Distributed Memory for Both Short and Long Term - Science Direct

NEUROBIOLOGY OF LEARNING AND MEMORY ARTICLE NO.

70, 268–274 (1998)

NL983852

Distributed Memory for Both Short and Long Term Joaquin M. Fuster Brain Research Institute and Department of Psychiatry and Biobehavioral Sciences, School of Medicine, University of California at Los Angeles, Los Angeles, California 90024

Neuropsychology points to the wide distribution of cortical memory networks. Electrophysiology and neuroimaging indicate that working memory, like long-term memory, is a widely distributed function, largely neocortical. Most of the evidence available from those three methodologies suggests that both working memory and long-term memory share the same substrate: a system of broad, partly overlapping and interconnected neocortical networks. Working memory appears mostly, if not completely, characterized by the sustained activation of one widely distributed network of long-term memory. That activation is at least in part sustained by reentrant excitatory loops through the different neuronal assemblies that constitute the network and that represent the associated features of the memorandum. q 1998 Academic Press

Here I present briefly a few general notions concerning the cortical topography and dynamics of memory that lately have been gaining empirical support and plausibility. They derive to a considerable extent from recent research on so-called working memory in the human and nonhuman primate. Some of this research is summarized below. The ideas it has generated are not only consistent with the data presented in these proceedings by several other authors, but suggest potentially fruitful lines of future research for all of us. The main idea I wish to offer here is that working memory consists of the sustained but temporary activation of a network of cortical neurons, much more extensive than has been recognized, at the service of ongoing behavior, speech, or reasoning. By virtue of its connections between cortical sites, that network represents the associated components of long-term memory and procedural knowledge that together constitute a ‘‘working memorandum.’’ For their continuity and the attainment of their goals, behavior, speech, and reasoning require the timely and temporary retention of many such memoranda in succession, all with wide though differing cortical distribution. Indeed it appears that the activated networks that represent those memoranda are not localized in any given cortical area but, instead, straddle distant and large domains of posterior (postrolandic, postcentral) and anterior (prefrontal) cortex. Lashley’s (1950) neuropsychological experiments provided the first indications of the distributed character of memory in the cerebral cortex. Thus the evidence that he obtained in rats and monkeys supported certain distributed Address requests for reprints and correspondence to J. M. Fuster, MD, UCLA Neuropsychiatric Institute, 760 Westwood Plaza, Los Angeles, CA 90024. Fax: 310–825–6792. E-mail: joaquinF @ucla.edu.

1074-7427/98 $25.00 Copyright q 1998 by Academic Press All rights of reproduction in any form reserved.

AID

NLM 3852

/

268

6v16$$$161 09-10-98 09:36:43

nlmoas

AP: NLM

DISTRIBUTED MEMORY FOR BOTH SHORT AND LONG TERM

269

constructs of the neural substrate of human memory that were emerging at about the same time that those experiments were performed. Prominent among these constructs was the one proposed by Hayek in his book The Sensory Order (1952). He was the first to enunciate a network concept of perception and memory. According to Hayek, memories are stored in widely distributed networks (‘‘maps’’) of interconnected cortical neurons. Such networks are formed by facilitation of synapses as a result of the temporal coincidence of inputs on cortical neurons, in accord with principles of synaptic change and modulation similar to those proposed by Hebb (1949). The reactivation of networks that have been formed in this manner would be the basis of both perception and the acquisition of new memory. New memory would result from the reactivation of old memory networks and from their modification and expansion by new experience. Edelman and Mountcastle (1978), based on knowledge of cortical physiology better than that possessed by former theorists, also postulated a wide distribution of cortical memory, though not as wide as that envisioned by Hayek and supported by recent research. In their theory, the essential unit of representation is the cortical column, a vertical—transcortical—array of equifunctional neurons, such as those that have been identified in sensory cortices; reentry of impulses through recurrent fibers within and between such neuronal groups plays a critical role in determining the selectivity and individuality of both learning and sensory perception. There is now considerable evidence that the hippocampus and other structures of the medial and inferior temporal lobe play an important role in the formation of long-term memories, thus presumably in the making and expansion of cortical networks (Squire & Zola-Morgan, 1988). Memory formation is probably mediated by the reciprocal fiber connections between the hippocampus and the areas of association cortex (Amaral, 1987). The precise mechanisms of memory formation through transactions between hippocampus and neocortex are still unknown; they may involve, among other things, certain local synaptic processes such as long-term potentiation (LTP) and certain glutaminergic receptors, such as NMDA receptors (for review, see Fuster, 1995). It remains an open question to what extent, in primates, the hippocampus plays its memory role inasmuch as itself is an ancient part of cortex. It is possible that in a lower species, such as the rat, the hippocampus plays in spatial (McNaughton, this issue) and auditory (Sakurai, this issue) memory the role that the primate neocortex plays in such memories. Because long-term memories are made of innumerable associated elements of sensorium and of organismic action with widely dispersed cortical representations, the networks of long-term memory can be expected to be distributed over large expanses of the cortex. Furthermore, as the associated elements of a given memory can be also components of other memories, any cortical neuron or group of neurons can be part of many memory networks, as noted by Sakurai. As we see below, both these corollaries of network memory theory find support in PET imaging and in the recordings of cortical unit activity during working-memory tasks, inasmuch as working memory can be reasonably construed as the temporary activation of long-term memory. Another plausible yet unproven hypothesis, this one supported by neuropsychology, is that memory develops in the neocortex along connective and ontoge-

AID

NLM 3852

/

6v16$$$162 09-10-98 09:36:43

nlmoas

AP: NLM

270

JOAQUIN M. FUSTER

netic gradients, that is, from primary sensory or motor projection cortex, toward and into cortex of association (Fuster, 1995). Thus the simplest and most concrete memories would rest on primary and secondary sensory or motor cortices. These cortices contain the basic innate structure that is ready to respond to sensory primitives and to effect motor primitives; in other words, they contain the memory of the species (phyletic memory) (Fuster, 1995). Individual memories of increasing complexity and generality would rest on progressively more distant—in connective terms—and later developing cortices of association. The result would be the development of two hierarchies of memory, one for perceptual memory and the other for motor memory, in posterior and frontal cortex, respectively. Perceptual memory is memory acquired through the senses. It ranges from the most concrete sensory and polysensory memories in networks of sensory and parasensory cortex, at the bottom of the posterior cortical hierarchy, to the most abstract (conceptual) knowledge at the top, in posterior association cortex. Neuronal evidence of auditory and visual memory in, respectively, auditory and inferotemporal cortex is presented hereby by Weinberger and Miyashita (this issue). Between sensory and abstract memory or knowledge, presumably also in posterior association cortex, lie the networks representing episodic and semantic memory. The transitions between stages of memory are gradual at all levels. Thus phyletic memory blends into individual sensory memory, the latter into episodic memory, episodic into semantic memory, and semantic into conceptual knowledge. In this theoretical framework, each stage provides the building blocks, the component networks, for the stage immediately above. At increasingly higher levels of the hierarchy, memories have a progressively wider cortical distribution. At the same time, they become progressively more robust and resistant to local cortical damage, as they become more widely anchored and accessible to more lines of associated input. A comparable hierarchy is suggested by neuropsychology in frontal cortex. This hierarchy is for motor or executive memory, that is, the memory for organismic actions. This cortical hierarchy for action memory is probably the extension of the subcortical motor hierarchy, which is based at its lowest level in the spinal cord and includes such structures as the cerebellum, the lateral thalamus, and the basal ganglia. Its lowest cortical level is the primary motor cortex of area 4, still part of phyletic motor memory. This cortex represents the most elementary motor acts, defined by muscles and muscle groups. Above it, the premotor cortex represents more complex motor acts, defined by goal and trajectory. Above the premotor cortex, the dorsolateral prefrontal cortex represents the schemas, programs, and general concepts of sequential action, these also defined by goal, but of a higher category and projected into the future. Anatomically, the connectivity that links vertically and reciprocally the successive cortical stages of the perceptual as well as the motor memory hierarchy is complemented by the horizontal and also reciprocal connectivity that links the stages of the posterior, perceptual hierarchy to those of the frontal, motor one (for review, see Fuster, 1997). It is through this posterofrontal connectivity that perceptual memories extend to, and are integrated with, motor memories. If, as I postulate, long-term memory is represented in widespread and inter-

AID

NLM 3852

/

6v16$$$162 09-10-98 09:36:43

nlmoas

AP: NLM

DISTRIBUTED MEMORY FOR BOTH SHORT AND LONG TERM

271

connected networks of posterior and frontal cortex, and if working memory consists in the temporary activation of such networks for ad hoc behavioral purposes, then the study of working memory should help us clarify the topography and dynamics of long-term memory. The cortical recording of the activation of long-term memory in the behavioral context of a working-memory task should reveal its functional architecture. So far, it has been possible to test three specific hypotheses that stem directly from those general postulates: 1. The activation of an old memory for the short term will be manifested by the temporary activation of a widespread network of posterior and frontal cortex. That network will extend into all areas of cortex representing the associated sensory and motor components of the memorandum, that is, among others, the sensory aspects of the cue to be retained in working memory and the motor aspects of the impending response that is contingent on that cue. It is now well documented that the retention of the memorandum in any visual memory task is accompanied by the sustained activation of cells in widespread regions of posterior and frontal cortex. To be sure, the spread and degree of that cellular activation depend on the nature of the visual information in memory and of the motor response that it calls for, but in any case the activated cortical network appears extensive. For example, if that information is a simple color or pattern in a delayed matching to sample task, sustained neuronal activation will be observed in broad inferotemporal regions (Fuster & Jervey, 1982; Miyashita & Chang, 1988) and, in addition, in areas of dorsolateral and ventral prefrontal cortex (Fuster, Bauer, & Jervey, 1982; Quintana, Yajeya, & Fuster, 1988). Neuroimaging in the human also provides a picture of extensive cortical activation in visual working memory. By PET measurement of fluor-deoxyglucose (FDG) uptake, we (Swartz, Halgren, Fuster, Simpkins, Gee, & Mandelkern, 1995) were able to observe the activation of large regions of posterior (occipitoinferotemporal) and frontal cortex during the retention of abstract visual memoranda in a working-memory task. 2. A cortical cell that is part of several—partly—overlapping networks (memories) will be activated during working memory of several sensory modalities and in several memory tasks. This activation of memory cells by different memoranda can occur anywhere where networks—and tasks—share representational elements, in perceptual or motor cortex. The best evidence we have obtained of network-sharing cells is from the prefrontal cortex of monkeys performing memory tasks with different visual memoranda, such as delayed response (a spatial task) and delayed matching to sample (a nonspatial task). Some dorsolateral prefrontal units show activation during memorization of either a spatial cue or a nonspatial one, though with some preference for one or the other (Fuster et al., 1982). The evidence that a cell is active in both spatial and nonspatial memory can best be explained by assuming the participation of that cell in a common component of two intersecting networks. For example, a prefrontal cell may be part of two memory networks, one for spatial memory and the other for nonspatial memory, by virtue of a common association with a motor response that is part of two

AID

NLM 3852

/

6v16$$$162 09-10-98 09:36:43

nlmoas

AP: NLM

272

JOAQUIN M. FUSTER

memory tasks, spatial and nonspatial. If that cell is involved in the representation of that common motor association of the two memoranda, it will appear to be involved in the retention of both. 3. A cell anywhere in the cortical network of an activated memory will be subject to inputs from other parts of the network, that is, from areas of representation of stimuli and responses that by prior association (by training, learning) have become part of the memorandum. Possibly, the reentry of impulses through the areas of representation of the associated properties of the memorandum helps maintain the network active for the short term, which is the essence of working memory. Next is some empirical evidence that this is indeed the case. The importance of functional interactions between distant parts of the cortex that belong to the same network, possibly involving reentry, can be exposed by cooling one cortical region while recording from another during workingmemory performance. Thus, in a visual memory task, the cooling (reversible depression) of either cortex, inferotemporal or prefrontal, will induce changes in the cell activity of the other (Fuster, Bauer, & Jervey, 1985). Most notably, some cells that are differentially activated by different visual memoranda will show a diminution of that differential activation during their mnemonic retention. That cryogenically induced loss of cellular differentiation of memoranda is accompanied by a drop in correct performance of the task, a sign of impaired working memory. These results can best be interpreted as manifestations of the interruption of corticocortical loops of excitatory reentry that is necessary for the maintenance of active memories in vast cortical networks. A more general point: Any cortical cell that is part of many networks or of a complex network with many components will be, during their activation in working memory, presumably subject to presynaptic inputs of great variety, excitatory and inhibitory. These inputs originate in other cortical and subcortical sites of associated representation that are jointly activated in working memory. During that activation, some inputs may reenter by recurrent circuits coursing through various other cortical areas, as suspected in the cooling experiments above. In sum, the cell will be ‘‘solicited’’ by a wider variety of inputs than in a baseline state of passive, unused memory. An immediate consequence of this increased advent of impulses of excitatory and inhibitory nature will be to unstabilize the firing frequency of the cell. Firing may not only become faster or slower on the average, but become more subject to transient variations. In fact, the cell may not change its average frequency and yet show an increase in firing fluctuations, or frequency transitions, during active memorization. As I explain next, we have been able to verify this prediction (Bodner, Zhou, & Fuster, 1997). For that purpose, we submitted to binary mapping the spike trains of cells in somatosensory cortex during performance of a haptic (tactile) memory task. Binary mapping basically consists of (a) the segmentation of time into equalsize bins, and (b) the assignment of 1 or 0 to each bin depending on whether it contains any spike or not. In the resulting temporal map, or binary curve, a transition of frequency is defined as a transition from 0 to 1 or vice versa. By segmenting the record into bins of multiple sizes, i.e., durations, the method

AID

NLM 3852

/

6v16$$$162 09-10-98 09:36:43

nlmoas

AP: NLM

DISTRIBUTED MEMORY FOR BOTH SHORT AND LONG TERM

273

becomes essentially a low-pass filter of spike frequency changes or transitions. In our analysis we used all bin sizes between 1 and 100 ms in 1-ms increments. As we had hypothesized before the analysis, transition numbers turned out to be greater, at more bin sizes (in the range 20 to 50 ms), during ‘‘delays’’ (memorization periods) than during intertrial baseline periods. Thus, it appeared that, overall, in working memory there were more fluctuations of firing frequency than in baseline periods. The cells appeared indeed subject to increased inputs during the periods of memorization of the surface features of objects perceived by touch. This phenomenon need not be accompanied by average frequency change, even though in somatosensory cortex, as elsewhere, many cells are commonly activated, their firing frequency increased, during working memory. In sum, cells in active memory appeared to fluctuate more often between different ‘‘attractor’’ frequencies than in baseline conditions. That indication of increased frequency ‘‘attractors’’ and fluctuations would be consistent with our computer model of the cortex in active short-term memory (Zipser, Kehoe, Littlewort, & Fuster, 1993). The essence of this model is a profusely recurrent network of cortical neurons interconnected by synapses of fixed weights established by learning. During the activation of such an artificial memory network, its ‘‘hidden units’’ exhibit patterns of behavior extraordinarily similar to those of real cortical cells in working memory. Both the model’s unit and the real cells generally show increased average firing. On fine analysis, however, the records from the model and from real cells reveal rapid transitions or fluctuations between ‘‘attractor’’ frequencies. Those transitions may reflect the reentrant circulation of excitatory impulses through network components that represent the various associated features of the memorandum in working memory. REFERENCES Amaral, D. G. (1987). Memory: Anatomical organization of candidate brain regions. In Handbook of physiology: Nervous System, Vol. V: Higher functions of the brain (Part 1, pp. 211–294). Am. Physiol. Soc., Bethesda, MD. Bodner, M., Zhou, Y., & Fuster, J. M. (1997). Binary mapping of cortical spike trains in shortterm memory. Journal of Neurophysiology, 77, 2219–2222. Fuster, J. M. (1995). Memory in the cerebral cortex. An empirical approach to neural networks in the human and nonhuman primate. Cambridge, MA: MIT Press. Fuster, J. M. (1997). Network memory. Trends in NeuroSciences, 20, 451–459. Fuster, J. M., Bauer, R. H., & Jervey, J. P. (1982). Cellular discharge in the dorsolateral prefrontal cortex of the monkey in cognitive tasks. Experimental Neurology, 77, 679–694. Fuster, J. M., Bauer, R. H., & Jervey, J. P. (1985). Functional interactions between inferotemporal and prefrontal cortex in a cognitive task. Brain Research, 330, 299–307. Fuster, J. M., & Jervey, J. P. (1982). Neuronal firing in the inferotemporal cortex of the monkey in a visual memory task. Journal of Neuroscience, 2, 361–375. Hayek, F. A. (1952). The sensory order. Chicago: Univ. of Chicago Press. Hebb, D. O. (1949). The organization of behavior. New York: Wiley. Lashley, K. S. (1950). In search of the engram. Symp. Soc. Exp. Biol. 4, 454–482. Miyashita, Y., & Chang, H. S. (1988). Neuronal correlate of pictorial short-term memory in the primate temporal cortex. Nature, 331, 68–70. Mountcastle, V. B. (1978). An organizing principle for cerebral function: The unit module and the distributed system. In G. M. Edelman & V. B. Mountcastle (Eds.), The mindful brain. (pp. 7– 50). New York: Plenum.

AID

NLM 3852

/

6v16$$$162 09-10-98 09:36:43

nlmoas

AP: NLM

274

JOAQUIN M. FUSTER

Quintana, J., Yajeya, J., & Fuster, J. M. (1988). Prefrontal representation of stimulus attributes during delay tasks. I. Unit activity in cross-temporal integration of sensory and sensory-motor information. Brain Research, 474, 211–221. Sakurai, Y. (1998). Cell assembly coding in several memory processes. Neurobiology of Learning and Memory, 70, 212–225. Squire, L. R., & Zola-Morgan, S. (1988). Memory: Brain systems and behavior. Trends in NeuroSciences, 11, 170–175. Swartz, B. E., Halgren, E., Fuster, J. M., Simpkins, F., Gee, M., & Mandelkern, M. (1995). Cortical metabolic activation in humans during a visual memory task. Cerebral Cortex, 3, 205–214. Zipser, D., Kehoe, B., Littlewort, G., & Fuster, J. (1993). A spiking network model of short-term active memory. Journal of Neuroscience, 13, 3406–3420.

AID

NLM 3852

/

6v16$$$162 09-10-98 09:36:43

nlmoas

AP: NLM