The Memory System of the Brain

The following sections are included:

• PREFACE

• Acknowledgements

• CONTENTS

A prerequisite for a synthetic description of the learning and memory (LM), neural bases is a model of the brain. The model we have adopted is based on functional anatomy and consists of three interacting systems. The first, R for representation, is made of the neurons which code sensory information or motor programs with the highest precision. The second, A for activation, comprises the neurons which are most directly involved in arousal and motivation. The third system, S for supervision, controls goal-directed behaviors. The best illustration of its functions is the “voluntary act” in humans: it involves a representation of a goal and of the appropriate strategies, the evaluation of the results and the correction of errors.

The memory system is not an anatomically separate entity but a set of interactions between R, A and S. In these interactions, the three systems have different though complementary functions. R is mainly involved in encoding and storage. S plays a role in encoding and retrieval through the control of attention and cognitive strategies. Structures of A, which modulate R and S activities, are involved in all stages of memory processes. Different types of LM set into play different types of interactions between R, A and S.

Within this general scheme, we considered data from different levels of organization, from the whole brain to the molecule, through intermediates such as small networks (for example, the cortical column). Finally, an attempt is made at defining the perspectives for future research. Among its main objectives are the integration of LM bases in the neurobiology of the whole behavior, the genetic and developmental factors, new therapies for improving memory in aged and demented people, the design of formalisms able to represent large-scale neural networks.

Despite extensive experimental investigations of human amnesia, the basic nature of this vivid syndrome remains surrounded by controversy. The dynamics of amnesia, the rapid, selective and long-lasting plasticity of hippocampal synapses, and the connections between the hippocampal formation and association neocortex. all suggest that amnesia may result from damage to the medial temporal site where the recent declarative memory trace is temporarily laid down. Alternatively, amnesics' preserved capacity for procedural learning on indirect memory tests suggests that their deficit may rather be in intentional, sustained and directed (i.e., active) encoding/retrieval processes. It has been difficult to distinguish between these possibilities because amnesics are most impaired on direct memory tasks that involve both a new integrative trace and active processes. It is possible that different amnesics may have a relatively greater defect either in the memory trace, or in active memory processes, or both, and these differences could correspond to differences in their anatomical lesions. Specifically, hippocampal formation lesions may disrupt all recent declarative memory traces, whereas brainstem lesions could produce amnesia by impairing modulatory processes essential for encoding/retrieval or for storage. In this model, the different areas of association neocortex with bidirectional hippocampal connections would contribute specificity to encoding/retrieval, with posterior areas encoding the sensory/semantic aspects of events, and prefrontal cortex the ongoing context. Active modulatory processes arising in the brainstem would then function to integrate this extensive declarative memory system. The cognitive correlates and neural substrates of the evoked potentials recorded during declarative memory tasks suggest that they may embody such modulatory processes. Finally, since the prefrontal cortex and the medial temporal lobe appear to control the onset, intensity and duration of the ascending neuromodulation, lesions of these structures may impair aspects of both the trace and of the processes supporting declarative memory. In summary, a model is proposed in which the association neocortex (encoding/retrieval) and hippocampus (trace) are integrated by the brainstem (modulation) to produce the psychological properties of declarative memory.

Contextual changes have been found to influence the performance and expression of a wide range of learning tasks, including classically conditioned and instrumental procedures. However, attempts to uncover the neuropsychological and/or neurobiological concomitants to these demonstrations of context are meagre. In a series of experiments, we have attempted to describe a possible modulation exerted by catecholaminergic pathways upon the role of context in latent inhibition and sensory preconditioning, on the one hand, and spatial navigation and latent learning on the other. In these experiments, noradrenaline and dopamine were depleted either in the adult animal (noradrenaline) or in the neonate. It was found that noradrenergic pathways in the forebrain do influence certain complex tasks, such as latent inhibition and sensory preconditioning. Dopaminergic modulation of spatial learning tasks was demonstrated although the role of dopamine in latent learning was dependent upon the procedure applied (i.e., the number of latent learning trials presented). A tentative discussion of these findings with regard to monoaminergic involvement in the tasks examined is presented and it is suggested that these results may offer the continued development of the context concept within this particular field of endeavour.

Memory performance following callosotomy suggests that in normal subjects the hippocampus actively integrates information across the hemispheres (Phelps et al. in press). The present interpretation of hippocampal function from studies on humans with hippocampal lesions suggests that the hippocampus is involved in explicit access to declarative information (Squire & Zola-Morgan 1991). While this description may adequately describe the behavioral outcome of hippocampal lesions in humans, the work on split brain patients implies that the underlying processing of the hippocampus does more than allow conscious access to information. In the present chapter we propose that the hippocampus is necessary for forming relations between multiple cues. We present research from rats, monkeys and humans suggesting this to be the case and point out the similarities in the current theories of “relational” (Eichenbaum et al. 1990) or “configural” (Sutherland & Rudy 1989) processing in animals and “cohering” (Hirst 1988) processing in human amnesia. Finally, we discuss how this approach to hippocampal functioning may account for the behavioral consequences of disconnecting or lesioning the hippocampus in humans and hypothesize as to what mnemonic processes may be unique to humans.

All living systems are dependent on information from the past. While this information may in part be inherent and genetically coded, there was through evolution a steadily growing increase of flexible and individual-specific information encoding, storage, and retrieval. In mammals, and especially man, this biological tendency resulted in a largely environment-stimulated access to information most essential for survival of the individual and the species. Consequently, the remembrance of emotionally and motivationally flavored events was of greatest importance. The apparent result of this is that there is a substantial overlap of those brain structures implicated in the processing of emotional, motivational, and memory processes, a conclusion obvious from the roles attributed to the Papez circuit. How interwoven arousal, attention, mood, and affect are, can most directly be deduced from the assessment of brain damaged patients. Examples from cases with memory disturbances in whom mood and affect influence memorizing as well as some hypotheses on the possible or likely interaction of mood and memory are given.

In man, the localized thalamic lesions giving rise to memory disorders have generally a vascular origin (infarcts). They very often affect several nuclei or associated tracts, which makes the pathophysiologic interpretations more difficult. Disorders are furthermore partially regressive with time, in their severity but also in their quality, which explains why some of the classical dissociations in amnesia are only observed at the late stage. Memory disorders are variable according to the lesion site and their uni or bilateral character. The “purest” and most severe amnesia are observed in bilateral injury of the anterior or antero-internal structures, which is much more rare in unilateral ones. Severe deficits in “declarative” learning tasks and in long term memory are then observed, which predominate in free recall. Short-term memory is often deficient at the initial and secondary stages, even when late evaluation may find it to be relatively unimpaired. Retrograde amnesia is frequent, but less severe than anterograde amnesia and the temporal gradient, frequently described by the patients, is difficult to assess with the classical tests (questionnaires). These patients have important and often lasting impairment in the control of time, which would contribute to the explanation of their amnesia (source amnesia). In the case of paramedian and moreover subthalamic extension of the lesions, patients frequently have at the initial stage vigilance disorders and then attention impairment and cognitive slowing. Their short-term memory and retrograde memory disorders could be more severe. One of the other factors which may contribute to the severity of amnesia is the extension of the lesions in the more external structures, dorsomedial and lateral nuclei: patients then present supplementary cognitive disorders in language (left lesions) or treatment of spatial information (right lesions). At the late stage, a memory disorder corresponding to the main initial deficit may persist. This phenomenon, related to the hemispheric specialization, is more easily observed in unilateral lesions. Amnesia could also be more severe in the case of left thalamic injury, and this could be related to the “declarative” character of most of the tests used.

There has evolved a conventional approach to analysis of the effect of brain dysfunction, induced by anatomical damage or chemical disruption, on learning and memory. The results have indicated that brain dysfunction impairs some instances of learning and memory, but not all. Given verification that the sensory detection of events to be learned is not impaired, the conventional conclusion has been that brain dysfunction can disrupt the operation of some systems of learning and memory but leave others quite intact.

The clear and often dramatic dissociation between how the CNS is impaired and what aspects of learning and memory are impaired has had an appropriately forceful impact that has shaped this conventional analysis. Yet this analysis, and the general case for separate memory systems, is weakened by an often-overlooked consequence of brain dysfunction: the enhancement of learning and memory that sometimes occurs. Some instances of this enhancement are dismissible as method-specific artifacts that are relatively uninteresting, such as lesion-induced increases in activity imposed on learning that is especially indicated by high levels of activity. Yet the number and variety of enhancement examples argue for their serious consideration and for the reality of enhanced memory function as one consequence of brain dysfunction.

One could conceivably deal with this in terms of a multiple-memory-systems approach, but new and untested assumptions would be necessary. A much simpler form of explanation is possible. Perhaps it is not so much entire memory systems that are affected by brain damage, but subprocesses of the one (or perhaps more) memory system(s). We have in mind subprocesses such as orienting, attention, encoding and retrieval, with present emphasis on changes in encoding that lead to corresponding changes in what is learned and remembered. How might the array of memory changes induced by brain dysfunction be accounted for in these terms? Considerations of recent studies of the ontogeny of learning and memory provide a guide.

Similar to the case of brain dysfunction and memory (although by no means directly analogous), animals and people with a relatively immature brain during the early phases of ontogeny typically have been less effective in learning and memory than are those in later stages of ontogeny. This, too, has been explained by some theorists in terms of the differential availability of certain memory systems; the notion is that early in ontogeny, the critical memory systems needed for effective learning and memory have not yet developed. Like the case of general memory dysfunction, however, this interpretation is complicated by a variety of instances of more effective learning early in ontogeny than later on. An alternative interpretation focusing on ontogenetic changes in encoding and stimulus selection has proven useful. Perhaps it would be equally useful to consider similarly the encoding changes that result from general brain dysfunction.

Learning alters information encoded in the brain about the world, and hence constitutes a change in neuronal semantics. Yet current neurobiological research on learning is very limited in its ability to address problems of neuronal semantics, due to lack of appropriate methodologies that can deal with the coding of discrete items of information in neuronal systems. In contrast, molecular and cellular approaches disclose ample information on elementary and general mechanisms operating in neuronal systems that subserve learning and memory. Current advances in the neurobiology of learning are thus in the domain of general cellular syntactic rules rather than in the domain of molar semantics of neuronal systems. The syntactic approach has revealed the nature of elementary molecular and cellular devices that operate in learning, including cellular acquisition devices, conjuncture devices, and information storage devices. Cellular acquisition devices identified so far in neuronal systems that subserve learning are membrane-bound receptors for neurotransmitters and cellular receptors for Ca²⁺. Cellular conjuncture devices are exemplified by the NMDA receptor and the Ca²⁺ /calmodulin-sensitive adenylate cyclase. Cellular information storage devices are protein kinase systems and molecular cascades that alter gene expression and/or synaptic morphology. The identification of these elementary devices enables heuristic translation of behavioral phenomena, such as contiguity detection or consolidation, into molecular language.

Processing of information into long-term storage (consolidation of memory) and the retrieval of processed knowledge is not independent of the physiological state of animal and man. The neuroendocrine system which is composed of central nervous and peripheral components i.e., peptidergic neurons and forming of membrane-active steroids in the brain, on the one hand, and releasing hormones into the circulation, on the other hand, is the primary messenger of bodily states. The neuroendocrine system is a rapidly responding one to environmental changes and, in turn, assures optimal conditions for processing information into long-term storage. Retrieval of knowledge is then affected either by pro-active influence of neuroendocrine principles during learning and consolidation or by a simple presence (tonic actions) during retrieval. These general conclusions can be drawn from studies devoted to the mnemonic effects of circulating adreno-sympathetic catecholamines epinephrine and norepinephrine, adrenal corticosteroids and (neuro) peptide vasopressin. The action of these hormones is of central nervous nature via direct or indirect mechanisms involving the central nucleus of the amygdala and the hippocampus as major targets.

The most labile and most sensitive link in the chain of events that constitute the memory process is the phase of consolidation that follows after acquisition, by which memories are transformed, at a loss, from an unstable into a stable state. Therefore the pharmacology of consolidation has been better studied than that of the other phases of memory (acquisition, storage, retrieval). Pharmacological studies unveiled much information concerning the actual mechanism of consolidation. This involves excitatory glutamatergic and cholinergic muscarinic synapses in the amygdala, medial septum and hippocampus, inhibited by benziodiazepine-regulated GABA-A synapses and modulated by B-noradrenergic terminals. Other additional neutrotransmitter systems, possibly different in each structure, may also intervene. Peripheral hormones reflexly modulate these mechanisms. The amygdala, medial septum and hippocampus process different aspects or components of memories (spatial, aversive, etc.). The entorhinal cortex, which receives projections from these three structures, has a post-consolidational, presumably integrative function. The glutamatergic synapses involved in memory consolidation in the amygdala and hippocampus sustain this late role of the entorhinal cortex through the generation of relatively brief long-term potentiations.

Evidence supporting the concept of a role of central cholinergic mechanisms in learning and memory processes has mainly been obtained from (i) psychopharmacological investigations; (ii) correlations between deficits in cognitive functions and either pathological or experimentally-induced decrease in cholinergic markers in defined cholinergic pathways; (iii) attenuation of these deficits by drugs that directly (i.e., cholinergic agonists) or indirectly (through trans-synaptic regulation of cholinergic neurones) improve cholinergic transmission, and by cholinergic-rich grafts. Even though there is no question that systemic administration of cholinergic antagonists can produce deleterious effects on performance of a large spectrum of learning and memory tasks, any attempt at a unitary formulation of central cholinergic function using this data base seems difficult. However, the presently available data suggest globally that the ascending cholinergic neurones of the basal forebrain complex play an important role in modulating the activity of certain limbic and cortical structures that have been independently shown to subserve cognitive processes involved in different forms of memory. Recent experiments have emphasized the need to take into account the multiple trans-synaptic interactions between cholinergic neurones and other neuromodulatory systems, as well as to directly study the temporal dynamics of changes in the activity of cholinergic neurones occuring at different stages or resulting from different types of learning and memory testing. Together with neurophysiological investigations, results from these recent experiments will undoubtely contribute to a better and more integrated understanding of the role of central cholinergic synapses in memory processes; moreover they offer some promise in developing pharmacological strategies aimed at counteracting cholinergic hypofunction and thereby at ameliorating the memory deficits observed in normal aging and senile dementia. However, tacit acceptance of this current paradigm inevitably leads to such a ramification of the initial “integrative” causal scheme (i.e.,: acetylcholine → memory) that one may wonder if a full understanding of cholinergic involvement in learning and memory processes may be achieved in this manner.

Eyeblink conditioning and the in vivo and in vitro rabbit hippocampus have been used for increasingly powerful behavioral, neuroanatomical, neurophysiological, and biophysical analyses of associative learning. Hippocampal ablation was shown to block acquisition and/or extinction in eyeblink tasks, most prominently trace conditioning. Extracellular recordings of single- or multiple neurons have demonstrated increased excitability to the CS in eyeblink conditioning. Analyses of cellular events underlying this increased excitability in hippocampal slices demonstrated a conditioning-specific reduction in the post burst afterhyperpolarization (AHP). CA1 pyramidal neurons demonstrate this postsynaptic AHP reduction, while dentate granule neurons do not. The AHP is generated by a calcium-dependent outward potassium current. A calcium channel blocker, nimodipine, facilitates acquisition of trace eyeblink conditioning in the aging rabbit. Convergent data suggest that this learning enhancement may also be mediated by changes in hippocampal excitability, via reduced calcium influx through voltage-gated calcium channels and a secondary AHP reduction. Finally, the involvement of NMDA-mediated synaptic transmission in the acquisition of the hippocampally-dependent eyeblink response has been explored. Eyeblink conditioning is enhanced by agonists of the glycine site on the NMDA receptor and, conversely, is blocked by noncompetitive antagonists. The relevance of our findings to studies of learning mechanisms in invertebrate preparations and to potential clinical applications in aging and/or Alzheimer's disease is discussed.

Long-term potentiation (LTP) of synaptic efficacy following repetitive (tetanic) inputs was described originally in the hippocampus, and it has been studied extensively based on the hypothesis that it represents a synaptic model for learning and memory in the brain. In the developing visual cortex, long-term depression (LTD) as well as LTP was found to be induced by tetanic stimulation of afferents, and such a synaptic modification was proposed as a basis for experience-dependent change in functional properties of cortical neurons during the critical period of postnatal development. This chapter deals with the induction of LTP and LTD in visual cortex and their possible functional significances. Among possible molecular mechanisms for the induction of LTP and LTD, those including the involvement of NMDA receptors, Ca²⁺/calmodulin dependent protein kinase II, protein kinase C, phosphatidylinositol turnover and membrane-associated cytoskeletal proteins have been reviewed, although the results obtained so far in visual cortex are only fragmentary.

Long-term potentiation (LTP) was initially described as an increase of hippocampal responses after brief (seconds) tetanization of monosynaptic pathways. Characteristics of LTP (longevity, specificity, associativity) are analogous to those of behavioural memory. LTP induction in hippocampal CA1 area depends on simultaneous pre- and postsynaptic activity which presumably enables postsynaptic Ca²⁺ entry. The increased intracellular Ca²⁺ triggers a biochemical cascade which might lead to protein kinase phosphorylation (to support maintenance of initial LTP stages of several hours) and protein synthesis (to support later maintenance). It is suggested that additional modulatory (“reinforcing”) influences are necessary to trigger the later stages under normal conditions. Quantal analysis and other data indicate that an essential mechanism of LTP maintenance is an increase in transmitter release at least during initial one hour. The separation of induction and maintenance mechanisms suggests an involvement of retrograde messengers from the post to presynaptic side. Generally, the data on both induction and maintenance mechanisms are still controversial and suggest a variety of mechanisms especially for other neuronal structures than CA1. The hippocampal LTP is a convenient paradigm to study mechanisms of information storage. However, studies of simple conditioned reflexes with short pathways through the neocortex might be more promising for further studies of the link between LTP and behavioural memory.

Classical conditioning rapidly produces enduring frequency-specific modification of receptive fields (RF) in the auditory cortex (ACx) which favor the processing of the frequency of the conditioned stimulus (CS). Responses to the CS are increased whereas responses to the pre-training best frequency (BF) and other frequencies are decreased; tuning is often completely shifted so that the frequency of the CS becomes the BF. Such plasticity is observed both for single tone and for two-tone discrimination training. CS-specific RF plasticity may be reversed by extinction training. Sensitization training produces only general increases in responsiveness. Habituation produces frequency-specific decreased responses in the RF. Tuning shifts similar to those produced by conditioning can be produced by iontophoretic application of muscarinic agonists or cholinesterase antagonists to the ACx and pairing one tone with application of ACh to the auditory cortex produces receptive field plasticity which is specific to the frequency of the paired tone. Dual medial geniculate (MG) input to the auditory cortex consists of a frequency-specific non-plastic nucleus (MGv) and a broadly-tuned plastic nucleus (MGm). A preliminary model of receptive field plasticity and behavioral learning is presented. It links MGv and MGm influences on auditory cortex with cholinergic neuromodulation, and makes several predictions, some of which have recently been supported.

It is well-established from work on mammals that when input can be limited to one cerebral hemisphere, testing the other subsequently reveals that it also has access to memory traces created in relation to that lateralized input. What is unclear is whether this aptness derives from the trace having been imparted to both hemispheres, or whether it resides in the one and is called forth by the other. The situation is even less clear when, as is normally the case, input is likely to be distributed concurrently, but unevenly, to both hemispheres. The evidence reviewed herein suggests that there are many answers. Bilateral engrams can often be assumed, and in a few instances can be demonstrated. Unilateral engrams are also formed, as a consequence of lateralized input or manipulation, maturational age, “focussed” attention, or perhaps in many other instances not readily amenable to analysis. In mammals the unprivileged hemisphere has access via components of the forebrain commissures to unilateral engrams held by its partner; but fish and birds in some instances lack such ability. Tetanization of the hippocampal system can be used to reveal some of these effects; and data from split-brain macaques suggest that brainstem processes participate bilaterally in mnemonic processing.

Our best conception for the representation of learned information is the cell assembly, a self-exciting configuration of nerve cells selected from the larger network of the cerebral cortex. Despite its advantages, there is a difficulty that such assemblies may not be able to operate adequately because of insufficient connections. In the first part of this chapter, an attempt is made to define this limitation on cell assemblies, using probability calculations. The difficulty is most severe for large scale integration of cortical cell assemblies, and at the stage of formation of cell assemblies, rather than in local integration, and for the routine operation of well-established assemblies. It has previously been suggested that circuits of neural activity between cortex and hippocampus, resonating at the frequency of the theta rhythm, are important in reducing these difficulties in cell assembly function. This theory is briefly summarized. The second half of the chapter links these ideas about cell assemblies to the various concepts of the psychological impairment which follows damage to the hippocampus and related structures. These concepts include place representation, context representation, working memory, cross-temporal associations, configural representation, conditional responding and declarative memory. Despite the wide range of these various concepts, they can all be regarded as instances when the integration of activity in wide areas of cerebral cortex, and/or its standardization and stabilization over time are required. The prediction is made that psychological functions vulnerable to hippocampal damage should also be vulnerable to cortical damage in many cortical fields rather than single areas.

The primate hippocampus is needed for spatial memory tasks in which the location of objects must be remembered, and in which the location of places where responses must be made are to be remembered. It is also involved in some non-spatial memory tasks such as recognition memory for visual stimuli. Some single neurons in the hippocampus of macaque monkeys performing these memory tasks respond to the positions in space of the stimuli; or to a combination of a non-spatial stimulus with a spatial response when the monkey must learn to make spatial responses to visual stimuli, with the responses of these neurons becoming modified during this type of learning; or to only novel visual stimuli in the recognition memory task. On the basis of these and related findings the hypothesis is suggested that the importance of the hippocampus in spatial and other memories is that it can rapidly form “episodic” representations of information originating from many different areas of the cerebral association cortex. These neurophysiological findings have been combined with information on the microanatomy of the hippocampus, and on long-term potentiation in the hippocampus to produce a neuronal network theory of the operation of the hippocampus. A key aspect of the theory is that the CA3 pyramidal cells with their 4% interconnectivity and Hebb-modifiable synapses implement an autoassociation memory which provides the basis for “episodic” memories which are required for many spatial and non-spatial memory functions.

In this chapter I point at a broad analogy between observed behaviour of neuronal firing patterns and structures appearing in complex ecological systems, notably temperate pristine forests. Although such a connection seems at first remote, I will argue that the mosaic-cycle (patch dynamics) concept seems particularly apt at describing universal features common to a wide variety of self organizing hierarchical structures. A simple approach where neurons are represented by connected probabilistic Turing machines will be introduced and related to the mosaic-cycle picture. At a more general level, I will try to sum up some of the generic properties an information processing machine like the brain must possess, including error correcting coding, optimal control, and universal computability. Finally, simple mathematical arguments are given to explain why perception and learning must be active, dialogue like processes.

We will consider recent evidence for spatial and temporal modularity, as well as structured connectivity in the cortex. These modularities have consequences concerning the “code” of cortical function, i.e., the internal language of the cortex - “how it talks to itself”. In order to understand learning and memory processes in the cortex, we must also investigate this code and suggest that they are crucially dependent. We discuss several spatial and several temporal scales in the cortex, concentrating here on their role in learning and memory. For example, we have proposed that the concept of selective versus instructional learning depends strongly on these modularities, as does the memory storage of large sequential quantities of information as in music. Short-term plasticity in the cortex is discussed: we suggest a key method to induce these effects involves the presentation of spatially and temporally patterned stimuli. We will base much of our remarks on recent results and generalizations of the trion model which is a highly structured mathematical realization of the Mountcastle organizational principle in which the cortical column is the basic neural network of cortex and is comprised of subunit minicolumns, our idealized trions. A columnar network of trions has a large repertoire of quasistable, periodic spatial-temporal firing patterns, MPs, which can be excited. These MPs can be readily enhanced (as well as inherent categories of MPs) by only a small change in connection strengths via a Hebb learning rule. The MPs evolve in natural sequences (related by certain symmetries) from one to another in Monte Carlo probabilistic evolutions. As the synaptic (neurotransmitter) fluctuation parameter B is varied, there exist a series of specific values B(n) giving new repertoires of MPs. The learning properties of the MPs are especially enhanced at these B(n) and the nature of the Monte Carlo evolutions are qualitatively different at the B(n). Learning properties between coupled columnar networks are discussed. A recent testable prediction that an epileptic focus might be eliminated by spatially and temporally patterned electrical stimulation depends crucially on both the modularity of the cortex and the details of the Hebb learning rule.

The following sections are included:

• AUTHOR INDEX

• SUBJECT INDEX