I decided to publish another excerpt from the story "
Higher Mind ", which also concerns the information processing discussed at Habré. It sets out a somewhat unexpected look at the flow of information in the brain, than adopted in programming. But do not rush to criticize, it is just an attempt to take a different look at what seems to be known and taken for granted. Immediately, I note that this approach is based on cognitive psychology. I would be glad to have interesting opinions, since the question is open.
- Amy, I ran into a problem. How to create models? Information from objects is not enough. And she is noisy. You need a lot of presentations to correctly train the model.
- The model is not created according to the information from the subject. She is already in the brain. Only then can the information, as you say, be “accepted”, that is, choose an option for this model. Every object, situation, fact, phenomenon is only a variant of a more general model. And the model itself consists of other models. As a sentence describing a subject, from words that serve to describe not only this instance. You need to figure this out.
- Wait, that is, in your opinion, our ideas are not based on information from the subject?
- If this were so, then you would see only a small circle in the focal zone in color (in focus), everything else around you would be in shades of gray. That's the way the retina works, as you know. Information about the item is always incomplete, but you always see the item completely. Because the information from the subject only activates the version of the model that is already in the brain.
- How so, do not understand? How does the brain work then?
- Just like you solve equations. The equation already establishes relationships between variables. And you usually do not know all the variables at once. You substitute one to find others. And finally come to a decision. To one of the solutions, that is, the version of the model with these variables. In the same way it happens in the brain, only the models are more complex, distributed and multimodal.
“But isn't this information processing?”
- You need to change the way you think about this process in order to understand. The brain does not process information as a stream, but selects a variant of the model that is already in the brain suitable for triggers of the sensory system. Your neural networks are already working. They learn - that means they acquire a model of what they learn to recognize. And the picture is just a trigger at the input, which activates the chain of weights, closed to a particular output of the neural network.
- Where do we get new models from?
- You generate them in response to the lack of a suitable model with the predicted response of the medium. In response to an unrecognized (not suitable for any model) stimulus, a search is underway for options (indicative behavior), in animals this is the whole range of movements available. People have all the options for mental action. For example, mathematicians begin to compose mathematical theories, which then have to be confirmed in experience in order to explain a new phenomenon. And if the response of the medium (trigger) with the desired result turns out to be predicted, this option is remembered as a process model. This is if simplified.
- But we still use the idea of transferring information from the source.
- The transfer of information about a subject by radiation from it is a naive metaphor. It originated from the transfer of the subject, a letter with the text, which was interpreted as the transfer of information. But in order to read a letter, one must already have a model of letters, words, sentences (language). Otherwise, it is only a dash on paper. Letters are variants in a model, and it can be transmitted only if the addressee has the same model - he knows at least an initial. So, by the way, metaphors prevent you from finding the right solutions. And you find it difficult to abandon them, many of your scientists and you still use this metaphor.
“What should I do then?”
- Algorithm for generating models by triggers, and not vice versa. You had work on generative neural networks, this may be appropriate if you slightly adjust the approach. I'll show you how.
- Well, you say that you need to generate a model. But what to take as a basis? Anyway, we need some data from the environment. Or are the elements of the model completely arbitrary and in no way connected with reality?
- Is the scale of the thermometer somehow similar to body temperature? Does it somehow correspond to the Brownian motion of molecules? It only shows how it is in the units that you have taken for yourself. In degrees. All that corresponds to reality in these measurements is the simultaneous change in body temperature and the position of the pointer on the thermometer scale, since mercury expands from heat. This is all that you learn with the help of the device about reality, its change in scale units. And without a scale, the position of mercury in the thermometer says nothing to you! No rational being can know what reality really is, because all rational beings operate not with reality itself, but with models of it within themselves. Models - is the scale of the thermometer. At the same time, they were created for the convenience of operating them, and not to match reality. They are created by the means that are available to the brain.
- Do you also use reality models?
- Of course, but more complex than yours. Our models can take into account more parameters. Any subject can see, operate only with his own models in order to organize his actions. Therefore, what we see is the scale of the thermometer, and not the temperature itself. This was shown by your ancient philosopher Kant.
- It turns out that we are solipsists, closed in their imagination?
- You are not solipsists, because the model always has options. Like graduation on a thermometer scale. The choice of options occurs according to what your sensors are currently receiving from the environment. These are the triggers.
- All models are somehow connected?
- Not always. Take, for example, an apple. This is one object, but it can have many models. As a fruit, as a molecular object, as an atomic, subatomic. You have different models that are not directly related to each other, although they can describe the same substance at different scales.
- Why it happens?
- The brain cannot process all changes perceived from the environment. He cannot make a model that is so detailed as to treat a whole apple at the subatomic level, and he resorts to simplifications. The whole manifests new properties that cannot be deduced from the interaction of the parts, because this is a different model of the whole, what you call emergence. It would be possible to make a single model of the whole of the composite, and you could see an apple as an interaction of molecules. But it would be a huge model that would be difficult to operate. It is impossible for your brain.
- That is, it is practically necessary to create models of certain categories, classes, which may not intersect. But is there a connection between them?
- Yes, she is in the third model - how to make a chair out of sticks.
- Got it. Is that all I need to consider when creating models?
- Not. Any concept that you know, on closer inspection, will only be a fixation of change, a dynamic in the environment, which reflects the model. For example, even the color you see, only if there is another color nearby. If you stay in a room for a long time, where everything is of the same color (or wear colored glasses), after a short time you will no longer perceive it as a color. Red will disappear. Even the line you see, while the eye crosses it with its focus. You know about the experiments when the line disappears, if its image is synchronized with the movement of the lens. This is precisely what suggests that the models are designed to perceive changes. Therefore, the models reflect the options, and when switching between them, you see it. When the apple is ripe, the color in the brain also changes from green to red.
- But why such difficulties? Color can be determined and the usual light detector.
- Does the detector know what can be done with this color? Recognition of variants is necessary for the brain to know how it can change, how it can look different. By recognizing a horse, not a zebra, you will know that you can harness it and ride it, unlike a zebra. By recognizing the edge line, not the surface, you will learn that you cannot move further. The model contains not only recognizable, but also its variants, as well as possible actions. The detector won't help you with that.
- It is completely incomprehensible how to program modeling now. I'm confused.
- I warned that you are still very far from what could be called intelligence. You do not even know the basic tasks of modeling the environment and thinking. If you can understand what I'm talking about, you can make the intellect. Not earlier.