📜 ⬆️ ⬇️

Possible way to create a strong AI

Below I will describe the path (basic principles) that allows you to create an AI capable of passing the Turing test, that is, communicating with a person not “mechanically”, but “understanding” the essence of the conversation. This intellect will be in many ways similar to human, it will “experience” the same emotions as a person, he will have a memory, he will “think”. I will describe the processes and mechanisms inherent in the real brain, but point out the methods of implementation available in computer modeling, not claiming that they are “similar” to how nature solved similar problems.

Let us formulate several ideological theses underlying the model:

1. With regard to AI, the model should reproduce as much as possible all the properties of thinking available for observation, using not “software patches” for each property, but be such that all properties would flow from the very principle of organization of the model. In addition, since the real brain was created in the process of evolution, step by step, the “sequence of creation” should be visible. That is, it is possible within the framework of the model to show the sequence of devices with limited functionality, which, despite the simplification, remain operable in the sense of some "expedient" behavior.

2. Man and, accordingly, his brain is the result of natural selection. It is convenient to distinguish two ways of brain evolution. First, the complexity of the structure, the emergence of new functional systems, the increase in their "power". Second, the complication of the system of reflexes, the emergence of new unconditioned reflexes that cause actions and the appearance of reflexes that cause emotions and sensations. Some analogy can be drawn with the improvement of computers and their software. The importance of reflexes, emotions and sensations will be discussed below. In the proposed model, they are the basis of thinking.
3. The brain does not work directly with raw information. Previously carried out a strong processing of information. From the incoming information, essential features are distinguished, with which it is convenient to operate in the future. The functional systems responsible for this can be successfully modeled, for example, on McCulloch-Pitts neurons .
The main task at this stage is the classification of incoming information.
In nature, such networks are formed as a result of mutations. But it is not important in what way these networks will be trained. It is important to obtain at the output of these networks a convenient set of features that, on the one hand, will be compact enough, on the other hand, various phenomena that are important for recognizing the surrounding world will not “merge” into one. So our ears highlight a set of phonemes and signs of sound, with which it is convenient to operate, recognizing both the sounds of nature and the sound of speech.

4. Creating artificial neural networks, we make a certain network topology, and then we carry out its training. In this case, you can conditionally distinguish the stages:
  1. topology creation - before birth;
  2. learning is a teenage condition;
  3. exploitation is an adult condition.
But such an analogy can lead to strong delusions, if you try to extend it to brain training. The real picture looks like this:
5. The memory of the brain is not a structure that stores information and delivers it to some kind of "processor" that operates with it. Memory itself is a “processor” that is able to: perceive information, reflect the picture of the surrounding world, memorize, model, control behavior. When modeling memory, one must either reproduce the memory neurons, or create computer models with similar properties. And here it is important that memory neurons are fundamentally different from formal neurons and cannot be reproduced by them.

6. "Instinct" - in the everyday sense - the behavior that nature has endowed us. In reality, there are only unconditioned reflexes from birth. Reflexes cause reflex actions, emotions and sensations. Emotions and sensations govern the formation of memory. Memory drives behavior. “Instinctive behavior” is the result of learning, it is determined by the initial set of emotions and sensations and the environment.

7. Emotions do not control behavior directly. Emotions do not stimulate any actions. Emotions only give an assessment of the current situation. The situation can be real, or it can be fantasy. Memory captures everything that happens to us or what we represent, all our real actions and fictional acts. In this case, not only the “situation” is remembered, but also what kind of change in the emotional state corresponded to it. Subsequently, the memory recognizes “familiar situations” and tries to activate actions that promise a positive change in the emotional state, and blocks those that, in our experience, promise a negative change. All this happens on a “subconscious” level. Do not confuse this with memories, fantasies and awareness.

8. The associative memory in a very rough approximation is the following structure:
  1. Reserve;
  2. Memories. At the same time, the current picture of activity at synapses and the degree of change in the emotional state are recorded;
  3. Recognition. The neuron is not active until the picture on its synapses becomes “similar” to the one that corresponded to the moment of memorization.
  4. Activity In this state, the neuron is part of the picture describing the world. In the model, information about the activity is distributed both along the axon and along the synapses. An active neuron allows other neurons, part of the synaptic space of which it is to recognize patterns of a higher order of complexity. An active neuron "causes" or "inhibits" motor activity associated with "his experience", depending on the stored memory of the emotional background (positive emotional memories - causes motor activity that led to these emotions; negative emotional memories - inhibits motor activity, which led to negative emotions in the event of this activity).
- A set of active neurons describes a picture of what the brain perceives and thinks about at a particular moment. We will call the picture of the activity of neurons Current View.

9. Each memory neuron is, in fact, a memory of a certain event and experience gained as a result. Description of the picture of the world takes place "in terms of existing experience." The “meaning” that stands behind each “term” is not determined by the neuron itself, as such, but by the structure of its connections, which define the “meaning of the concept”. "All roads lead to Rome." There are no concepts of “hanging in the air”, all concepts going down the chain of associative links lead us to the sensory field formed by receptors or exits of functional systems, preprocessing information.

10. Speech by itself does not “contain meaning,” speech is a way of conveying information about the current view. The speaker has a memory, formed by life experience, and a certain set of emotions. If the interlocutor has a memory formed in similar conditions and a similar set of emotions, then the speech can cause him a picture of the current presentation, the “meaning” of which will correspond to the “meaning” in the speaker's picture.
In other words, it is impossible to create an AI simply by putting words into it, and generating a stream of communication. When a person associates the meanings of a word with a subject or phenomenon, he already has significant experience, which serves as the foundation. The child first learns a lot of information about food, its properties, taste, how it looks, what happens, how to get it, etc. and only then connects these concepts with the word food. The word food is one of the associative features of the “food concept”, which in turn is built on the sensory field.
The main disadvantage of the existing semantic networks is their isolation from the extra-semantic base.

11. Information about the surrounding world is first processed by functional systems that recognize individual phenomena and highlight their signs. Memory deals with "convenient" information. The same phenomenon may be part of different memories. The set of such memories pertaining to one phenomenon, in fact, describes all our experience and our knowledge regarding this phenomenon. Each picture of the world consists of a set of phenomena that cause recognition by memory elements. As a result, a certain picture of the activity of memory neurons appears reflecting our perception of the picture of the world.
It is convenient to select groups of neurons associated with a single phenomenon. Activation of neurons belonging to the same grouping can cause activation of neurons associated with another phenomenon, if in our experience these phenomena were present in shared memories. In other words, if an associative connection is established between them. The strength of the associative connection is determined by the number of joint memories and the strength of emotions that corresponded to these memories.

Associative memory and emotional apparatus allow us to describe how the process of thinking. Perhaps this is the main element in the present description and I will focus on it in more detail.

The picture of the current presentation determines what we think and what we perceive. Imagine that the change of "thoughts" occur tact.
We describe one measure:
The description deliberately omitted many important and fundamental points, the description of a workable structure is somewhat more complicated, the goal was to formulate the main ideas.

It is easy to trace how logical, heuristic and creative thinking takes place in such a construction.

12. For teaching AI to human language, it makes sense to “skip” through one of the stages of training and immediately create a ready-made functional system that will parse the speech, whether written or oral, extracting words from it and preparing preliminary information.

13. If we want to get an AI, which will more or less adequately perceive us, then we cannot do without setting the matrix of emotions. That is, descriptions of emotions inherent in man and the task of reflexes, describing the conditions of their appearance. Unfortunately, this task is practically not worked out by psychologists and is poorly developed in existing formal models of emotions.

14. When computer simulation of AI, one can begin with the task of simple-to-implement sensory stimuli, for example, an analogue of tactile sensations and modeling of internal sensations, such as hunger, boredom, etc. Subsequently, it will be possible to supplement the model with technical vision and auditory analyzers. It is important that reflexes are initially present, which will activate certain emotions. Through a fairly small information channel that has an emotional response, is not easy, but can be taught to complex concepts. Such training schemes are well developed and have long been used in teaching blind-deaf-dumb children.

The above principles do not pretend to be a complete guide to creating a strong AI. But their understanding is the key to starting the development of AI, able to pass the Turing test.

Source: https://habr.com/ru/post/105110/

All Articles