Reprints 121-144

Date:
1992-2001
Reference:
PP/CRI/M/1/7
Catalogue details

Licence: In copyright

Credit: Reprints 121-144. Source: Wellcome Collection.

    38/284
    dreams are more frequent, more hallucinoid and (if the subject is woken abruptly) more vivid than such dreams as are reported in non-REM. not remembered. The observed facts are that if a person is allowed to sleep freely they report rather few dreams, especially if they wake infrequently in the night. On the other hand, if a person is constantly awoken during REM, they report a relatively enormous amount of dreaming. These facts are usually interpreted along the following lines. It is assumed that, during REM, immediate or very short-term memory storage operates, but that the mecha nism for inputting long-term memory is temporarily inac tive. On awakening, this latter mechanism slowly recovers. Thus, most dreams are necessarily forgotten, since they This interpretation, if correct, leads to the following important conclusion. It is unlikely that the biological function of REM is related to remembered dreams. That function, whatever it is, is more likely to be related to our host of unremembered dreams. There is no evidence to suggest that remembered dreams are anything more than an accidental by-product of this function. The characteristic features of REM dreams could be • called 'bizarre intrusions'. These are brief incidents that often appear to arise for no apparent reason. It has long been believed that the content of these intrusions is often related to day residue, that is, to incidents occurring in the last day or so and especially pertaining to those matters which the dreamer has had on his or her mind. However, it is important to notice that one rarely dreams of a pre vious event in correct detail. More typically, the intrusion consists of a mixture of features, all or most of which can be related to events which have occurred recently. These bizarre intrusions seem to follow at short intervals, per haps every second or so. The other main feature of the dream appears to be the narrative, which possesses greater continuity than the in trusions. It is as if a part of the brain is trying to make some sense of the bizarre intrusions, as indeed one would if such events happened while one was awake. The nar rative often has a particular emotional tone (erotic, filled with anxiety, etc.) which one suspects is related to other causes. It is unclear whether the first few intrusions set up the narrative or whether the narrative exists in some latent form before the first intrusion. When one speaks of 'a dream' one is usually referring to a single fairly continu ous narrative. We tend to regard such mentation as a separate dream if it has a distinct narrative. As will be seen, our theory provides a good explanation of the nature of the bizarre intrusions. It has nothing useful to say about the narrative. It has been known for many years that during REM sleep a series of impulses, called PGO (ponto-geniculo- occipital) waves, appear in the brain. These impulses origi nate in the pons and spread to the neocortex via the thala mus. These waves are very frequent, in the cat, for example, as many as 16,000 per day. Moreover, they affect most cortical neurons in the sensory and motor areas. Hobson and McCarley [ 12] have suggested that in part of the pons there is a dream state generator which produces these impulses. They postulate that it is these impulses which provide the driving force for REM dreams. Our views are based on their Activation-Synthesis hypothesis. We can summarize the main points of this Section as follows. (1) REM sleep performs some important biological func tion for most higher vertebrates. (2) This function is likely to benefit the developing animal as well as the mature animal. (3) While REM dreams may give some clues about this function, it is unlikely that remembered dreams have, in themselves, any major biological use, since the majority of dreams are unremembered. (4) REM dreams are probably influenced by the PGO 2. Memory storage and neural nets It seems likely that any memory system can usefully be described under these headings: (1) putting the pattern to be remembered into the system; (2) storing it over time; (3) accessing the system in order to recover it. While we have no solidly established theory of human memory, the following rather general account is at least plausible. It is assumed that the first and last processes above - that is, inputting and accessing - necessarily in volve neurons firing. For long-term memory storage it is assumed that neuronal firing is not required and that the memory is stored in some semi-permanent modmcation to parts of neurons and especially to synapses. The exact neurological basis of very short-term memory (immediate It is widely believed that the operation of the brain is radically different from the operation of a modern digital computer. The latter uses accurately pulse-coded mes sages. Information is encoded in the pattern of 0s and Is sent out at regular time intervals. This enables a particu lar message to be sent to a particular 'address' where information can be put in, stored and accessed. There appears to be no sign of such a system in the brain. Instead, the information a neuron sends out appears mainly to be contained in a somewhat irregular pattern of spikes in its axon, probably encoded as the average firing rate, although there may be some information in the firing of each spike relative to those from other relevant neurons. This makes it likely that such a system cannot send pre- Instead, it is generally thought that memory in the brain is 'content addressable' (see below). Another important difference between most neurons in the brain and the transistors in a computer is that most neurons receive input from very many different sources and each sends its output to many other neurons. For this reason theorists believe that the operation of the brain can only usefully be modeled by systems in which many units (a unit is an idealized neuron) interact with each other. Such studies have led to the idea of neural nets, in which many units act in parallel and which, in some models, connect back on all the other units in the net. More elabo- together in various ways. For a general introduction, see [17]. These studies show that patterns of activity can be stored in the strengths of the connections (the synapses) between all the different units. Thus, a single associative net, suitability adjusted, if given a pattern of activity A (in which some inputs are firing and some are inactive), will produce as its output a different pattern of activity A'. If a sufficiently large part of the pattern A is used as input, it will output the whole of pattern A'. It is the latter be haviour which is described as 'content addressable', since fragments of the memory can be used as an address to call up the total memory. One can ask where the memory is stored in such sys tems. As we have said, it is stored mainly in the synaptic weights-the strength of all the connections between the neurons. (Notice if there are n neurons in a net, there are likely to be n 2 synapses between them.) Memory is not stored exclusively in one or even a few connections, but in the net as a whole. That is, the memory is 'distributed'. Moreover, if we make a few random alterations in the weights, this usually makes little difference to perform ance, primarily because each unit has a threshold (as does a neuron). The unit is non-linear, in that it will not fire or only fire at a very low rate, if its effective input is below this threshold. Thus, the system is 'robust'. However, most nets have a further remarkable property. Let us assume that the synapses have been adjusted (and there are usually simple rules of how this should be done) so that with an input A it produces an output A'. Now let us make further adjustments to the system so that a dif ferent input B will produce some other output B'. Then, if the net is sufficiently large and if A and B (and A' and B ' ) are sufficiently distinct, it is found that the system now can do both jobs. An input A produces A' ; input B pro duces B '. In both cases the information is contained in the strengths of all synapses, but the information for the two associations (A with A' and B with B') is superimposed, so that any one synapse is likely to contribute to both associations. This behaviour is not limited to just a pair of associations. Further ones (C with C, D with D', etc.) can be added. The reader will readily surmise that associations cannot be added indefinitely. For any net there is a limit to what it can store before the system begins to misbehave. Broadly speaking, the larger the net and the more distinct the pat terns are from each other, the more associations can be When a particular net is overloaded it usually misbe haves in a special way. Instead of outputting the required stored pattern x', it may produce a pattern which can be seen to be a mixture of several of its stored patterns. This is especially likely to happen if the patterns are not totally distinct but have some parts in common. To summarize, neural nets can store information in a way which is distributed, robust and superimposed. When overloaded, such nets often output a pattern which is a mixture of some of its stored patterns, especially if the patterns are somewhat related. It is important to realize that these properties have not been directly imposed on the net by its designer. They are emergent properties of this general type of memory stor age system. Whether these models are significantly simi lar to the arrangements of neurons in the brain remains to be seen, but they appear to be consistent with much that is known, in a broad way, about the neuroanatomy, the neurophysiology and the overall behaviour of the brain, although in detail the present generation of models is almost certainly highly oversimplified. It should not be thought that we believe the brain is merely a set of simple 3. Reverse learning The process of reverse learning is designed to make the storage in an associative net more efficient. The hope is that this process will reduce somewhat the mixed outputs produced by overlapping memories, while leaving intact the unmixed memories which the net was supposed to store. It may also help to remove inappropriate connec tions made by the somewhat random nature of neural development. To understand what has been proposed, the reader must grasp the idea behind the normal storage of memories. Let us take a simple example. Consider a pattern of inputs, but for simplicity let us consider only three of the units par ticipating in this input. Let us assume that in this particular
    digitised image 1
    digitised image 2
    digitised image 3