Gerald M.Edelmann
A UNIVERSE OF CONSCIOUSNESS
Basic 2000

Nonrepresentational Memory 
pg. 93
 
 
Memory is a central component of the brain mechanisms that lead to consciousness. 
It is commonly assumed tbat memory involves the inscription and storage of information, but what is stored? Is it a coded message? When it is "read out" or recovered, is it unchanged? These questions point to the widespread assumption that what is stored is some kind of representation. This chapter takes the opposite viewpoint, consistent with a selectionist approach, that memory is nonrepresentational. We see memory as the ability of a dynamic system that is molded by selection and exbibits degeneracy to repeat or suppress a mental or physical act. This novel view of memory is illustrated with a geological comparison; memory is more like the melting and refreezing of a glacier than it is like an inscription on a rock.
______________________________________________________

We have argued that the brain is not organized like a computer, that its functioning rests instead on such properties as variability, differential amplification, degeneracy, and value.

But if the brain is not like a computer, how does memory work? 

There is a widespread assumption that, at least in its cognitive functions, the brain is fundamentally concerned with representations and that what is stored in memory is also some sort of representation. In this view, memory is the more or less perrnanent laying down of changes that, when appropriately addressed, can recapture a representation and, if necessary, act on it. 

In this view, learned acts are themselves the consequences of representations that store definite procedures or codes.

The idea that representational memory occurs in the brain carries with it a huge burden. Although it allows an easy analogy to informational transactions embedded in computers by humans, that analogy presents more problems than it solves.

In the case of humans working with computers, semantic operations occurring in the human operator's brain, not in the computer, are necessary to make sense of the coded syntactical strings that are stored physically in the computer. Coherence must be maintained in the code (or, if not, error correction is required), and the memory capacity of the system is naturally expressed in terms of storage limits. Above all, the input to a computer must itself be coded in an unambiguous fashion.

The problem the brain confronts is that signals from the world do not generally represent a coded input. Instead, they are potentially ambiguous, are context-dependent, and are not necessarily adorned by prior judgments of their significance.' 

The signals entering the eye of an animal in the jungle­patches of green and overlapping browns and of movements in the wind­can be combined in countless ways. An animal must nevertheless categorize these signals for its own adaptive purposes, whether in perception or in memory, and somehow it must associate this categorization with previous experiences of the same kinds of signals. In the case of humans, we would most likely report seeing "trees." 
Because of the enormous number of changeable combinations, to do so with a coded or replicative storage system would require endless error correction and a precision at least as great or greater than that of computers. There is no evidence, however, that the structure of the brain could support such capabilities directly; neurons do not do precise floating-point arithmetic. Such mathematical capabilities are not directly represented in brains but have arisen in human culture as a consequence of linguistic interactions and the application of logic,all, of course, because we have brains.

Representation implies symbolic activity, an activity that is certainly at the center of our semantic and syntactical language skills. It is no wonder that in thinking about how the brain can repeat a performance­that it can, for example, call up what may appear to be an image already experienced­ we are tempted to say that the brain represents. 
The flaws in yielding to this temptation, however, are obvious: There is no precoded message in the signal, no structures capable of the high-precision storage of a code, no judge in nature to provide decisions on alternative patterns, and no homunculus in the head to read a message. For these reasons, memory in the brain cannot be representational in the same way as it is in our devices.
 
How, then, can one conceive of a nonrepresentational memory? 
An analogy will help. Consider the immune system. An antibody is not a representation of a foreign antigen, yet through the system of immunological memory, it and other antibodies can recogruze that antigen. An animal can be well adapted to an environment but is not a representation of that environment. Similarly, a memory is not a representation; it is a reflection of how the brain has changed its dynamics in a way that allows the repetition of a performance.

In a complex brain, memory results from the selective matching that occurs between ongoing, distrituted neural activity and various signals coming from the world, the body, and the brain itself. 

The synaptic alterations that ensue affect the future responses of the individual brain to similar or different signals. These changes are reflected in the ability to repeat a mental or physical act after some time despite a changing context, for example, in recalling an image. 
It is important to indicate that by the word act, we mean any ordered sequence of brain activities in a domain of perception, movement, or speech that, in time, leads to a particular neural output. We stress repetition after some time in this definition because it is the ability to re-create an act separated by a certain duration from the original signal set that is characteristic of memory. And in mentioning a changing context, we pay heed to a key property of memory in the brain: that it is, in some sense, a form of constructive recategorization during ongoing experience, rather than a precise replication of a previous sequence of events.
 
GLOBAL MAPPINGS
The cerebral cortex alone is not sufficient to bear the burden of perceptual categorization and control of movement. That burden is carried out, according to the theory of neuronal group selection (TNGS), by a structure called a global mapping.(see figure 8.1). A global mapping relates an animal's movement and changing sensory input to the action of the hippocampus, basal ganglia, and cerebellum as they connect to the cerebral cortex. 
 
It links the first two topological arrangements of anatomy, the thalamocortical system and the subcortical appendages that we considered in chapter 4. 
A global mapping is thus a dynamic structure containing multiple reentrant local maps (both motor and sensory) that interact with nonmapped regions, such as those of the brain stem, basal ganglia, hippocampus, and parts of the cerebellum. The activity of a global mapping reflects the fact that perception generally depends on and leads to action. When one moves one's head to follow a moving target, the motor and sensory portions of a global mapping continually readjust. 
In other words, categorization does not occur solely in a cortical sensory area that then executes a program to activate a motor output. Instead, the results of continual motor activity are considered to be an essential part of perceptual categorization. 
This consideration implies that the global mappings carrying out such categorization must contain both sensory and motor elements. 
Neuronal group selection in global mappings occurs in a dynamic loop that continually matches gesture and posture to several kinds of sensory signals. In other words, the dynamic structure of a global mapping is maintained, refreshed, and altered by continual motor activity and rehearsal.
 

FIGURE 8.1 DIAGRAM OF A GLOBAL MAPPING. This structure is made up of multiple brain mops. These mops are connected to subcortical appendages, such as the hippocampus, basal ganglia, and cerebellum. Note that signals from the outside world enter into this mopping and tRat outputs lead to movement. This movement, in turn, aiters how sensory signals are picked up. A global mapping is thus a dynamic structure, one tbot changes with time and behavior. Its reentrant local maps, which correlate features and movement! make perceptual categorisation possible.
 
Global mappings provide a necessary substrate for relating categorization to memory. This relationship cannot generally be accounted for by the activity of any one small neural region, for, by their nature, global mappings must inclnde large portions of the nervous svstem. 
Within a global mapping, long-term changes in synaptic strengths tend to favor the mutual reentrant activity of those groups whose activity has been correlated across different maps during past behavior. When we prepare to grasp a glass and drink for example, a whole set of different circuits, modified by previous synaptic changes, are called up. 
 
Such synaptic changes over large portions of a global mapping provide the basis for memory, but memory in global mappings is not a store of fixed or coded attributes to be called up and assembled in a replicative fashion as in a computer. Instead, memory results from a process of continual recategorization, which, by its nature, must be procedural and involves continual motor activity leading to the ability to repeat a performance­in our example, grasping a glass. The ongoing synaptic changes in global mappings that occur as a result of such rehearsals favor degenerate sets of pathways with similar outputs. The contribution of global mappings to memory also carries the major burden of unconscious performance in the brain. In chapter 14 we discuss how such unconscious activity can be linked to the processes responsible for consciousness.
 
 
MEMORY AND SELECTION
What characteristics of the brain give rise to a dynamic memory without coded representation. We believe the answer is just those characteristics that one would find in a selectional system. These characteristics are a set of degenerate neural circuits making up a diverse repertoire, a means of changing the synaptic populations upon receiving various input signals, and a set of valne constraints that increase the likelihood of repetition of an adaptive or rewarding output regardless of which degenerate circuit is used. 
Given these constraints, signals from the world or from other parts of the brain act to select certain circuits from the enormously various combinatorial possibilities available. Selection occurs at the level of synapses through alteration of their effficacy or strengths. Which particolar synapses are altered depends on previous experience, as well as on the combined activities of the ascending value systems that we mentioned before (the locus coeruleus, raphe nucleus, cholinergic nuclei, and so forth).
Thus, the triggering of any set of circuits that results in a set of output responses sufficientlv similar to those that were previously adaptive provides the basis for a repeated mental act or physical performance. In this view, a memory is dynamically generated from the activity of certain selected subsets of circuits. These subsets are degenerate: A comparison would indicate that different subsets contain circuits that are not the same; nevertheless, activation of any of them can result in a repetition of some particular output. 
 
Under these conditions, a given memory cannot be identified uniquely with any single specific set of synaptic changes because the particular synaptic changes associated with a given output and eventually with an entire performance are subject to further change during that performance. So what is called forth when an act is repeated must be any one or more of the neural response patterns adequate to that performance, not some singular sequence or specific detail.
We see that synaptic change is fundamental and essential for memory but is not identical to it. There is no code, only a changing set of circuits corresponding to a given output. 
The more or less equally effective members of that set of circuits can have widely varying structures. It is this property of degeneracy in neural circuits that allows for changes in particular memories as new experiences and changes in context occor. Memory in a degenerate selectional system is recategorical, not strictly replicative. 
There is no prior set of determinant codes governing the categories of memory, only the previous population structure of the network, the state of the value systems, and the physical acts carried out at a given moment. The dynamic changes linking one set of circuits to another within the enormously varied neuroanatomical repertoires of the brain allow it to create a memory. The probability of creating a memory is enhanced by the activity of value systems.
In our example of reaching for a glass, the satisfaction of thirst will activate value systems and lead to the selection of a number of circuits appropriate for performing that action. By these means, structurally different circuits within the degenerate repertoires are each able to produce a similar output, leading tO repetition or variation of the act of reaching. Their activity gives rise to the associative properties of memon; for example, an act can trigger another act, a word can trigger other words, or an image can provoke a narrative. These associative properties arise materially from the fact that each different member of the degenerate set of circuits used at different times has different alternative network connections.
In this view, there are hundreds, if not thousands, of separate memory systems in the brain. These systems range from all the perceptual systems in different modalities­sight, smell, touch, and so on­to the systems that govern intended or actual movement, to the language systems that organize speech sounds. 
 
This view is compatible with various types of memory described and tested by experimentalists in the field­so-called procedural, semantic, episodic memories, and the like­but it does not restrict itself to these types, which are defined mainly by operational criteria and, to some degree, by biochemical criteria.
Although such individual memory systems differ, the key conclusion is that whatever its forrn, memory itself is a system property. It cannot be equated exclusively with circuitrv, with synaptic changes, with biochemistry, with value constraints, or with behavioral dynamics. Instead, it is the dynamic result of the interactions of all these factors acting together, serving to select an output that repeats a performance or an act. 
The overall characteristics of a particular performance may be similar to a previous performance, but the ensembles of neurons underlying any two similar performances at different times can be and usually are different. This property ensures that one can repeat the same act, despite remarkable changes in background and context, with ongoing experience.
Besides guaranteeing association, the property of degeneracy also gives rise to the remarkable stability of memorial performance. In a degenerate system, there are large numbers of ways of assuring a given output. As long as a sufficient population of subsets of circuits remains to give an output, neither cell death nor changes in a particular circuit or two nor switches in the contextual aspects of input signals will generally be sufficient to extirpate a memory. Thus, nonrepresentational memory is extraordinarily robust.
 
 
 
 
 
HOME