Neuronal Mechanisms of Memory Formation

Dr E. J. Gorzelanczyk Dr P. A.Wozniak July 1, 2001

Book Review: Cambridge University Press 2001, 490 pages About the reviewers: Dr Edward J. Gorzelanczyk, Laboratory of the Applied Research at the Department of Health Sciences, Karol Marcinkowski Medical Academy, ul. Dabrowskiego 79, 60-529 Poznan, Poland Dr Piotr A. Wozniak, SuperMemo R&D, SuperMemo World, ul.R.Maya 1, 61-371 Poznan, Poland Book availability: ($90, July 2001)

Neuronal Mechanisms of Memory Formation edited by Christian H????pan>lscher revolves largely around the concept of long-term potentiation (LTP) and its implications for our understanding how memory traces are formed and sustained. The book was written by a selection of authors that represent many prominent personalities in the field of memory research (incl. Nobel Prize winner Dr Eric Kandel). The book is composed of nineteen independent papers with an introduction by Christian H????pan>lscher which lays ground for critical analysis of the most pivotal questions, on which the current research on memory formation is focused. All individual papers, without an exception, provide a very well-written, carefully proofed, and well generalized analysis of a selected research problem. Although independent authoring has resulted in a slightly repetitive treatment of introductory themes, each chapter of the book can be picked for selective reading and forms a coherent entity on its own. On the downside, the book does not in any way discuss the role of sleep in memory formation and optimization (research in this field, in authors’ opinion, could lay ground for breakthroughs in understanding high-level memory systems). Secondly, it is also regrettable that the concepts of the spacing effect and spaced repetition in learning have not as yet attracted much attention of researchers dealing with neuronal and molecular aspects of memory. The fact that cognitive and behavioral memory research seems to be run in disjoint research communities is thus also reflected in the book. A notable exception is a paper by Lisman et al. that abounds in behavioral parallels and illustrats the richness of potential inspiration. Similarly, the paper by Matzel and Shors, although strongly criticized in this review, illustrates the power of behavioral research in supporting or falsifying molecular and neural theses.

Besides the introduction, concluding remarks, and a somewhat scanty index, the book is divided into five sections that were clearly molded to accommodate for individual authors’ interests. As a result, the overall coverage may have lost on systematic analysis, but it has certainly gained on the in-depth creative quality of individual articles. Consequently, this is less of a student textbook, and more of a guiding light for further research in the field.

In the introduction, Christian H????pan>lscher emphasized that memory research is inevitably guided by the available methodology and historical implications. He noted the need to expand the present research beyond its traditional focus on the LTP in the hippocampus. He also noticed that one of the lesser studied aspects of memory formation is its time dynamics. Similarly, HFS protocols which induce LTP in a robust way still dominate the present research despite their being quite different from the natural firing patterns. Slicing the hippocampus deprives it from natural input from modulatory systems (one of the authors compared hippocampal slices to memory circuits in the NREM state). Similarly confounding can be the removal of inputs from the raphe nuclei or the locus coeruleus, which may act as master trigger circuits of synchronized neural activity in sleep. More precise classification of the forms of learning is needed for the analysis of the underlying neural structures as learning protocols might be associated with varying molecular mechanisms. Confusion in the protocol area may obscure the implications of experimental findings in which varying learning techniques are used to yield different results.

 Here are selected highlights of research papers included in the book 

  1. In one of the most captivating articles, Edmund T. Rolls describes the three types of neural networks underlying the function of the brain: pattern associators, autoassociators and competitive networks. Simple description of the architecture and the rules of operation of such networks are presented in the context of learning, recall and saturation forgetting. Network properties are discussed: capacity, generalization, error recovery, etc. A role of non-linear NMDA properties in forming well-defined sparse memories is hinted at. Competitive neural nets are shown to learn to categorize input pattern vectors. Their role in removing input redundancy is presented. Neural input subject to, what the author calls “sparsification”, can later be efficiently applied to associative networks. Competitive networks respond to correlations in input patterns and require no teaching stimuli. As such they seem to be an excellent model for sensory filters. Upon this neural network introduction Rolls goes on into demonstrating how this rudimentary knowledge combined with LTP/LTD and/or LTP/LTD-like mechanisms can greatly enhance our understanding of specific neural structures in the brain. He presents an extensive analysis of the hippocampal circuitry and how it can be modeled computationally with far reaching conclusions about the importance of individual network components. For example, evidence is presented that CA3 cells might operate as an autoassociative network capturing episodic memories (a conclusion that differs slightly from a similar one by Lisman et al.). Network properties combined with neurohistological data can be used to estimate network capacity. For example, rough approximation of autoassociative CA3 capacity based on the number of inputs and the estimated sparseness was found to be 36,000 separate memories. Rolls also picks examples of cortical structures whose function can be explained in terms of neural network types. For example, face recognition in the temporal visual cortex is shown as implementable as a multi-layer competitive network. Rolls successfully demonstrated the power and applicability of neural computing and computational models in bringing the understanding of the brain to new levels
  2. John Lisman at al. make an ingenious effort in bridging LTP research with behavioral data in developing their model of short-term memory named the Lisman-Idiart-Jensen model (later abbreviated to LIJ) in which the role of theta and gamma oscillations in expressing a short-term serially-searchable memory buffer is suggested. The model is then extended by an effort to pinpoint individual hippocampal structures that might be involved in short-term recall in word-list learning. Finally, the NMDA-mediated LTP is presented as a step towards consolidating long-term memories. The authors begin with a look at behavioral memory research in a historical perspective through serial recall, recency and primacy effects, and the Atkinson-Shiffrin short-memory buffer model. In the course of this analysis the authors fail to emphasize the importance of mnemonic strategies on probability of recall that has traditionally obfuscated research on short-term memory. This ultimately leads to undermining some of the evidence presented in support for their model. The authors quote research that supports the existence of short-term memory buffer in PET and fMRI cortical scans which correlate neural activity with memory loads in abstraction from attention levels. In studying the effects of amnesia on the recency effect, the authors note that hippocampal injury can selectively affect the pre-recency curve. However, they conclude that “the hippocampal region is required for LTM storage, but not STM”. This implication arrives at a time when, in support of their own model, the authors could have equally well concluded that such pre-recency curve deflection could indicate the damage to STM capacity in hippocampal injury (lower capacity entails a more pronounced recency effect). Interestingly, just a few pages later, the same authors note that surprising cases of patients have been found that showed short-digit span accompanied with unaffected long-term memory. Following this historical review, the authors introduce the Sternberg probe-recognition task. The LIJ model is presented as derived straight from the Strenberg task in which the reaction time increases linearly with the list length. In 1966, Strenberg suggested serial exhaustive scanning (SES) to explain this effect. However, the authors draw a striking parallel between the reaction time increment (of about 40 ms) in the Strenberg task and the gamma wave period at 25Hz frequency. The LIJ model draws on this parallel by replacing a spatial separation of neural patterns with a gamma-wave frequency-based temporal separation of the same patterns. The model and its implications for developing long-term memories have been analyzed by means of computer simulation. The authors draw a bold conclusion on the link between the “magic seven” and the fact that there are about seven gamma cycles per one theta cycle. However, the dependence of the STM span on knowledge representation extends “magic seven” far beyond the number of gamma subcycles achievable at frequency limits even though the involvement of cortical representations could argue for this to be a transgression beyond the strictly short-term LIJ model. The repeat time in LIJ model is determined via external theta input, while gamma oscillations arise from feedback inhibition in which currently active patterns extinguish the remaining memory traces until their own turn in the cycle. Using simulation experiments, Jenson demonstrated feasibility of such superimposed gamma-theta oscillations. Simulation also made it possible to explain the phenomenon of multiple cueing. Incidentally, the authors use a not-so-fortunate example to illustrate multiple cueing which refers to list learning. The recall of the successor of N in the alphabet does not usually come easier after the cue of LMN due to multiple cueing but rather from the fact that LMN or KLMN are, for most people, the actual recall cues while O can be concluded as the successor of N via reconstructive recall. In other words, the N-O link is not weaker. It is simply rare to store it explicitly in memory (this can be tested in each individual by simply registering the reaction time and whether reconstructive recall could be identified). In conclusion of their analysis, the authors make another bold proposition by associating hippocampal layers with highly specific neural functions. The autoassociative information is proposed to be stored in the recurrent connections of the dentate, the heteroassociative inter-item links are found in the recurrent collaterals of CA3, and the contextual input arrives at CA3 from the EC through the perforant path. Although the evidence for such a claim is behavioral, derived from neural computing, and highly speculative, the proposition itself provides a good focal point for further research that should decide on its validity
  3. Matthias H.J. Munk presents a detailed and lucid analysis of possible functions of synchronized neuronal activity in brain functions. The article begins with separating the concepts of “smart neurons” and neuronal assemblies. Synchronized neuronal responses are shown as a code for relations in input patterns. Gamma-frequency oscillations are of particular interest as they correlate with pattern recognition and synchronized responses. The role of the reticular formation in inducing synchronization is discussed. Separate roles and modes of activity of stable feed-forward connections and plastic reciprocal long-range tangential corticocortical connections are discussed. Possible role of synchronized activity in massive reorganization of cortical representation is found as of potentially monumental importance. A mechanism for such a reorganization based on the time-dependent enhancement or suppression of neural activity is outlined. The link between the cholinergic systems, attention, learning and gamma waves is analyzed, esp. in the light of determining the relative saliency and relevance of sensory input patterns. In this context, the self-organizing nature of synchronized activity is of particular interest for pattern recognition
  4. Matzel and Shors responded to the editors invitation to question the established dogmas of memory research. In particular, LTP as the prime experimental model of memory and learning has accumulated a collection of confusing research outcomes that should make the scientific community pause and reevaluate the role of LTP in learning. Most authors in the book accept LTP as a valuable research model (e.g. Abraham, Cho, Rogan and Kandel, etc.). Matzel and Shors took on a challenging task of proving LTP irrelevant in associative learning. Unfortunately, the rich body of evidence presented in their paper is highly flawed and the strongest counterevidence comes from the areas that were hardly noticed in the presented book: spacing effect and spaced repetition. Authors derive some of their evidence from the fact that “associative learning and its lasting expression does not require multiple trials” – a claim refuted evidently by the existence of the optimum spacing of repetitions in learning. In the section entitled “Associative memories can persist indefinitely” the authors quote “everyday life” evidence, “anecdotal” evidence, as well as the three most popular misconceptions related to forgetting (1) using reconstructive recall, implicit memories, and retrieval failure as a typically abductive evidence for the persistence of memories, (2) confusing the dynamics of procedural forgetting with the dynamics of declarative forgetting, and (3) using the stochastic nature of forgetting to formulate inductive theses without probabilistic evidence. Rich literature references mostly predate the currently established research trends. Similarly flawed is the comparison of the acquisition kinetics of LTP and associative learning. The dynamics of the development of synaptic potentiation will depend on stimulation protocols, which indeed in the case of LTP are often highly un-physiological. The authors write that “LTP exhibits strength through repetition which is not a defining characteristic of associative learning”. Here the volatile short-term instantaneous acquisition of memories in associative learning is confused with its long-term expression that is always trial-dependent. In contrasting the spacing effect of associative learning with the effect of inter-stimulation intervals in LTP, the authors again confuse the short-term conditions needed to establish LTP with the conditions that are suitable for retaining memories for months and years. Memory reacquisition was used as another piece of evidence. It is known that relearning takes much less effort than learning anew. Matzel and Shors note that LTP can produce reacquisition savings only if relearning occurs before LTP decays back to the baseline. This fact should actually be used to draw the opposite conclusion, i.e. in support for the value of LTP for LTM research. Long-term memories in declarative learning behave in the exactly same manner. Relearning brings benefit only within a strictly limited period of time (even if this period might last years depending on the current status of memory traces). The relearning effect, which for shorter intervals should rather be called a retrial effect, adheres to a well-defined curve with a maximum at a strictly determined point in time. In laboratory conditions, this point may range from minutes to days for LTP. Finally, Matzel and Shors indicate that non-associative induction of LTP questions its relevance in associative learning while ignoring the fact that the actual synaptic mechanism by which LTP is generated is not well understood. Even if LTP was indeed non-associative, its most important component, synaptic potentiation, has been an invaluable tool in investigating molecular cascades that result in the synthesis of proteins thought to be responsible for the strength of memory traces. No molecular biologist would question the relevance of this body of knowledge to memory and learning, consequently reaffirming the immense value of LTP as a research tool. There is little indication that associatively and non-associatively induced memories use essentially different molecular systems to justify discarding non-associative models in the study of associative ones. In conclusion, the authors state that monothematic preoccupation with LTP may deter efforts to elucidate other mechanisms that might be better suited to subserve associative memory. It is true that creative molds imposed by peer-review process, memetic nature of human communication, and historical implications of scientific research process should always be carefully taken into consideration; however, to facilitate the suggested departure from LTP the authors will have to equip memory researchers with alternative experimental models, which have not therein been submitted
  5. Rogan et al. (incl. Eric D. Kandel) discuss fear conditioning and the LTP in the amygdala as a valuable research model for memory formation. The authors stress the associative nature of fear conditioning with wide implication for learning in general. Fear conditioning circuits are presented as well as their study with the use of single unit recording and fMRI. Methods for evoking experimental LTP in amygdala are presented. Research into possible molecular substrates of amygdaloid LTP is discussed. The authors conclude that unlike the hippocampus which seems to be a well-studied structure with a hazy relationship to overt behavior, the amygdala seems to be one of the most promising areas with well-defined link to behavioral variables
  6. Stephen Maren compares the hippocampal and amygdaloid LTP in the context of emotional learning. The author notes that Pavlovian fear conditioning is robust and rapidly acquired in a similar way in various mammals, and indeed it has formed the core of the associative learning theory. Those characteristics make it an ideal candidate for studying time-dependent memory processes such as memory consolidation. The basolateral amygdaloid complex (BLA) is presented as the primary sensory interface of the amygdala, while the central nucleus of amygdala (CEA) is depicted as the final common pathway for the generation of acquired fear responses. Consequently, BLA makes up the system of sensory convergence while CEA a system of executive divergence in fear conditioning. The role of the hippocampus in the same learning tasks is then discussed with a clear indication of its contextual and temporally-limited role. The hippocampus is seen as complementing the amygdala as an accessory decoding unit without actually taking an intimate part in the learned reflex
  7. Wickliffe C. Abraham analyses differences in LTP in layers of the EC-hippocampus circuit and in the neocortex, and how these could be involved in autoassociation, decoding steps for complex memories, and the development of sparse memory representation characteristic of higher learning. The role of hippocampal place cells is discussed as well as the possible role of hard and soft synapses. The paper includes a presentation and discussion of the three families of LTP decay curves for dentate gyrus cells
  8. Sabrina Davis at al. discuss the role of gene activation in investigating synaptic plasticity. The article begins with presenting the behavioral evidence for the role of LTP in learning (e.g. McNaughton’s dentate gyrus saturation experiments, AP5 antagonist experiments, measurements of molecular learning correlates, etc.), and the current understanding of the way memories are encoded in distributed networks. The role of LTD in preventing memory overload is stressed. Then the series of molecular events leading to memory formation is described. From calcium influx to the protein synthesis in 3-6 hours window with protein delivery via synaptic tagging. A longer section is devoted to the characterization of immediate early genes activated in learning and the resulting upregulation of protein synthesis (CaMKII, syntaxin, synapsin, NMDA subunits, etc.). Finally a separate section is devoted to gene deletion technique in developing a wide range of knock-out mice deprived of specific genes thought to play a role in learning
  9. Donald P. Cain presents a discussion of recent research on spatial learning with the water maze, esp. the unclear role of NMDA-dependent LTP. The article emphasizes the importance of the classification of task difficulty as essential for comparing data from various labs, as well as detailed behavioral analysis of individual tasks. Only by thus minimizing the number of variables in the equation can the physiology of spatial learning be elucidated and the role of NMDA stated as either central or rather related to non-spatial functions
  10. Paul F. Chapman presents a detail analysis of genetic research techniques applicable to studying LTP and learning with an emphasis on a recent flurry of new data obtained with transgenic mice, as well as the promising future applications of gene-altering technology
  11. Christian H????pan>lscher, the chief editor, proposes a novel stimulation protocol able to reliably induce LTP that is closer to natural activity of the living brain in contrast to HFS frequently used in LTP research. Superiority of the new protocol is demonstrated by a better match in the profile of glutamate receptor agonists effect on LTP and actual learning tasks. The status quo of research methodology is analyzed. An outline of desirable yet realistic prospects for improvement looming on the horizon is presented
  12. Kathryn J. Jeffery suggests how we can circumvent our inability to record the input onto individual neurons by studying cell responses to strictly defined environmental stimuli that result in learning. The role of the hippocampus in representing space and in spatial learning is discussed
  13. In a refreshing departure from the standard mold, Neil McNaughton (do not confuse with Bruce McNaughton from the University of Arizona) presents his unorthodox model of the hippocampal function which is based on “preventing rather than promoting the formation of connections within the brain”. McNaughton lists a dozen of theories on the actual function of the hippocampus and shows that they cannot explain why amnesics show little recovery from memory interference in paired-associate learning. Instead McNaughton proposes his own model in which the hippocampus works to detect negative associations (associations of conflict). The model in part entails a controversial separation of the program from its data in a neural network along a far-reaching metaphor derived from linear digital computers. In conclusion, McNaughton compares hippocampal LTP to programming but still insists on tagging his model as nonmemorial


Neuronal Mechanisms of Memory Formation provides an excellent review of the current state of research into memory formation and will greatly satisfy both researchers in the field as well as students with a particular interest in memory and learning. At the same time, with minor exceptions (e.g. as found in papers by Munk, Rolls or Lisman), the book cannot claim to take a wide inter-disciplinary effort of holistically exploring lesser tested hypotheses and models of memory and cognition.

In concluding remarks, the editors note that “it is possible that LTP is only one of many memory mechanisms used in the brain”. We suggest that this conclusion should be reformulated to: “LTP is only one of many expressions of the same memory mechanism used in the brain”. Once the common denominator is understood, the confusion coming from the effort to equate LTP with memory mechanisms should be cleared out