What is important for creating AI 2: “Dependence. View from the neural network »

Addiction

Every time when television was broadcast about the dependence on the Internet or computer games I was not alone. These dependencies have always been equated to a drug one, emphasizing that, in practice, there is no difference between these dependencies. And what could be worse? I often played a lot of computer games, but I spent most of my time studying the PC itself and the software, spending a lot of time experimenting on hardware and software, and I’m fond of it even now. And, moreover, I did not have dependence on the Internet. He was very expensive, used it only to search for information on the essay, coursework, thesis. But he often caught the slanting views of his parents when they once again talked about these addictions on TV. It always surprised me exactly how the “scientists” (I would not be surprised that in fact, not scientists, but who the hell are) searched for and measured the common in these dependencies. In this article I will try to prove that each addiction is “practically no different from drug addiction,” and that this similarity does not at all prove the addiction of addiction. I will show that dependence, habit and hobby are “practically no different from drug addiction.” I will also “prove” that the AI ​​can be addicted and will have habits.
Since gambling, addiction and Internet addiction are completely different things, the only thing that is common is psychological dependence. Therefore, I will consider the dependence on the part of the “nervous” system, namely, neurons and neural networks (I take dependencies only for the sake of example, and what kind of dependence is absolutely irrelevant). I will not describe what a neural network is - the materials on this topic on the Internet are just heaps. The only thing that I will clarify is that the brain is a self-organizing neural network, therefore such a network will be discussed.

Attention, this is not the article where you can read and chew popcorn!
Caution! A lot of difficult to understand conclusions! Therefore, do not translate in vain tomatoes and rotten eggs ...
If you are ready, please under the cat!

First terms and abbreviations:

GNS is a global neural network (the entire neural network).
LNS is a local neural network (part of the GNS), the neurons of which can be both connected together and scattered across the global neural network. It is implied that the neurons of this network are united by some property, for example, this network is some kind of skill or memory.
NA - a neural network (it does not matter, the STS or LNS).
Negative signal - error signal in NA (weakens synapses).
Positive signal - confirmation signal (strengthens synapses).
Interference - third-party signals that complicate the work or lead to incorrect operation of the system on which they affect.

Skill ~ Habit?

To begin, let's start from the bottom - let's consider what a skill is:
Skill - an activity formed by repetition and bringing to automatism.
And also what is a habit:
Habit is an established mode of behavior, the implementation of which in a certain situation acquires for an individual the nature of a need that induces them to perform some actions and deeds.
(Both definitions are taken from Wikipedia) To perform an action you need to be able to do it. Logically thinking, we can conclude that a habit is a skill that works (“turns on”) “in a certain situation”. It turns out that a habit is a skill to pick up a situation in order to use another skill. From what it turns out that from the neural network, the habit of the skill, in fact, does not differ - it is the same skill, the same "local" network of neurons, organized in the process of learning. On the other hand, the entire neural network consists of the interconnection of such “local” networks - some were from the very beginning, others were built in the learning process, I think it was no secret to anyone. What else can a complex neural network consist of if it is not from smaller networks combined into one large one? Automatically closing an important window or message, by clicking on OK, out of habit, means that at this moment one of such LNSs “brought to automatism” received “control”.

')

LNS destruction mechanisms

Any LNS can be destroyed, for example, if someone hits you on the fingers when you automatically close the necessary window again (at this moment I remembered Demoshean), as a result of numerous negative signals, this local network will begin to weaken the synapses and, at the end ends, will start breaking ties. As a result, of course, you can “build” a network that will “force” you to squint every time you close a window, but it will be a completely different network (squinting in anticipation of hitting your fingers, or remembering the sensations of hitting ). It will not be so easy to destroy the network if it has sufficiently strong connections, which is why you will squint and close the windows at the same time.
Let's consider the "natural" destruction of LNS. Suppose you drowned your old mobile phone, to which you are used to and with your eyes closed, you know what is in it, where and how, but you had to buy a new one. A lot of habits and skills acquired from using the old phone are not suitable for it. From time to time your fingers do something, repeating what they did on the old phone, but on this phone these actions do not give the desired results, resulting in a lot of mistakes and stupid actions. Your LNS often receive negative signals and begin to break down, break ties, create new ones. New connections may be completely wrong, leading to illogical actions. LNS data from ordered networks turn into chaos, creating interference in the GNS. A skill has been gained - a new LNS has been built, but for some time, there is an old one, which, from time to time, forces us to do old actions that are inapplicable to the new phone. In other words, it creates "interference" in the STS, bringing us discomfort, wrong actions and a lot of trouble. The old LNS will still search for its place in the GNS for some time, destroying and creating new connections both inside and outside. It is either reorganized into new actions, which is unlikely (except that the old and new actions will be very similar), or it will crumble into chaos and the neurons will be “forced” to look for their place in the GNS in other LNS, inevitably interfering with their work.
In any case, the destruction of LNS, there is a high probability of neuron death. There is no guarantee that every neuron will find a new place in the STS, especially if it consists of self-contained LNS that no longer need to learn. Any neuron added to it will only interfere and will more likely interfere than help. Such a network or collapse (as a last resort, will be damaged), if it is weak or simply does not let in a new neuron. A free, non-irritating neuron will simply die. It is difficult to imagine what chaos is created when LNS is destroyed, it is still possible to take into account that each such LNS is connected with other LNS and its destruction leads to the fact that other LNS will be “forced to look for” new connections for themselves. LNS destruction can be equalized to a local explosion in the GNS. From this we can conclude that part of the neurons of the destroyed LNS will become attached to the free conclusions of other LNS. Free conclusions will create and destroy synapses for a long time until new correct synapses are found. You can also mention that in the process of learning in any case, some neurons can be released, but this does not compare with the destruction.

Some lyrics

A small lyrical digression before continuing to reflect on the topic. Imagining the destruction of the LNS, it dawned on me: "this is why every part of the brain has its place." This, of course, does not explain why it is exactly on which it is located, but it explains why. One has only to imagine what will happen if the destruction of some LNS affects the LNS, which is responsible for breathing, the heartbeat, or any equally important LNS. Death is what will happen (the released neurons from the destroyed LNS will search for their place in other networks, creating interference and, as a result, the networks in which they will be implemented will not work correctly). Another thing I thought about is that the destruction of the network occurs only with a sufficient number of negative signals. How does a person understand that he is doing something wrong? It happens differently. In one case, it is pain in some other emotions (him or someone from others), the person’s conclusions are the result, but, for example, the opposite of what was expected, because of which the person received stress. Stress itself, in this case, is a negative signal. Stress can be so strong that this person can lose the gift of speech - the LNS responsible for speech has collapsed (even if not completely, only a part of the synapses). And here speechless to his actions? Stress - a negative signal, spread through the GNS (albeit not uniformly), “caught by distribution” synapses were broken. This is a normal reaction to a negative signal in the NA. As already mentioned, the destruction of the neural network is unlikely to do without the death of neurons. Recall the expression: "Do not be nervous! Nerve cells do not regenerate. ” Here, I also note that in a natural NA mechanism the propagation of a negative signal (error) is not entirely transparent, as in an artificial one. Firstly, it is not so local and clearly acts; secondly, in addition to signals, there are still many catalysts / inhibitors in NS, etc. etc. I’m not going to climb this jungle - this is another topic and not one ... Well, enough of the lyrics, back to the topic ...

Some more terms

Having considered two simple concepts (skill and habit), you can step on the slippery path to begin to consider and the third concept - dependence.
Addiction (addiction) - an obsessive need that motivates a person to a certain activity. (from wikipedia)
Here I returned to the definition of habit (and highlighted the word need). We can say that dependence is the same habit and we will be right. But this concept is a bit more complicated than it would seem. Imagine the situation that you want to drink. The body generates its own kind of irritable signals, which as a result should lead to the satisfaction of the need to drink. You can drink a glass of water without even thinking, and not even remember that you did it.
And now we will more deeply go into the question and consider several terms (from Wikipedia):
The addiction (habituation) is a gradual decrease in the response as a result of continuing or repeated stimulation under normal conditions (not accompanied by reinforcement of the stimulus).
A receptor is a complex formation consisting of the terminals (nerve endings) of dendrites of sensitive neurons, glia, specialized formations of the intercellular substance and specialized cells of other tissues, which together ensure the transformation of the influence of external or internal factors (irritant) into a nerve impulse.

Addictive

Receptors are able to perceive not only the stimulus, but also its intensity. Let's make an addictive experiment. Put a brick coin on your hand and sit for a while, without moving it. Over time, we will stop feeling the coin. If we put a brick on the arm, then we would hardly be used to it, the mass of nerve endings would “scream” about injuries, i.e. changes would occur at the micro level, over time it would be visible to the naked eye: redness of the arm, etc. The reason for this is not at all receptor fatigue. Imagine a neural network in which the same signals arrived. What happened to the network? It changed until it learned under these signals. And suddenly the signals have changed. What will happen? Yes, the same as in the previous case - the network will be rebuilt until it learns. Presumably, our nervous system is able to "feel" only changes in stimuli. A hand pressed by a brick can no longer feel or move until blood circulation is resumed. If we move our hands (change the stimulus), we will again begin to sense the coin. We may not even feel the pain of a wound until we move it or touch it. And now let's take, for example, listening to music while listening to the same song for a long time, we will no longer receive the pleasure that we received when we first listened. Listening to music, we “teach” LNS, receiving a lot of positive signals. At some point, the network will learn and stop issuing positive signals. If you continue to listen to music that is already fed up, then learning will not stop, it will continue, only in other neural networks and if the network starts to receive completely “not acceptable” signals, it will signal very strongly with negative signals, it may even start to break down, in other words, you can go crazy. Let us imagine a network that is already sufficiently trained and here it is “told” that each of its actions is erroneous. That is why when we are sad we want to listen to sad music. Because music stimulates those neural networks in which the search for the right solutions occurs. Free outputs and neurons interfere with the NA, while listening to music, these neurons and outputs communicate and receive positive signals - confirmation of correct actions. Hence the interest in all sorts of things, the desire to learn, to get new sensations. There is never a balance in the system - there are always free conclusions that interfere with the National Assembly - “forcing” us to act, in search of a new one, to create and consolidate new synapses. On the other hand, if the signals contribute to breaks, first of all, at first, to negative signals, then such signals are a great discomfort for the National Assembly, such signals will be simply unacceptable. In other words, the more signals signaling that the synapse is superfluous and the stronger the synapses, the greater the discomfort for the NA , the synapse was 100% sure that it works correctly, and he is told that everything he does is wrong in the root is just a catastrophe .
All this is nothing more than addictive. Accustoming through learning but by its nature is not the only type. Consider another example. We are sitting in the car that drives along the pavement. If we sit in it for some time , we simply glue the teeth, we will no longer feel the same irritation as at the beginning of the trip (this is addictive through training). And when we stop the car, then for some time we will feel that we are shaking, when we are still going. What does it mean? This means that some processes are occurring in the NN that cause us to feel the shaking in its absence. It is possible that there is the same habituation as in the previous examples, but this type of sticking is impossible without the internal activity of the NA. Irritants are within the network itself. Maybe these stimuli are those neurons that have free conclusions that are looking for the right connections, but I suspect that these are not the only stimuli. The presence of stimuli inside is the main difference between natural and artificial neural networks. An artificial NS acts only in the presence of external stimuli. I dare say that the natural NA, the absence of external stimuli will first lead to insanity, and then death.

Thinking about addiction ...

Addictive ... I can not say exactly what it is but there are several versions of what is happening (I think they are all correct):

Version One "Changes"

We feel only those actions that lead to the creation or breaking of synapses. Not so ... Actions that weaken or strengthen the synapses, and we just feel the intensity of the changes in the synapses. For any impact, neither the network, i. arrival of signals into it, synapses will change in it (synapse weights, breaks, connections of new ones, etc.) When the network is trained on synapses, practically do not change, we are accustomed to the stimulus and feel it.

Version Two "Search for a nook"

LNS begins to search for the position of connections in which the impact on it will have the least intensity.

Version Three "Action-counteraction"

Each impact on LNS is suppressed by opposition in NA, i.e. NA sends signals to suppress its effects. As a result, we are still shaking after we left the car, which was driving along the pavement.
In my opinion, these are not separate versions, but parts of one whole.

Addiction

Consider a typical addict. After trying a drug and getting pleasure from it, he takes it again and again. Each time he gets less and less pleasure, which makes him increase the dose. Finally, he wants to throw a pernicious occupation, but he cannot throw it - breaking causes him to take the drug again.
It's no secret that the nervous system is a complex computer that controls all the processes in the body. Receiving signals about the emergence of needs, the nervous system sends signals to its satisfaction. For example, I want to drink - a whole chain of skills and habits (what? Where? How?) Is launched, because the National Assembly “knows” what to do (she is trained in this).With natural needs, everything is clear, but as with the "wired" everything happens.
Take for example gaming addiction. You sit down playing a game, get pleasure from it, in consequence of which the LNS responsible for this is built.
What happens next? We have LNS, in which there are quite strong synapses (we played a lot of games). If you do not use this LNS, it will begin to collapse.
Imagine a large LNS built no matter drug addiction, or gaming, or some other occupation ... This LNS begins to break synapses. Free conclusions of neurons again begin to look for their place in the network, connecting with other free conclusions. In all this chaos in the STS, a huge number of negative signals and interference arise. This, of course, also depends on the size of the LNS, which is destroyed: the more neurons in it, the more it creates errors and interference. How can we conclude that a small LNS can not be addiction. (By this, to the article, she is different from habit or skill), and that dependence is just a big LNS. You can also conclude that any frequent occupation can overgrow in addiction and this is not always bad. Diligent occupation and learning something will inevitably entail the creation of a large LNS,if it does not collapse with stress and bad emotions, as a result of learning this LNS against desire. It turns out, if you make an effort, even against unwillingness, the LNS can grow and “drown out” the unwillingness itself (or the LNS of “unwillingness” drowns it, then you can think and argue forever). I dare to assume that everyone has their own dependencies and cannot get away from this, the main thing is that they do not interfere with our lives.

Again, some lyrics ...

I dare to suggest that the countermeasure mechanism can arise as a simple training of some LNS perceiving signals of pleasure or pain as negative signals and training in the development of anti-active signals. And under the circumstances, the counter signal begins to be produced, even in the absence of the signals themselves, which leads to the strongest discomfort of the National Assembly. This is just a preview of my version. Here you can conclude that for one network the signal can be positive and at the same time be negative for another.
Here it is necessary to deal with the mechanisms of pleasure and discomfort. Getting pleasure is a positive signal for LNS, discomfort is negative. Most likely, with the mechanisms of error propagation, not everything is as simple as we think.
The mechanisms for the emergence and spread of positive and negative signals are not fully understood. It is not clear why the structure of the brain remains almost unchanged throughout life. In other words, what controls the neurons without letting them roam anywhere, but remain in their territory.

findings

- I was once again convinced of the importance of emotions (stress, joy, etc.) for the development of intelligence and it is not even clear what it is on the part of the National Assembly.
- Once again it was brought that the signals arrive in the National Assembly not only outside, but also from within it.
- The same signal can be positive for one LNS and negative - for another simultaneously.
“We still do not fully understand the principle of the work of a natural NA. Error propagation mechanisms are not fully understood.
- Modern INS never become an AI, because it is quite primitive (based on previous findings).
- If everything was clear - there would not be this article, because for a long time there would be working copies of AI.
- Addiction is just a big LNS.
- Everyone has their own dependencies and this is not always bad.
- It became clear why it is so hard to do nothing. Because as a result of inaction, synapses are broken and free neurons or their conclusions begin to look for where to connect, closing the different parts of the NA, which makes us do something or suppress actions by other LNS. It makes you wonder that, perhaps, this is the “generator” of signals inside the NA and thinking itself.
- Neuron, this is not just a cell, but a kind of single-celled organism with its own needs and behavior (which no one took into account when modeling ANN).

The source of the beginning of reflection was the video:


Thank you all for your attention. Waiting for interesting comments.