|
From Complexity to Creativity -- Copyright Plenum Press, © 1997 |
Part III. Mathematical Structures in the Mind
CHAPTER 10. DREAM DYNAMICS
CHAPTER TEN
DREAM DYNAMICS
Scientifically as well as experientially, dreaming is a confusing phenomenon. It is known to be necessary for healthy functioning, in the sense that dream deprivation leads to harmful effects. Yet the biological and psychological reasons for this necessity are largely unknown.
Clinical psychologists, since well before Freud, have believed dreams to hold deep secrets about the unconscious mind. Jung believed even more: that they were a doorway to the collective unconscious. The importance of dreams is affirmed by many ancient wisdom traditions. On the other hand, recent theorists have proposed that the contents of dreams are just "random noise," generated by primitive parts of the brain.
In recent years, however, it has become possible to simulate some aspects of dreaming behavior using formal neural networks. In this chapter I will review some of this work -- the Crick- Mitchison hypothesis and its extensions in the work of George Christos. I will discuss the shortcomings of this work, and propose an alternate approach based on the dynamics of mental processes. The new approach proposed here salvages many of the ideas of the neural network approach, while also tying in neatly with the older, content-oriented theories of dreams.
10.2 The Crick-Mitchison Hypothesis
In 1983 Crick and Mitchison coined the catch-phrase, "We dream in order to forget." This audacious statement, now called the Crick-Mitchison hypothesis, was primarily formulated not on the basis of neurological experiments, but on the basis of experience with a simple formal network model called the Hopfield net (Hopfield, 1985). Crick and Mitchison observed that Hopfield nets, when used as associative memories, quickly become overloaded with useless old memories. Why, they asked, does the human brain not suffer from the same problem? The answer they arrived at was: because the human brain dreams.
In a 1986 paper, Crick and Mitchison retracted their original statement, replacing it with the weaker "We dream to reduce fantasy and obsession." However, some authors, such as Christos (1994), feel that the original Crick-Mitchison formulation was more accurate. Computer simulations by Christos (1992) and others (Nadel et al, 1986; Parisi, 1986) give partial support to the original hypothesis. By passing a Hopfield net through alternate phases of learning and forgetting, one does indeed solve the problem of overloading, though at the cost ofdrastically reducing the memory capacity of the network.
Neural network models of dreaming are typically interpreted to contradict the psychoanalytic account of dreams, according to which dreams serve deep emotional functions (Freud, 1900). I will argue, however, that there is no such contradiction. If one transplants the idea of "dreaming as forgetting" to the level of neuronal groups instead of neurons -- i.e., the level of complex process dynamics instead of merely switching network dynamics -- the apparent contradiction disappears. One obtains a theory that is quite consistent with the Freudian view of dreams, as well as being, in some respects, less conceptually problematic than the original Hopfield net theory. This new and improved theory of dreaming behavior, while speculative, is conceptually quite satisfying. By invoking a "process dynamics" view of the mind, it closes the gap between a physicalistic model of dreams (the Hopfield net model) and a mentalistic model of dreams (the Freudian model).
10.3 TESTING THE CRICK-MITCHISON HYPOTHESIS
According to the most reasonable estimates, the brain takes in about 100 billion billion bits of information during an average lifetime. But the brain can store only 100 thousand billion bits at a time. Clearly a great deal of information compression is going on in the brain, so that these numbers are not strictly comparable. One cannot justifiably conclude that only one out of every million pieces of information can be "permanently" retained. But nevertheless, it seems clear that the brain has a large incentive for forgetting irrelevant information. A very important part of learning and remembering is knowing what to forget.
But how does the brain do this forgetting? Crick and Mitchison were the first to suggest that dreaming might play a role in this process. To see the inspiration for this idea one must think a little about the limitations of Hopfield nets as associative memories.
Hopfield Nets
The Hopfield net is perhaps the simplest formal neural network model of any practical use (Hopfield, 1985). It is a natural elaboration of the original neural network model of McCullough and Pitts (1943). In the simplest case, each neuron has a state of either 1 or -1, and each neuron accepts input, potentially, from all other neurons. At each time step, each neuron sums up all its inputs, arriving at a certain "internal sum." The input which neuron i receives from neuron j is multiplied by amount wij before it is added onto the internal sum of neuron i. Then, after it has constructed its internal sum, a neuron tests whether this sum is over its threshold or not. If so, it sends out charge to all the other neurons. If not, it doesn't. In the simplest model, before the neuron fires again, it wipes its memory clean, sets its internal sum to zero.
The firing of neurons is the dynamics of the network. Thisdynamics may be managed in at least two different ways: asynchronously, whereby the neurons fire one after the other, in some random order; or synchronously, whereby all neurons fire at the same time. Under either scheme, the dynamics may involve a variety of complex circuits. In many applications, however, the weights are taken symmetric (wij = wji), which guarantees that, from any initial condition, the network will eventually converge to a fixed point attractor (this is not obvious, but may be shown by algebraic manipulations). Relaxing the symmetry conditions gives the full spectrum of dynamical behavior -- periodic points, limit cycles, etc.
To view a Hopfield network as an associative memory, one assumes one is given a series of "memories" A1,...,AM, each one of which is presented as a vector of -1's and 1's. The synaptic weights are then defined in terms of these memories. To find the correct wij, for each pair (i,j) of potentially interconnected neurons, one sums up the products Aki Akj for all the memories vectors Ak. If all goes well, then this choice of weights will lead to a network which has the desired memories as fixed-point attractors.
The catch is that, like a human brain, a Hopfield net can only store so much. Once the number of items stored exceeds about 15% of the number of neurons, the memory ceases to function so efficiently, So-called "parasitic memories" emerge -- memories which are combinations of parts of real memories, and which are fallaciously associated with a number of different inputs.
So, suppose the Hopfield net in question is part of a living, learning system, which constantly needs to store new information. Then the network will need to have a way of unlearning old associations, of forgetting less crucial memories so as to make room for the new ones. This, Crick and Mitchison suggest, is the role of dreaming. Instead of adding terms Aki Akj on the synaptic weights, as one does to train the network, dreaming subtracts these products, multiplied by a suitable constant factor, from the synaptic weights.
But which memories get their contribution to the weights decreased in this manner? Why, of course, the ones that are remembered most often! If a real memory is remembered often, and one decreases its weight a little bit, this won't really hurt it, because it has a deep basin. But if a memory is remembered often simply because it is parasitic, then a slight decrease in the weight will destroy it -- because its basin is so shallow. Or so the theory goes.
Hopfield, independently of Crick and Mitchison, did some simple reverse-learning experiments with a computer-simulated network. He filled the network just slightly above its capacity, and then tried to winnow it down, by feeding it random inputs, observing its responses, and subtracting some percentage of the memories thus obtained. In order to avoid completely obliterating the memory, the percentage had to be taken very small, around 1%.
The First Christos Experiment
Intrigued by the reverse-learning approach to dreams, West Australian mathematician George Christos decided to put the Crick-Mitchison hypothesis to the test (Christos, 1992). In his first experiment, Christos simulated a repeatedly dreaming Hopfield network by "random sample analysis." At each stage, to determine what the network should dream about, he presented it with a large number of random inputs (100-200 times the number of stored memories), and observed what memories the network associated these inputs with. This statistical procedure gives a good qualitative picture of the "energy landscape" of the network. Those memories that are retrieved more often must have larger basins of attraction, and are therefore the ones that should be removed by "reverse-learning" or dreaming.
The results of this experiment were somewhat surprising. At first things work just like the Crick-Mitchison hypothesis suggests: dreaming reduces the incidence of spurious, parasitic memories. But then, as the network continues to dream, the percentage of spurious memories begins to increase once again. Before long the dreaming destroys all the stored memories, and the network responds to all inputs with false memories only.
For instance, in a network with five memories, using a weight-reduction factor of 1%, the first fifty dreams behave according to Crick-Mitchison. But by the time five hundred dreams have passed, the network is babbling nonsense. The same pattern is observed in larger networks, and with different reduction percentages. Intuitively one might think that dreaming would tend to even out the basins of attraction of the stored memories. But, as Christos put it, "intuition is not guaranteed in a nonlinear process such as this."
Dreaming, removing part of a memory "attractor," may actually cause new spurious attractors, new false memories. In this sense the Crick-Mitchison hypothesis is based on a conceptual fallacy: it is based on the idea that one can remove a single false memory from a Hopfield net, in a localistic way. But in fact, just as false memories emerge from holistic dynamics, attempts to remove false memories give rise to holistic dynamics. The reverse-learning mechanism inevitably affects the behavior of the network as a whole. Linear intuition does not apply to nonlinear dynamics.
The Second Christos Experiment
But this is not the end of the story. Human dreaming is obviously a cyclic phenomenon: we dream, then we wake, then we dream, then we wake, and so on. For his second experiment, Christos sought to simulate this process with a Hopfield net. He trained the network with memories, then put it through the reverse-learning, "forgetting" process. Then he trained it with new memories, and put it through the forgetting process again. And so on.
The experiment was a success, in the sense that the network continues to function as an associative memory long after its"fifteen percent limit" has been reached. Dreaming does, as Crick and Mitchison predicted, allow the network to function in real-time without overloading. Since the older memories are subject to more dreams, they tend to be forgotten more emphatically, making room for new memories.
The only catch is that the overall memory capacity of the network is drastically reduced. A network normally capable of storing twenty memories can now only store three or four, and even for these it has a low retrieval rate. The reason is that, just as in the first Christos experiment, dreaming actually increases the production of false memories.
So the Crick-Mitchison hypothesis is salvaged, but just barely. If one keeps the number of memories very, very small -- around .05 of the number of neurons in the network -- then the cycle of learning and forgetting is indeed effective. Given the huge size of the brain (100 billion neurons is a reasonable figure), the number .05 is not particularly worrying from the neuroscience point of view. A network with intermittent periods of dreaming and learning provides much less efficient memory than a network which is simply re-started every time it gets full, but real brains may not have the option of periodically wiping themselves, or even parts of themselves clean..
10.4 A MENTAL PROCESS NETWORK APPROACH
So, is dreaming really forgetting? Based merely on these Hopfield net simulations, this seems a rather grandiose conclusion, which is perhaps why Crick and Mitchison eventually reformulated their statement in a weaker form.
It is my contention, however, that there is actually some psychological sense to be found in the idea that dreaming is forgetting. This sense may be found by transplanting the idea from the domain of Hopfield nets to the domain of neural maps. A neural map is a network formed from neuronal groups rather than single neurons; it is thus a network of complex processes rather than a network of simple mathematical functions.
In some ways, I will argue, the "dreaming is forgetting" idea makes more sense in a neural map context than it does in terms of Hopfield nets. In the domain of neural maps, there is no problem with the forgetting of useful instead of useless memories, a fact which leads one to suspect that the memory capacity limitations of the dreaming Hopfield network may not apply to the brain. Furthermore, the transplantation of the theory to the domain of neural maps leads to a connection with the psychoanalytic theory of dreams. It explains how dreaming might accomplish both the role of forgetting and the role of highlighting or resolving emotional problems.
This account of dreaming is speculative but, in truth, no more so than the original Crick-Mitchison hypothesis, the Freudian theory, or any other current theory of dreaming behavior. The key asset of the present theory is its ability to bridge the gap between the physicalistic dynamics of neurons and the mentalistic dynamics of thoughts and emotions. The processdynamics of neuronal groups are used as a "middle ground" between these different levels of explanation.
Memory and Consciousness
One thing sorely lacking in the Crick-Mitchison hypothesis is any reference to the phenomenon of consciousness. The peculiar feature of human dreaming, after all, is that when one dreams one is conscious while sleeping. In a sense, the absence of consciousness from the Hopfield net automatically makes it inadequate as a model of dreaming behavior.
It was argued in Chapter Eight that one of the psychological roles of consciousness is the creation of memories, as wide- basined autopoietic magician systems. To understand the relation between this aspect of consciousness and dreaming, recall the classic experiments of the neurologist Wilder Penfield. Penfield noted that by stimulating a person's temporal lobe, one can cause her to relive memories of the past. Merely by touching a spot in the brain, Penfield could cause a person to have a detailed, intricate experience of walking along the beach with sand blowing in their face, or lying in bed in the morning rubbing their eyes and listening to the radio. The most remarkable aspect of these memory experiences was their detail -- the person would recite information about the pattern of the wallpaper, or the color of someone's shoes ... information which, in the ordinary course of things, no one would bother to remember. And if he touched the same spot again -- the same experience would emerge again.
Penfield's conclusion was that the brain stores far more information than it ever uses. Every moment of our life, he declared, is filed away somewhere in the brain. If we only knew how to access the relevant data, we could remember everything that every happened to us with remarkable accuracy.
Recent replications of Penfield's experiments, however, cast serious doubt on this interpretation. First of all, the phenomenon doesn't occur with everyone; maybe one in five people can be caused to relive the past through temporal lobe stimulation. And even for this select group, a careful analysis of the "relived memories" suggests that they are not exact memories at all.
Current thinking (Rosenfield, 1988) is that, while the basic frameworks of the relived memories are indeed derived from the past, the details are manufactured. They are manufactured to serve current emotional needs, and based on cues such as the places the person has visited earlier that day. This, of course, ties in perfectly with the view of consciousness in terms of a memory-creating Perceptual-Cognitive Loop.
And what distinguishes those people who are susceptible to the Penfield phenomenon? The answer, it appears, is limbic system activity. The limbic system is one of the older portions of the brain -- it is the "reptile brain" which is responsible for basic emotions like fear, rage and lust. When a susceptible person has their temporal lobe stimulated, the limbic system is activated. But for a non-susceptible person, this is not thecase; the limbic system remains uninvolved. The message seems to be that emotion is necessary for the construction of memory. This is a very Freudian idea, but nonetheless it is supported by the latest in neurological research.
Certain people, with a particularly well reinforced pathway between the limbic system and the temporal lobes, can be caused to construct detailed memories by temporal lobe stimulation. Penfield's original conclusion was that the brain stores all its experience, like a vast computer database. But the fact of the matter seems to be that the brain, quite unlike any computer database known today, holds the fragments of a memory in many different places; and to piece them together, conscious, emotional intervention is required.
Memory and Dream
So consciousness produces memories. What does this have to say about dreams?
When awake, consciousness is constrained by the goal- directed needs of the brain's perceptual-motor hierarchy. It pieces together fragments of different ideas, but always in the service of some particular objective. This means that, unless they are almost omnipresent, false memories will quickly be rejected, due to their failure to solve the practical problems at hand.
When the body is sleeping, on the other hand, consciousness can act unfettered by the needs of the perceptual and motor systems. It has much less lower-level guidance regarding the manner in which it combines fragments from the mind's associative memory. Therefore it will tend to produce false memories just as readily as real ones. In particular, it will combine things based on how well they "go together" in the associative memory, rather than how well they serve pragmatic goals.
In other words, to use the language of the psynet model, what I am suggesting is that dreaming is inefficient in the Hopfield network precisely because the Hopfield network is merely a static associative memory, rather than a dynamic associative memory coupled with an hierarchical perception/control network. In the human brain, during waking hours, the perceptual-motor network reinforces real memories due to their greater utility. During sleeping hours, dreaming decreases real and false memories alike, but as the false memories do not have so much waking reinforcement, they are eventually obliterated.
Furthermore, in the brain, old memories are constantly interacting and recalling one another. Old memories do not necessarily fade just because they are not explicitly elicited by external stimuli. This is one very important role of feedback in the brain: memories activate one another in circular networks, thus preventing the dreaming dynamic of reverse learning from chipping away at them.
Dreaming as Forgetting, Revisited
To put these abstract ideas in a more concrete context, let us turn to the psynet model. In biological terms, this means one is stepping up from individual neurons, as in the Crick-Mitchison hypothesis, to neuronal groups which represent pattern/process magicians.
In the psynet perspective, memories are autopoietic systems, so it is easy to see that many useless memories will naturally "forget themselves." Thought systems with no positive external input will often just disapppear. In biological terms, one may say that, if a certain cell assembly is of no use, it will simply not be reinforced ... it will dissolve, and other more useful assemblies will take its place.
But of course, this does not always happen. Some thought systems, like political belief systems, may proceed of their own accord; continuing to exist in perpetuity because of their own autopoiesis. Neurally speaking, one may say that, in the more abstract regions of the mind, maps which are of no objective use can sustain themselves by self-reinforcing dynamics, by map-level feedback loops. These will not be forgotten by simple neglect.
This leads us to our main idea. Suppose that dreaming served as a special dynamic for forgetting useless, self- perpetuating thought systems? This hypothesis forms a connection between the "dreaming as forgetting" idea and the psychoanalytic theory of dreams, in which dreams represent unresolved problems and anxieties. All that is needed to draw the connection is the association between neuroses or "complexes" and complex, self- perpetuating systems of neuronal groups.
While sleeping, consciousness is present, and yet it is dissociated from the ordinary external world. In place of the external world, one has a somewhat sloppily constructed simulacrum of reality. And on what principles is this simulacrum created? The main difference between dream and reality is that, in the dream-world, expectations are usually correct. Whatever the controlling thought-system foresees or speculates, actually happens. Thus the disjointed nature of dream life -- and the peculiarly satisfying nature of dreams. In dreams, we can get exactly what we want ... something that very rarely happens in reality. And we can also get precisely what we most fear. The image in the mind instantaneously transforms into the simulated perception.
In dreams, in short, thought-systems get to construct their own optimal input. This observation, though somewhat obvious, leads to a rather striking conclusion. Suppose a thought-system has evolved a circular reinforcement-structure, as a tool for survival in a hostile world -- for survival in an environment that is constantly threatening the system with destruction by not conforming with its expectations. What will be the effect on this thought-system of a simulated reality which does conform to its expectations?
This question can be answered mathematically, using the formalism of magician systems. But this is hardly necessary; the upshot is easily perceived. A thought-system, presented with acomfortable world, will let down its guard. It will relax its circular reinforcement dynamics a little -- because, in the dream world, it will no longer need them. The dream world is constructed precisely so that the thought-system will be reinforced naturally, without need for circularity. Thus dreaming acts to decrease the self-reinforcement of circular belief systems. It weakens them ... and thus, after a fashion, serves as a medium for their "forgetting."
Let's say a married man has a dream about his wife strutting down the street in a skimpy dress, swaying her hips back and forth, loudly announcing that she's looking for a good time. Meanwhile he's running after her trying to stop her, but his legs won't seem to move fast enough; they suddenly feel like they're made of glue. The interpretation is obvious enough, but what is the purpose?
According to the present model, the purpose is precisely to feed his suspicious thought-system what it wants to hear. Apparently a certain component of his mind is very mistrustful of his wife; it suspects her, if not of actually committing adultery, at least of possessing the desire to do so. This mental subsystem may have no foundation in reality, or very little foundation; it may to a great degree sustain itself. It may produce its own evidence -- causing him to interpret certain behaviors as flirtatious, to mis-read certain tones of voice, etc.
The thought system, if it is unfounded, has to be circular in order to survive daily reality. But in the dream world it is absolutely correct: she really is looking to have sex with someone else. While getting input from the dream world, then, the thought system can stop producing its own evidence. It can get used to receiving evidence from the outside world.
Temporarily, at least, the dream-world breaks up the circularity of the thought system, by removing the need for it to make up its own evidence. Whether the circularity will later restore itself is another question; but one may say that the expected amount of circularity in the system will be less than it would have been had the dream not occurred. In other words, his suspicions, having temporarily had the liberty of surviving without having to create mis-perceptions, may forego the creation of mis-perceptions for a while. And in this way, perhaps they will begin to fade away entirely. Or, alternately, the thought- system may rejuvenate itself immediately, in which case he can expect to have similar dreams again and again and again.
The big question raised by the second Christos experiment is, how could a neural network be configured to tell real memories from false ones? How could one get it to "subtract off" only the bad guys? If this problem were solved, then the extremely low storage capacity of dreaming Hopfield nets would be solved. But in the mental process network theory, there is no problem with this. If a useful thought system is subjected to dream treatment, and given a simulated reality which fulfills all its expectations, it will not be jarred as much as a lousythought-system, which relies more centrally on self-reinforcement for its survival. Its dream world will not be as drastically different from the real world. It will get in the habit of not producing its own evidence ... but then, it never had this habit to an excessive degree in the first place.
Let's say the same married man mentioned above has a dream about his wife kissing him passionately and giving him a new gold watch with a huge, fist-sized diamond on it. This is produced by the thought-system that knows his wife loves him and treats him well. It temporarily permits this thought system to relax its defenses, and stop actively "twisting" reality to suit its preconceptions. But in fact, this thought system never twisted reality nearly so much as its suspicious competitor. So the dream has little power to change its condition. Dreams will, according to this line of reasoning, affect predominantly circular belief systems much more drastically than they will affect those belief systems which achieve survival mainly through interaction with the outside world.
And what about the role of the body in dreaming? One must recall that perception and action are inseparable -- they are carried out by a unified perceptual-motor hierarchy. Or, in the aphorism of Heinz von Foerster, "If you want to see, learn how to act." To fool a thought-system into thinking it is perceiving its ideal environment, one must often feed it perceptions that involve actions as well.
It is disappointing to realize that the Hopfield net is too unstructured to serve as a useful psychological model of dreaming. For, after all, the Hopfield network is so easy to analyze formally, and to simulate on the computer. When one starts talking about consciousness and inter-reinforcing networks of memories, things become much murkier, much less elegant and mathematical. But on the other hand, perhaps that's just the way mind is -- ....
One concrete, if speculative, application of these ideas is to the phenomenon of Sudden Infant Death Syndrome (also known as SIDS, or simply "crib death" or "cot death"). Nearly two in every thousand births result in a death classified as SIDS. A huge amount of biological data has been gathered regarding this "disease," but even so, researchers have been unable to arrive at a good explanation. George Christos has proposed an intriguing explanation based on dreaming.
It has been suggested that respiratory obstruction is the culprit, or else reduced circulation in the neck arteries. But these theories fail to account for the observed information: why, then, do infants sleeping on their back also die from SIDS? George Christos (1992a) has suggested that the flaw is not in the data but rather in the physicalistic bias of medical researchers. He suggests that SIDS is fundamentally a psychological problem, the understanding of which requires an analysis of the infantmind.
During ordinary, non-dreaming sleep, the brain and the body are fairly well "disconnected." But during dreaming sleep, this disconnection becomes more extreme. This fact has long been used by researchers to study the effect of dream deprivation in cats: the cats are placed on tiny steeply sloped "islands" in a room full of water. They can sleep on the islands, but when they sink into dreaming sleep, their muscles relax even further and they slip into the water.
The results of the failure of this brain-body disconnection are well-known. Walking and talking in one's sleep are two prime examples. One does not want the brain hooked up to the body during dream sleep; otherwise the dreams will control the body's motions, resulting in bizarre and dangerous activity.
But even during dreaming sleep, however, some degree of brain-body connection remains. Active contact is maintained with the muscles regulating the eyes, the heart and the lungs. And it is these connections, Christos claims, which are responsible for Sudden Infant Death Syndrome.
In the Stanford University sleep lab, an experimental subject was observed to actually hold his breath while dreaming about swimming underwater. Similarly, Christos proposes, SIDS occurs when infants dream about being in the womb and, consequently forget about breathing.
We dream about our past experiences. But what past experiences has an infant had? It is almost inarguable that, when an infant dreams during its first few weeks after birth, it dreams about the first nine months of its life: its life as a fetus. And this is eminently sensible in terms of the dreaming- as-forgetting hypothesis, since an infant really does need to unlearn its memories of the womb. They are not very useful in its new, independent life.
But if the body does tend to "enact" the contents of its dreams as much possible, then when the infant dreams about the womb, its body will re-enact its fetal state. And one distinctive quality of the fetal state is the fact that breathing, in the usual sense, does not occur. The hypothesis is then that cot death is caused by infants dreaming that they are back in the womb and consequently forgetting to breathe.
This simple idea, Christos argues, explains all the circumstantial data regarding cot death: the exponentially decaying age at death curve, the higher risk in the prone sleeping position, and the climatic variation in incidence.
SIDS infants also demonstrate a greater amount of dreaming sleep than other infants, and a reduced motility during dreaming sleep. Furthermore, most SIDS deaths occur in the early morning, which is precisely the time of the longest dreaming sleep, corresponding to the most intense dreams. SIDS infants demonstrate a higher than average heart rate, meaning that their heart rate is closer to the fetal heart rate. And, most significantly of all, SIDS infants who survive past one month demonstrate a higher heart rate only during dreaming sleep.
There is, in this view, no cure for SIDS -- the best that a parent can do is to make their infant's life as "un-fetus-like"as possible. Avoid wrapping the baby tightly, and put her to sleep face up. This is not a terribly comforting view of crib death, and it does not provide funding agencies with much cause to support research work on the alleviation of crib death. Instead, it suggests, funds would be better spent on research into the biological and psychological foundations of dreaming.
SIDS and the Mental Process Network
Given the mental-process-network theory of dreams outlined above, one can easily form a plausible story of the psychological events underlying Christos's proposed mechanism. When the infant emerges from the womb, her neural patterns are still configured for womb life. She has developed a rudimentary "belief system" for dealing with life as a fetus. This system must be destroyed, or at least tranquilized, as quickly as possible. The tool for this destruction is the dream: dreams of fetal bliss fool the womb-life thought-system into complacency, making it easier for the new, outside-world-life thought-systems to grow and prosper.
But in order to most effectively fool the old thought- systems into thinking they're in their ideal environment, the womb, certain actions must also be re-enacted (after all, the thought-system perceives what the body is doing, and will recognize if something isn't right). Sudden infant death occurs because this process gets out of hand -- in other words, it is a consequence of overenthusiasm for life outside the womb, of an attempt at overly rapid adjustment.
This is, of course, a speculation -- as are all other proposed explanations of SIDS, at this stage. But, at very least, it is a nice illustration of how the dreaming is forgetting idea can be lifted from the Hopfield network context, and transplanted into a Neural Darwinist, psynet-style, process dynamics view of the mind.