Making and Unmaking of representations are both necessary. Without unmaking representations and representation structures, it becomes increasing difficult for any signaling process or developmental process to occur. This note discusses some these issues.
Here are images of a fruit fly neuron from the above paper:
"Sleep and Synaptic Homeostasis: Structural Evidence in Drosophila" argues reduction of synaptic connections is done for energy conservation purposes. But this may not be the only explanation. If cells continued to expand their synaptic connections, why should their be an upper limit to their growth? The simple effects of size and mass may play some role in affecting reduction during sleep.
Below I discuss that reducing connections changes the structure of neuron groups and that alters the representations they instantiate and it is the representations which must be unmade to build a sensible representation simulation that is coherent. Many new and spontaneous connections would produce much useless information because the information forms only as a function of it's novelty. But utility is also important, and utility of information can be achieved through spontaneous representation formation, or connection making but must be achieved through trial and error. Trial and error, is making and forgetting.
In terms of representaitons it is natural to see there is making of representations and unmaking of representations as remembering and forgetting, learning and forgetting. the simulated world and the cellular stigmergic system are making and unmaking together. when something is unmade in the cellular system it is "forgotten" in the simulation, or it is "rediscovered" when connections are reconnected in the cellular system once again. thus to rediscovery requires two networks that suddenly start signalling each other again. this is like a time circuit
unmaking is arbitrary. making and unmaking are both arbitrary, because by definition, initial connections and representations are arbitrary. We do not learn things perfectly the first time, because what perfect is, is itself unknown.
To learn, making must win out of over unmaking structurally. a singular experience will produce new connections, new representations, new pathways, new networks. and then there must be a period of unmaking when those representations and connections will be undone. The making and unmaking are part of the same larger process.
unmaking connections must exist to produce learning. otherwise we oversaturate connections and that leads to useless network structures and many useless ideas. in a stigmergic sense, we must get rid of signal making, signal reception, and signal pathways we receive. otherwise we explode with too much material. It may take too much energy to maintain expanded synaptic connections and dendrite growth, but it also takes too much mass.
In terms of learning and synaptic growth, a single learning event will likely leave little trace of new connections. but a repeated learning experience will definitely leave a trace in the form of new and stable dendrites and synaptic connections between neurons. what does that look like when making and unmaking both occur?
what happens with unmaking is that connections are undone. unmaking of connections must follow a different kind of curve than making. they cannot be undone right away, otherwise there is no spontaneous adaptation. but connections cannot be kept around indefinitely, otherwise, the processing connections and signaling between cells will become very odd and confused. representations must be developed that make sense over the long term. that is how representations have power. we must learn to walk, and not just learn all the different leg twitches, but a particular set of leg twitches to walk.
sleep is the primary unmaking period. some period of unmaking must follow a period of making. and that unmaking must not eliminate all of the made connections. but it must eliminate a lot of those connections. Notice that learning is repeating process. connections are made and and lost (or forgotten). This process repeats over and over. Why? because it is the validity of repeated connections made arbitrarily (or randomly) that gives us a sense of what connections are valid over time.
The making and unmaking process will repeat, but the unmaking should never overwhelm the making process. this insures that repeated connection makings would endure, even if they are randomly eliminated by an unmaking process (just as they are randomly made by an impetus of making connections). I suspect that unmaking actually eliminates a large percentage of new connections, but what percentage should be something testable with a model that shows this behavior.
if making and unmaking occur at the same time, what remains? the making wins. it's a stronger signal. remember it's about making network connections. As we see in experience, we often can do something one time, and then have difficulty repeating it, as if that connection has just disappeared. This is unmaking in action. how do we create perfect games? by repetition.
What connections are preserved and what connections are not preserved? It is purely arbitrary. But unlike connection making which can be intentionally arbitrary, unmaking is the opposite, it is unintentional. Why should we see unmaking occur primarily during sleep? Because unmaking is explicitly about removing connections. unmaking explicitly will remove ways to move and act. If you forgot how to walk while first learning to walk, you would fall over immediately. new skills are de facto attentional. and losing new skills must necessarily be done without attention, eg. when we are asleep.
For instance, I am horrible at remembering names. I can be told a name, and talk to a person and 15 minutes later I have forgotten it. Such behavior would be catastrophic if I was just learning to run or to crawl and couldn't get away from a dangerous situation, or just learned not to avoid something hot, and then put my hand on it again because I had forgotten what I had just learned. But we must erase learning to determine what is valuable and a time delayed unmaking process appears to be a reasonable solution to that problem.
thus learning and forgetting something immediately is grossly dis-functional. learning something and forgetting it tomorrow or next week is less dis-functional and provides a mechanism to relearn it to establish what things learned are valuable over time.
there may be a second reason most unlearning happens during sleep, and why we dream. There must be a way to weight or reward or preserve important connections. the reward mechanism, the indication something was done correctly or we got it right. How does that work? Notice that it is innate to organisms, and it is slower than an immediate response. we feel when we get something right. it "feels right."
for instance in soccer, or golf, or any game, we get a sense of when me make perfect kick, or perfect swing. we can sense this after it happens, or before it happens, or as it's happening. but is this a signal process, in the same way neurons are signaling to do the action or is it some other signal mechanism. and is the recognition of perfection a posterior realization of the feeling we have or is it the feeling of doing something perfectly?
I think it's always a posterior realization. we feel a certain way, and then we connect that feeling to "getting it perfect". we recognize that feeling when we have it again, but that feeling of perfetion is a thing of it's own.
So in the making of connections we have two signaling processes going on. One is about the efficacy of the signals being propagated and the connections being made. a failure to propagate a signal likely produces an impetus to form new connections. extra stigmergic material that would be expelled by the cell but is not expelled probably has an effect on he production of new connections form the cell with extra neuro-transmitters.
perhaps extra neuro-transmitters at the axon tend to induce axonal growth. extra neuro-transmitters at the dendrite may tend to reduce dendritic growth, while a shortage of neuro-transmitters may play a role in some other process that induces dendritic growth. [there are many neuro-transmitters and they induce a variety of stigmergic responses (polymorphic) ] the point is, that the signal process itself induces some making of connections and some unmaking of connections.
the neurotransmitters also induce how the other cells behave. that is, some neurotransmitters inhibit and some excite receptivity. for the sake of argument, this feature is likely a cell wide regulatory behavior. like an automata or artificial neural network, the cell responds in the same way regardless of the signal, once it's trained or established it's behavior. of course we know that cells behavior can change, but can simple signaling cause that change to take place?
probably not. their must be a mechanism that affects connection making and unmaking and even the behavior of cells interaction.
signaling can induce this behavior, and signaling can induce actions in the cell to change it's behavior. this is basic stigmergic function. but what is the mechanism of making and unmaking connections? it is primarily feeling and not signaling. it is feeling that preserves a signaling pathway and one that undoes or expands a pathway. if we feel we are not walking right, we can change how we walk. if we feel we walk just fine, we do not change how we walk.
certainly this feeling process is stigmergic to the cell, but it is a secondary signal process that alters the cellular behavior. this process is likely both via neurotransmitters and generally in the fluid of the neural system.
Stimulation produces pathway development of neurons. Activity is the most stimulating. Pathway development cannot purely be about signaling. it's also about feelings. There is a loop of feeling, and of signaling. We make connections and drop connections because we feel they are important or unimportant. Importance probably derives because we do signaling through connections and this signaling produces confirming feelings. [Simultaneous Induction of Pathway-Specific Potentiation and Depression in Networks of Cortical Neurons 1999 Y. Jimbo,* T. Tateno,* and H. P. C. Robinson]
notice that unmaking should produce feelings too. unmaking connections probably produces bewildering feelings and confusion. it's likely a good thing that happens when we sleep.
The Representation Simulation System is an ecology of purposes which form loops of feeling. We keep what supports the ecology. How our body stays alive is though a good working simulation. It's basic function is an ecology of purposes that direction body function and spur action. We keep the connections that support this ecology of purpose. What is advantageous to the simulation and fulfills purposes is kept. it's not that we keep what is rational, or logical, but what is representationally valuable to that ecology, the the function of the simulation.
For instance, learning to read has no survival value... until written language was invented and it offered such an advantage. At that point, learning to read became a purpose that is adopted into a persons ecology. If a person is resistant to reading, the resistance is spurred from some other purpose that spurns reading in lieu of some other purpose intended to offer a survival advantage. Purposes may be obvious, as with reading, but many purposes will be non-obvious and subconscious. And in organisms at least, purposes originate in chemical actions in cells.
There are many low level features of an ecology of purpose and low level features of the representational simulation we do not have conscious access to. But these features are explicitly representational and not features of the cellular stigmergic system that instantiates the representational simulation. However, there must be low level features of that RSS that are induced by the CSS directly. And vice versa. This is the basic model of interaction between representation and some non-aware process that instantiates representation.
Moreover, there is no single model of neural cells that will describe all neural behavior. there are many different types of cells with many different types of behavior to give us a wide range of experience and expression. the simple neural net model is one set of algorithms to determine function. and it may work as a reasonable way to do image recognition but certainly it wouldn't be a good method to develop images from concepts. The learning of new concepts and the development of neural networks in an AI by itself requires some other mechanism.