), such as picture naming, word–picture matching (pointing to a picture from eight other semantically related foils), category fluency [producing as many exemplars from a semantic category (e.g. animals) in 1 min], naming an item when given a description (e.g. `an electrical kitchen appliance that is used for browning bread'), picture sorting (i.e. grouping black-and-white line drawings depending on various pre-specified criteria, such as living versus non-living; electrical versus non-electrical and so on) and generating verbal descriptions from a spoken label. On non-verbal testing of semantic memory, patients show deficits when asked to select the appropriate colour for a black-and-white line drawing of a familiar object (e.g. yellow for banana), to draw animals and objects from memory, to use previously familiar objects (e.g. a box of matches, kitchen utensils, etc.) and to match common object and animal sounds to the appropriate picture (Bozeat et al., 2000; Hodges et al., 2000; S. Bozeat, M. A. Lambon Ralph, K. S. Graham, K. Patterson, H. Wilkin, J. Rowland et al., unpublished results). In contrast to the impairments seen on tests of semantic knowledge, the patients perform well on tests of visuo-perceptual and spatial ability, non-verbal problem solving and working memory, even at the relatively late stages of the disease (Breedin et al., 1994; Hodges et al., 1994, 1995; Graham et al., 1997b; Waltz et al., 1999).
The relatively selective loss of semantic memory shown in this disease has led many researchers to adopt the term `semantic dementia' as opposed to the designation `progressive fluent aphasia' (see Snowden et al., 1989; Breedin et al., 1994; Hodges et al., 1994; Snowden et al., 1994). Readers should also be aware that patients with semantic dementia may be described clinically as having the temporal variant of frontotemporal dementia (Miller et al., 1993; Edwards-Lee et al., 1997). Criteria for the diagnosis of semantic dementia, very similar to those employed in Cambridge, have been proposed by the consensus study group on frontotemporal lobar degeneration (Neary et al., 1998). The two major subtypes of frontotemporal dementia—reflecting the major locus of pathology, predominantly frontal versus temporal—have distinct cognitive profiles, including the status of semantic and episodic memory (Hodges et al., 1999; Rahman et al., 1999; Perry and Hodges, 2000). In this paper, the neuropsychological data we will review pertains only to the temporal variant of frontotemporal dementia (i.e. semantic dementia).
2.2 Neuroradiology and neuropathology
In all cases of semantic dementia, including more than 30 studied in Cambridge, there is focal atrophy of the inferolateral aspect of the temporal lobe, which is typically most marked on the left, but may be bilateral (Breedin et al., 1994; Hodges et al., 1995, 1998). A recent voxel-based morphometry analysis of the MRI scans of six patients with semantic dementia demonstrated that the most significant and consistent locus of atrophy was the left polar and inferior temporal lobe [Brodmann area (BA) 38/20] (Mummery et al., 2000) and a neuropathological study by Harasty et al. (1996) found significant bilateral atrophy to the inferior and middle temporal gyri. The status of structures in the medial temporal lobe is more controversial: it has been widely thought that the hippocampi are relatively spared in semantic dementia, at least in the early stages of the disease (Graham and Hodges, 1997; Graham et al., 1999b, 2000). In support of this view, Mummery and colleagues' voxel-based morphometry study found no evidence of hippocampal atrophy in their group of patients with semantic dementia (Mummery et al., 2000).
Evidence concerning the pathological basis of semantic dementia is still scant, although earlier speculation that this was unlikely to reflect a variant of Alzheimer's disease is currently upheld: two cases that have been subject to post-mortem in Cambridge have shown changes consistent with Pick's disease and a meta-analysis of 13 cases from the literature revealed that all had either Pick bodies or non-specific histological changes without either Alzheimer's or Pick's pathology (Hodges et al., 1998). In terms of the integrity of neuroanatomical structures at post-mortem, while on the one hand there is consensus about the involvement of the anterior temporal regions, temporal pole and inferomedial temporal cortices (Snowden et al., 1996b; Hodges et al., 1998), there is variability in the degree of reported hippocampal pathology. Graff-Radford and colleagues noted severe involvement of the hippocampus, but Harasty and co-workers and Scheltens and colleagues found absolute and relative sparing of this structure, respectively (Graff-Radford et al., 1990; Scheltens et al., 1990; Harasty et al., 1996).
This brief overview of the literature on neuroradiological and neuropathological studies of semantic dementia illustrates that further investigations of the status of temporal lobe structures are needed in semantic dementia, in particular those that compare methods of hippocampal measurement in vivo and that correlate neuropsychological performance and neuroanatomical damage. Three recent studies have attempted to address this issue and provide more informative data about the neuropathological basis of the disease. Galton and co-workers used a visual rating scale (validated against volumetric measures) to assess the extent of atrophy in the hippocampus, parahippocampal gyrus, anterior and lateral temporal lobe in 30 patients with probable Alzheimer's disease, 30 patients who fulfilled consensus criteria for frontotemporal dementia (17 with semantic dementia and 13 with the frontal variant of frontotemporal dementia) and 18 control subjects (Galton et al., 2001). The major findings from this study were: (i) 50% of the patients with Alzheimer's disease had moderate to severe bilateral hippocampal atrophy compared with the control subjects, but little involvement of other temporal lobe structures; (ii) patients with semantic dementia also had some hippocampal atrophy, particularly evident on the left, in conjunction with significant atrophy bilaterally of parahippocampal regions, lateral temporal lobe and temporal pole; and (iii) the frontal variant frontotemporal dementia group had atrophy in the temporal poles, hippocampi and right parahippocampal gyrus (a pattern that was largely indistinguishable from that seen in Alzheimer's disease). Almost identical findings emerged from a recent study by the London Dementia Research Group (Chan et al., 2001), which obtained volumetric measures of temporal atrophy in patients with Alzheimer's disease and semantic dementia. Taken together these studies confirm that atrophy of the polar and infero-lateral temporal regions differentiates patients with Alzheimer's disease and semantic dementia, but that atrophy to hippocampal and parahippocampal regions can be present in both diseases, with bilateral involvement in Alzheimer's disease and predominantly left-sided atrophy in semantic dementia.
To date, only one empirical study has compared episodic memory performance and degree of temporal lobe atrophy in semantic dementia (Simons et al., 2000). Simons and colleagues used the validated visual rating scale described above (developed by Galton et al., 2001) to assess the degree of involvement of temporal lobe regions in the disease, and on the basis of results from this scale, characterized the laterality of a patient's atrophy as either predominantly left (n = 4), predominantly right (n = 4) or bilateral (n = 5). Subsequently, the authors analysed which of these groups showed a significant episodic memory deficit on the faces component of the Warrington recognition memory test (Warrington, 1984). It was found that so long as atrophy was restricted to the left temporal lobe, patients showed preserved recognition memory. In contrast, patients with right temporal lobe damage, in particular with involvement of the hippocampus and parahippocampal gyrus, were significantly impaired on the test. This study is important for two reasons: (i) it confirms that episodic memory, at least as measured by the faces version of the Warrington recognition memory test, can be preserved in some patients with semantic dementia; and (ii) it reveals that performance on episodic memory tests in the disease is affected by both the extent of atrophy and the laterality of the pathology.
In summary, recent studies that have used rating scales and volumetric measures of temporal lobe regions support the view that anterior and inferior areas of the temporal lobe are affected early in the disease and that this profile is different from that seen in Alzheimer's disease. With respect to medial temporal lobe regions, Chan and colleagues' and Galton and co-workers' investigations clearly demonstrate that the hippocampi and parahippocampal gyri are involved at some point in the progression of the pathology of the disease, although it is unclear how neuropathological involvement correlates with neuropsychological profile (Chan et al., 2001; Galton et al, 2001). Simons and colleagues' experiment reveals that the status of episodic memory, as measured by recognition memory, correlates negatively with atrophy in medial temporal lobe structures and that asymmetrical patterns of pathology may play a critical role in explaining the pattern of performance exhibited by a patient on tests of episodic and semantic memory (Simons et al., 2000). To date, however, a number of important issues relating to the neuropathology in semantic dementia remain unclear and must be considered along with the issues discussed in detail later in this review. For example, at what stage in the disease are medial temporal lobe regions implicated and does the progression of pathology from more lateral to medial areas in the temporal lobe correspond to increasing difficulty on tests of episodic memory? Furthermore, it is important to note that measures of atrophy, such as the rating scales and volumetrics used by Galton, Chan, Simons and colleagues, are not measures of functionality and that we must be careful in extrapolating directly between these structural measures and our cognitive models of memory function.
The only functional imaging study in patients with semantic dementia illustrates how we need to be cautious about interpreting structural MRI data (Mummery et al., 1999). The authors measured regional cerebral blood flow (rCBF) using PET in a semantic decision task in four patients compared with six control subjects. Surprisingly, the patients showed a significant reduction in activity in the left posterior inferior temporal gyrus (BA 37). Voxel-based morphometry, however, revealed significant anterolateral temporal lobe damage (especially on the left side) but no significant structural damage to BA 37. Mummery and colleagues propose that the reduced rCBF seen in BA 37, which is thought to be involved in lexical retrieval (Moore and Price, 1999), is consistent with a loss of activation from more anterior, structurally damaged, temporal regions. The study also shows that lack of structural damage in neurodegenerative patients (as measured using volumetrics or rating scales) does not necessarily conform to normal functionality in that neuroanatomical region.
2.3 Case history
The following case-history, from a patient who was studied longitudinally over 3 years (1994–97), illustrates the pattern of cognitive deficits commonly seen in semantic dementia (see also Hodges and Patterson, 1996; Graham and Hodges, 1997; Knott et al., 1997).
A.M. (d.o.b. 1930), an ex-works manager, presented in April 1994 with an informant-confirmed history of progressive word-finding and comprehension difficulties. The following transcription illustrates this point: while A.M.'s speech was fluent and contained few phonological or syntactic errors, it was strikingly devoid of content.
E.:Can you tell me about a time you were in hospital?
A.M.:Well one of the best places was in April last year here (ha ha) and then April, May, June, July, August, September and then October, and then April today.
E.:Can you remember April last year?
A.M.:April last year, that was the first time, and eh, on the Monday, for example, they were checking all my whatsit, and that was the first time, when my brain was, eh, shown, you know, you know that bar of the brain (indicates left), not the, the other one was okay, but that was lousy, so they did that and then doing everything like that, like this and probably a bit better than I am just now (indicates scanning by moving his hands over his head).
A.M. was a well-educated man with a wide-range of sporting and academic interests. After leaving school at the age of 16 years, he went to night school and then on to university to complete an undergraduate degree in engineering and a master's degree in science. For the rest of his working life, he was employed by the same internationally renowned company, where he eventually became a manager with responsibility for over 450 employees. During his career he travelled extensively, including a 2-year period in the southern hemisphere. The number of awards received by A.M. during his career is testimony to the success he had in his chosen profession.
Formal neuropsychological testing in April 1994 revealed that A.M. was severely impaired on tests of picture naming. He was able to name only three out of 48 black-and-white line drawings of highly familiar objects and animals. On a word-picture matching test based on the same 48 items, A.M. scored 36 out of 48 (25 age-matched controls score, on average, 47.4 ± 1.1; see Hodges and Patterson, 1995). On the picture version of the Pyramid and Palm Trees Test, a test of associative semantic knowledge in which the subject has to decide which of two pictures (a fir tree or a palm tree) goes best with a target picture—pyramid (Howard and Patterson, 1992), A.M. scored 39 out of 52. Control subjects typically score close to ceiling on this test. On non-semantic tasks, e.g. copying the Rey Complex Figure (Osterrieth, 1944), however, A.M.'s performance was faultless. When asked to reproduce the Rey Complex Figure after a 45 min delay, A.M. scored 12.5 (control mean = 15.2 ± 7.4). On non-verbal tests of problem solving, such as Raven's Coloured Matrices (Raven, 1963), a multiple-choice test of visual pattern matching which requires the subject to conceptualize spatial relationships, A.M.'s performance was also remarkably unimpaired.
A.M. was tested approximately every 6 months for the next 3 years. Figure 1A shows his longitudinal performance on the 48-item picture naming and word–picture matching test mentioned previously. A.M. was profoundly anomic when he first presented. This pattern remained remarkably consistent over the six testing sessions (for a more detailed discussion pertaining to A.M.'s anomia see Knott et al., 1997). With respect to word–picture matching A.M. showed a dramatic loss of semantic knowledge over time, which resulted in him performing close to chance by 1996 (5 out of 48; controls = 47.4 ± 1.1). Figure 1B reveals a similar pattern on the pictorial version of the Pyramid and Palm Trees Test: in May 1996, A.M. scored 25 out of 52 (controls = 51.2 ± 1.4).
|
Despite this rapid loss of semantic knowledge, A.M. showed no significant decline on tests of non-verbal problem solving, such as the Raven's Coloured Matrices (Raven, 1963
), scoring 36 out of 36 in November 1996. As noted previously, his copying of the Rey Complex Figure (Osterrieth, 1944) was similarly preserved and stable over time (Fig. 1C
): in November 1996 he scored 33 out of 36 (controls = 34.0 ± 2.9). Figure 1C also shows A.M.'s longitudinal performance on delayed recall of the Rey Figure. While it is clear that A.M.'s performance on this non-verbal memory test has declined over time, even in November 1996 he was still able to reproduce some of the figure (scoring 5 out of 36; controls = 15.2 ± 7.4). Auditory-verbal short-term memory, as measured using forwards and backwards digit span, was initially normal, but has since declined: A.M.'s forwards digit span dropped from 7 to 4 (controls 6.8 ± 1.0), while his backwards span fell from 6 to 0 (controls 4.8 ± 1.2) over the six testing sessions. A.M.'s loss of semantic knowledge had a considerable impact on his everyday activities. On various occasions he misused objects (e.g. he placed a closed umbrella horizontally over his head during a rain storm), selected an inappropriate item (e.g. bringing his wife, who was cleaning the upstairs bathroom, a lawn mower instead of a ladder), and mistaken various food items (e.g. at different times, A.M. put sugar into a glass of wine, orange juice into his lasagne and ate a raw defrosting salmon steak with yoghurt). Activities that used to be commonplace have acquired a new and frightening quality to him: on a plane trip early in 1996 he became clearly distressed at his suitcase being X-rayed and refused to wear a seat-belt in the plane.
Since 1997, A.M. has deteriorated generally, becoming increasingly withdrawn, time-obsessed and disinhibited. Like another patient (J.L.) described by Hodges and colleagues, A.M. showed a fascinating mixture of `preserved and disturbed cognition' (Hodges et al., 1995, p. 467). J.L. would set the house clocks and his watch forward in his impatience to get to a favourite restaurant, not fully comprehending the nature of the relationship between the clock and world time. In A.M.'s case, his wife reported an incident in which she secretly removed his car keys from his key-ring to stop him taking the car for a drive. At this point A.M. was rather obsessive about his driving and very quickly noticed the missing keys. He solved the problem by taking his wife's car keys off her key-ring without her knowledge and going to the locksmiths, successfully, to get a new set cut. At no point did A.M. realize that his wife had taken the keys from his key-ring. Despite his semantic problems, A.M. continued to play sport (particularly golf) regularly each week, remembering correctly when he was to be collected by his friends. In late 1997, however, A.M. lost his sense of time and took to sitting up at night waiting for his golfing partners. His wife reported, however, that despite these profound comprehension difficulties he still possessed problem-solving skills: for example, in 1997, he successfully took up dominoes.
Figure 2 shows two coronally orientated, T1-weighted MRI images of A.M.'s brain, obtained in November 1995. The upper panel (A) illustrates severe left temporal pole atrophy with a lesser degree of shrinkage to the right temporal pole (note that the frontal lobes appear normal). The lower panel (B) shows that the left-sided temporal lobe atrophy involves the inferolateral region and to a lesser degree the superior temporal gyrus. The left hippocampal complex is relatively preserved, but the entorhinal cortex and perirhinal cortex are almost certainly affected, as evidenced by the gross enlargement of the collateral sulcus (indicated with an arrow). The right medial temporal lobe appears normal. The rating of A.M.'s brain from an MRI scan (February 1995), using measures developed by Scheltens and colleagues and Galton and co-workers, confirmed, to some extent, these observations (Scheltens et al., 1992; Galton et al., 2001). Notably, however, there was evidence of significant atrophy to all regions (including the hippocampus and parahippocampal gyrus) bilaterally (see Fig. 1 from Simons et al., 2000).
|
In summary, A.M.'s case history illustrates a number of characteristics (see Hodges et al., 1992) associated with semantic dementia: (i) selective impairment of semantic memory causing severe anomia, impaired single-word comprehension, reduced generation of exemplars on category fluency tests and an impoverished fund of general knowledge; (ii) relative sparing of syntactic and phonological aspects of language; (iii) normal perceptual skills and non-verbal problem-solving abilities; and (iv) relatively preserved recent autobiographical and day-to-day (episodic) memory.
In some senses, the neuropsychological profile observed in patients with semantic dementia (i.e. poor semantic knowledge with relatively preserved episodic memory) seems to provide support for a fractionation between episodic and semantic memory, at least at the psychological level. Patients with semantic dementia provide us, therefore, with a unique opportunity to investigate the organization of human memory and to test the validity of current models of memory consolidation. Before summarizing some of the recent work in semantic dementia, which has addressed this issue, we will briefly review current computational models of amnesia and discuss how these networks can inform us about neurological disorders of long-term memory.
|
3. Models of long-term memory |
|---|
|
TopAbstract1. Introduction2. Disorders of long-term...3. Models of long-term...4. Can computational models...6. ConclusionReferences |
|---|
3.1 Neuroanatomy and global architectures of connectionist models
At present, only one connectionist model of long-term memory, TraceLink, makes specific reference to semantic dementia (Murre, 1996
, 1997). As we will discuss, however, several other computational models can easily be extended to incorporate the neuropsychological data from this disorder. In this section, we will concentrate on the model by Murre and on two other connectionist models that are felt to be particularly promising in explaining the cognitive phenomena arising from semantic dementia, namely those described by Alvarez and Squire (1994) and McClelland et al. (1995). A more extensive and general review of computational models of amnesia, the hippocampus, and long-term memory can be found in the papers by McClelland et al. (1995) and Gluck and Myers (1995). Recent special issues of Hippocampus (Gluck, 1996) and Memory (Mayes and Downes, 1997) also provide a good overview of the current state of computational modelling in this area.
Recent developments in neuroanatomy suggest that the neocortex is densely and bidirectionally connected with the hippocampus and related areas. Although the connectionist models (and non-connectionist theories) of long-term memory stress the importance of the hippocampus, there is little doubt that other medial temporal lobe structures (e.g. entorhinal and perirhinal cortices, parahippocampal gyrus, etc.) are crucially involved in long-term memory consolidation. As will be discussed in more detail later in this article, it is possible that the hippocampus is critical for episodic retrieval, but that other regions of the medial temporal lobe can support learning and consolidation of semantic facts. Felleman and Van Essen position the hippocampus at the top of the neuroanatomical hierarchy of interconnected brain areas, with the sensory and motor organs at the bottom (see Fig. 3A, and Felleman and Van Essen, 1991). Figure 3B illustrates a similar hierarchy described by Squire and Zola-Morgan, in which the entorhinal area is bidirectionally connected to the perirhinal (BA 36) and the parahippocampal cortices (Squire and Zola-Morgan, 1991; Squire, 1992). The lower part of the hierarchy is less specified and represented by a single box called `unimodal and polymodal association areas' (frontal, temporal and parietal lobes).
|
The three computational models described in this paper assume this basic neuroanatomical framework and subscribe to the view that the neocortex and hippocampus play distinct, but complementary, roles in long-term memory storage (i.e. memories are acquired and stored in the brain via the process of memory consolidation). Although different with respect to some details (see later), the general principles underlying these three models are very similar. Whereas initially the retrieval of a recently experienced event is reliant upon the hippocampal system, repeated reinstatement of the hippocampal neocortical ensemble over time results in the formation of a more permanent, hippocampally independent, memory representation in the neocortex. The nature of the consolidation process is largely unknown, but some recent evidence suggests that it may take place during sleep: Wilson and McNaughton (1994) proposed that, during slow-wave sleep, hippocampal representations may be re-activated and used to reinstate the entire cortical activation pattern that was present during the initial learning experience.
At this point, it is important to consider the reasons why we may have evolved a complementary learning system. Murre and Sturdy (1995) argued that this process was necessary because there is compelling neuroanatomical evidence that the neocortex does not have sufficient connectivity to rapidly link activated areas in the short time that an individual experiences an event. This cortical connectivity problem has two important effects: (i) it is highly unlikely that neural connections will be in place when two brain sites must be associated in order to form (part of) a new memory representation; and (ii) the formation of synaptic connections in the neocortex will be slow and accumulative. The neuroanatomical and modular hierarchy of the cortex (as illustrated in Fig. 3A and B) solves this connectivity problem because the hippocampus functions as an intermediate site, which can (i) link distant neocortical regions (see also neuroanatomical models by O'Reilly and McClelland, 1994; Treves and Rolls, 1994; McClelland and Goddard, 1996) and (ii) reinstate patterns of cortical activation (compare Abeles, 1991; Murre, 1996, 1997). Murre and Sturdy suggest, therefore, that memory consolidation is an essential biological process which allows us to circumvent the problems of sparse connectivity and create the long-range neocortical connections necessary for storing permanent memory representations.
McClelland and colleagues have proposed an alternative, although not necessarily mutually exclusive, hypothesis: memory consolidation helps prevent catastrophic interference in sequential learning (McClelland et al., 1995). Using computer-based simulations of semantic memory, they demonstrated that integrating new semantic facts into our existing neocortical database of semantic knowledge can have a negative impact on the integrity of that database. For example, they describe an experiment in which they compared presenting the concept of `penguin' (a bird that swims but does not fly) directly to semantic memory (focused learning) as opposed to indirectly, via a training set (interleaved learning). In the focused learning condition, the network learnt about `penguin' rapidly, but this acquisition of information had a detrimental effect on other bird concepts (e.g. `robin'). Interleaved learning, while slower, did not result in such catastrophic interference.
Although initially counterintuitive, therefore, it might actually be desirable, both neuroanatomically and functionally, for the human brain to have evolved a complementary learning system in which the hippocampus and neocortex interact in the formation of long-term memories. Furthermore, as will become clear, there is strong evidence from studies of amnesia and semantic dementia that the retrieval of memories is influenced by time, a direct prediction from models in which memories are consolidated in the human brain.
3.2 TraceLink model
A schematic drawing of the TraceLink model is shown in Fig. 4. Its three main components are (i) a trace system, (ii) a link system, and (iii) a modulatory system. We will briefly summarize some of the details of the model's implementation, in particular specifics relating to the interaction between the three systems and the formation and maintenance of episodic (and semantic) memories.
|
Trace system
This system represents the neocortical basis for memories. The input to the trace system originates in the sensory areas, although these regions are not represented in the model. Similarly, output or motor areas are not included but are assumed. In the model, we abstract from the complex hierarchical structure of the cortex, reducing it to a system with many modules. Each of these modules contains a number of interconnected `nodes', which are highly abstracted neurones. An episodic memory representation in the trace system is simply a number of activated (firing) nodes (for more details about the implementation of the nodes see Murre, 1996). It is presumed that a relatively small percentage of the nodes participate in a given representation and that nodes in the trace system are sparsely connected in a random fashion. Each node also has connections to and from a random subset of the nodes in the link system. A central assumption in the TraceLink model is that the formation of new long-range associations (i.e. associations between faraway neurones that are not directly connected) is a slow process compared with the formation of associations between the trace and link system.
Link system
The link system's function in the model is to interconnect remote trace nodes (i.e. those without direct cortico-cortical connections). The anatomy of the link system includes the hippocampus and certain other structures, such as the anterior medial temporal lobe structures of the brain. In the model, the link system has a much smaller number of link nodes than the trace system (i.e. it is of limited capacity) and each link node is connected to a random subset of trace nodes. These features are in accordance with our knowledge of the neuroanatomy, as described earlier in the review (Felleman and Van Essen, 1991). Link nodes are also interconnected within the link system (i.e. there are link–link connections). If the modulatory system is functioning well (see below), link–trace and link–link connections are formed more rapidly than trace–trace connections. Other models of amnesia also make a similar assumption of rapid hippocampal learning and slow cortical learning (e.g. Alvarez and Squire, 1994; McClelland et al., 1995).
Modulatory system
The modulatory system includes certain basal forebrain nuclei, especially the nucleus basalis with its cholinergic inputs to the hippocampus via the fornix (see Hasselmo, 1995, 1999) and several areas that have a more indirect, controlling function. The role of the system is to trigger increased plasticity in the link system and hence the ability of the link system to record rapidly a new episodic representation. The modulatory system may be activated directly through central states such as arousal and attention, and through stimulus-specific factors such as novelty and biological relevance (i.e. emotional stimuli involving danger, food, sex, shelter, etc.; these aspects are thought to be processed through the amygdala).
In the following section, we will discuss briefly the processes of normal episodic learning and retrieval, and comment on how TraceLink accounts for some of the data on retrograde amnesia.
Normal learning and recall
Under normal circumstances, a memory representation passes through roughly four stages (see Fig. 5).
|
Stage 1 (Fig. 5A
). Sensory and motor information, comprising an episode-to-be-remembered, activates a set of trace nodes (filled circles).
Stage 2 (Fig. 5B).
The trace elements activate a set of link nodes within seconds. The modulatory system is sufficiently active (e.g. a new or interesting pattern is present), allowing an increase in plasticity in the link system and a subsequent strengthening of connections between activated link nodes and trace nodes (shown by the thickening of the connections). In both the link and trace system, the activation of new memory ensembles will gradually result in the overwriting of older representations (i.e. forgetting). This interference process is more evident in the link system compared with the trace system, however, because the hippocampal system has lower capacity and higher plasticity.
Stage 3 (Fig. 5C).
This stage represents the initial consolidation process. Repeated activation—via rehearsal, reminiscence and perhaps sleep—of the link–trace ensemble leads to the gradual formation of trace–trace connections. These are initially weak, but grow in strength with each consolidation episode.
In TraceLink, the repeated re-activation of learned representations that drives the consolidation process is simulated as follows. The link system is given a burst of random activation, which initiates a random search for the nearest representation. During the search, both the trace system and the link system are active. After a stable representation has been found, the representation remains active until the next burst of random activation in the link system. The search time for a representation is much shorter than the post-search time during which the actual consolidation takes place. Consolidation occurs through the formation and strengthening of connections within the trace system at a fixed base rate (i.e. the trace system's plasticity is not modulated and is the same at all stages). Unlike the trace system, the link system is not plastic during the consolidation process (i.e. there is no change to the link–link or the link–trace connections): if the link system was plastic this would result in runaway consolidation where one particular memory would become overly strong (see later discussion). We refer to Murre for more details on the implementation of the consolidation process (Murre, 1996).
Stage 4 (Fig. 5D).
In the final stage of consolidation, trace–trace connections have become very strong. Link–trace connections may have decayed or been reassigned to other memory traces and retrieval of the initial memory is now independent of the link system.
Initially, a memory representation is dependent upon the link system, but towards completion of the consolidation process it becomes predominantly reliant on the trace system. This `transfer' of memory representations is the basis for explaining retrograde amnesia. By making the link nodes inactive (i.e. modelling a hippocampal lesion) all memory representations at stage 2 are lost. Stage 3 representations may be preserved if they have received sufficient consolidation, whereas stage 4 representations will always be intact. If we assume that the majority of autobiographical memories are consolidated at about the same rate, taking into account the fact that emotional and environmental factors will result in some memories being reinstated more or less often, recent memories will be lost because these will be mainly at stages 2 and 3. Lesioning of the link system, therefore, results in a characteristic Ribot gradient (see Fig. 6A), which fits a power function [control (forgetting): P(t) = 1.1031t–0.8384, r2 = 0.9324; Ribot: P(t) = 0.041t0.5682, r2 = 0.922] and can be compared with remote memory data in patients with Alzheimer's disease and Korsakoff's syndrome (Fig. 6B) (Kopelman, 1989). The exponent of the Ribot curve of 0.5682 indicates that there is a clear increase of recall probability with memory age. More details on the explanations of amnesia by the TraceLink model can be found in other papers (Murre, 1996, 1997), where we describe anterograde amnesia, preserved implicit memory, shrinkage of retrograde amnesia and details of the consolidation process.
|
3.3 Other connectionist models of amnesia
Two other models of amnesia have been published, one by McClelland and colleagues and one by Alvarez and Squire (Alvarez and Squire, 1994
; McClelland et al., 1995). Alvarez and Squire (see also Squire and Alvarez, 1995) present a connectionist implementation based on earlier, non-computational, models of long-term memory by Squire and co-workers (e.g. Squire et al., 1984; Squire and Zola-Morgan, 1991; Squire, 1992). In this model, the entire medial temporal lobe system (see Fig. 3B), rather than just the hippocampus, functions as a site that binds together remote cortical areas (in a similar manner to that suggested by Murre, 1996). Each of these specialized cortical areas contributes to a complete memory representation. A consolidation process is assumed during which cortico-cortical connections emerge between the disparate cortical sites.
The model is implemented in a small connectionist network, where the medial temporal lobe system is represented by four neurones (see Fig. 7). Similar modules of four neurones make up two `cortical areas'. In each of these `winner-take-all' modules, only one out of the four neurones can remain activated. An assumption in this model is that links to and from the medial temporal lobe system have a higher learning rate than the cortico-cortical connections (50 times higher).
|
Simulations with the model showed an initial steep forgetting curve due to fast decay of the medial temporal lobe connections. The system continuously received intermittent consolidation trials, so that additional forgetting was lessened after the initial stage because the cortico-cortical connections, which have a slower decay rate, became more important for recall. Forgetting is simulated by weight decay in the medial temporal lobe initially and then in the cortico-cortical areas. Recall was tested by presenting half of a pattern and assessing how well the other half was recalled.
When the medial temporal lobe system was lesioned immediately after training, performance was near chance because the system had not yet received any consolidation trials and was thus still fully dependent on the (lesioned) medial temporal lobe connections. After sufficient consolidation, lesioning showed no effect on performance because pattern recall had become fully independent of the medial temporal lobe system. The simulations thus showed a clear Ribot gradient in that recent memories had a higher chance of being lost when the medial temporal lobe system was lesioned.
While Alvarez and Squire's (1994) and Murre's (1996) models stress that the hippocampus (medial temporal lobe) links geographically separate cortical regions, McClelland et al. (1995) view the role of the hippocampus slightly differently. These authors have suggested that recent memories are initially stored through fast-changing synapses in the hippocampal system. More specifically, during initial learning a copy of the cortical representation resides in the hippocampus. This `summary sketch' of the memory must be sufficient to retrieve the whole of the neocortical trace. Repeated reinstatement of the hippocampally based memory results in the accumulation of subtle neocortical changes, allowing the new memory to be integrated gradually into existing neocortical networks. McClelland and colleagues propose, therefore, that the hippocampus functions as a `trainer', teaching the hippocampal representations to the cortex, thereby achieving the necessary integration of new and old semantic knowledge and avoiding catastrophic interference. Without such gradual integration of new semantic information, via interleaved learning of new and old patterns by the hippocampus, the cortex part of the model would suffer from very strong interference, whereby new patterns would overwrite old patterns very rapidly. The reason for catastrophic interference in McClelland and colleagues' model is that learning is implemented with the backpropagation algorithm (Rumelhart et al., 1986). Other models of amnesia, such as TraceLink, do not use a backpropagation method and, therefore, do not show such catastrophic interference.
McClelland et al. (1995) report simulations with recall curves before and after lesioning (i.e. disabling) of the hippocampal system and show that while hippocampal lesioning results in a Ribot effect, whereby recent memories are harder to retrieve than older memories, the unlesioned model shows normal forgetting over time. The simulations are in excellent agreement with the experimental data by Zola-Morgan and Squire who obtained evidence for memory consolidation in monkeys over a period of ~10 weeks (Zola-Morgan and Squire, 1990).
In contrast to the TraceLink model, the `hippocampus' in the connectionist model by McClelland et al. (1995) is not implemented as a neural network, but simply as a set of stored patterns. As pointed out by McClelland and colleagues, this simulation should be viewed as a computational implementation of a more biologically and anatomically plausible model that has been analysed in some detail by McClelland and co-workers (O'Reilly and McClelland, 1994), although no amnesia simulations have yet been undertaken with this more biologically plausible model. The approach whereby the function of biological structure is simulated by a non-biologically plausible model is used with success in computational cognitive neuroscience. For example, a similar approach was taken by Gluck and Myers, who simulated a large number of experiments on learning in healthy and hippocampally lesioned animals using backpropagation to approximate the role of the hippocampus in mediating the compression of redundant representations in the brain (Gluck and Myers, 1993). Unfortunately, the non-connectionist implementation of the hippocampus in McClelland's model makes it less straightforward to predict certain characteristics of the neuropsychology of amnesia. It should be noted that the model by McClelland et al. (1995) is still being developed further. For example, O'Reilly and colleagues reported an up-dated form of the model, which includes recollection (implemented by the hippocampus) and familiarity (implemented by cortical regions) components (O'Reilly et al., 1998).
3.4 Modelling of other aspects of the amnesic syndrome: anterograde amnesia
Whereas all three of the models discussed above can simulate the temporally graded retrograde amnesia seen in patients with bilateral damage to the medial temporal lobe, anterograde memory deficits are more difficult to implement. Most theoretical views accept that retrograde amnesia can occur in isolation, predominantly due to damage to temporal neocortical areas and/or frontal regions (Levine et al., 1998; Kapur, 1999), but it is controversial whether such a phenomenon as truly isolated or pure anterograde memory exists. In the models by McClelland et al. (1995) and Alvarez and Squire (1994), anterograde amnesia is explained exclusively in terms of loss of the hippocampal component (see Table 1), a view which predicts that retrograde and anterograde memory should be strongly intercorrelated. From the neuropsychological literature, however, the degree of anterograde amnesia is not always a clear predictor of retrograde amnesia. For example, some studies note that retrograde amnesia and anterograde amnesia are correlated only weakly (ranging from 0.3–0.6) in populations with Alzheimer's and Korsakoff's disease (Shimamura and Squire, 1986; Kopelman, 1989, 1991; Schmidtke and Vollmer, 1997). More importantly, retrograde amnesia can occur in isolation (Kapur, 1993; Evans et al., 1996; Markowitsch, 1996). There are also reports of patients with varying aetiologies, such as fornix damage and transient global amnesia, who show moderate anterograde memory deficits with virtually absent retrograde amnesia (e.g. Hodges and Carpenter, 1991; Kazui et al., 1995). These results seem to suggest that a partial independence of anterograde and retrograde amnesia is possible (Kopelman, 1989; Meudell, 1992). Squire and Alvarez (1995) argue, however, that the findings of a relatively positive relationship between retrograde and anterograde amnesia in some patients is strongly indicative of a common lesion underlying both forms of amnesia. Whatever one's stance on this controversial topic, the experimental literature presents at least one very strong example of complete independence of anterograde and retrograde amnesia. Kopelman (1986) reported that cholinergic blockade results in a severe anterograde amnesia with virtually no accompanying retrograde amnesia. This study provides strong evidence in favour of a separate system that can trigger and modulate learning and that can be disrupted independently of other systems.
|
The TraceLink model does not show the same limitations, at least with respect to modelling anterograde memory, as the other two networks. This is because there are two main causes of anterograde amnesia in the model: (i) lesions of the non-hippocampal component of the modulatory system (e.g. through damage to certain basal forebrain lesions); and (ii) lesions of the link system. The latter lesion results in a correlated degree of anterograde and retrograde amnesia, but when the modulatory system is lesioned, this causes anterograde amnesia with little or no retrograde amnesia (for a more detailed discussion see Murre, 1996, 1997).
It is clear from the discussion above that all existing simulations of amnesia have severe limitations with respect to scale, generality and application (see Table 1 for a point-by-point summary of the three connectionist models described above). They do, nevertheless, allow us to investigate some of the assumptions present in existing, non-computational theories of amnesia. The example of consolidation strategy is especially illustrative here. It is extremely difficult to predict in advance how different factors, such as rate of decay from the hippocampus and rate of incorporation into the neocortex (McClelland et al., 1995), novelty (Murre et al., 1992), frequency of event re-occurence (see Anderson and Schooler, 1991) or biological relevance (LeDoux, 1996), would influence the length and outcome of memory consolidation. Fortunately, simulations exploring these topics enable the generation of more precise constraints on what we mean by the term `relevance' (see discussion in Murre, 1997) and also enable us to devise models in which the `relevance' of a memory representation (in some form) is a guiding factor in consolidation (rather than random selection).
Questions about the implementation of consolidation strategies may also benefit from computational investigations. For example, TraceLink simulations showed that it was important to keep link connections (i.e. within, to and from the hippocampal area) fixed during consolidation and allow only within-trace (i.e. cortico-cortical) connections to be strengthened (see also Hasselmo, 1995, 1999). If this strategy was not followed, randomly selected patterns would become stronger in the link system, making it more likely that they would be selected again for further consolidation. This self-reinforcing strengthening would typically degenerate to a state whereby a single pattern would become extremely strong in the model, overwriting all others. McClelland et al. (1995) did not encounter this problem because their hippocampus was not implemented as a neural network model. Alvarez and Squire (1994) used only two patterns, making the system much less sensitive to such runaway consolidation.
3.5 Multiple trace theory of memory consolidation
Before turning to the question of whether these computational models can usefully explain the neuropsychological patterns seen in amnesia and semantic dementia, it is worth considering another theoretical view of memory consolidation that is relevant to semantic dementia. Nadel and Moscovitch pointed out that very few documented amnesic cases actually show a short duration temporally limited retrograde amnesia after bilateral hippocampal atrophy (as would be predicted by the interactive hippocampal–neocortical view described above: Nadel and Moscovitch, 1997; Moscovitch and Nadel, 1999). Instead, many patients actually show retrograde memory deficits extending back as far as 25–40 years (Victor and Agamanolis, 1990; Kartsounis et al., 1995; Rempel-Clower et al., 1996). The authors proposed, therefore, that the neuropsychological literature is more in keeping with the view that the hippocampus is necessary for the retrieval of all episodic memories regardless of the age of the memory.
Considering Nadel and Moscovitch's (1997) view in more detail, the initial stages of memory encoding are similar to that of the standard model: the geographically separate neural components of a recently experienced memory are bound together by the hippocampal complex, creating a medial temporal–neocortical ensemble. The hippocampal constituent acts as an indexer pinpointing the different neocortical areas that need to be activated to produce the full content of the event [as is true of Murre's (1996) TraceLink model]. Unlike the standard model, whereby repeated reinstatement of memories results in the formation of a permanent, hippocampally independent, neocortical representation of the episodic memory, repeated retrieval of personal experiences in the multiple trace model creates recoded traces of the experience within the hippocampal complex. These traces are distributed throughout the medial temporal lobe and the number of traces is positively correlated with how often an event has been retrieved.
The implications of the formation of multiple traces within the hippocampal complex in Nadel and Moscovitch's (1997) model is that older memories have more traces and are more widely dispersed over the hippocampal complex, thereby reducing the vulnerability of these memories to selective hippocampal damage. In terms of retrograde amnesia, therefore, the extent of the lesion (selective hippocampal versus hippocampal and other medial temporal lobe regions) will be positively correlated with the extent of retrograde amnesia. Moscovitch and Nadel (1999) argue that this view is consistent with results from amnesic patients, such as those reported by Reed and Squire (1998), showing that patients with selective hippocampal lesions had only limited retrograde amnesia, while those with temporal neocortical damage had extensive, non-temporally graded impairments to autobiographical memory.
A noteworthy feature of the multiple trace model is that, unlike episodic memories, new semantic knowledge is consolidated in the cortex and thus becomes hippocampally independent. Consolidation of semantic information is useful because it helps prevent catastrophic interference in sequential learning [as demonstrated by McClelland and colleagues (1995) and discussed earlier in this paper]. The implication of this view is that a patient with a bilateral hippocampal lesion may show a flat gradient for the retrieval of autobiographical memories from the past, but a clear temporal gradient for remote semantic memory, e.g. knowledge of famous personalities. This pattern has not yet been demonstrated. At present, it is unclear how the multiple trace model maps onto theories about the role of hippocampal and perirhinal/entorhinal cortices in episodic and semantic learning, respectively, but it is important to distinguish between the permanent role of the hippocampus in the retrieval of autobiographical memories versus a more temporary, or perhaps weaker, involvement in the acquisition and storage of semantic knowledge.
Further evidence in favour of this compelling theoretical view was reviewed in a recent article (Nadel et al., 2000), including the results of analytical and connectionist simulations of the multiple trace model. The paper also describes a study of remote memory in patients with unilateral temporal lobe epileps