CN111582457A - Method for realizing general machine intelligence by simulating human memory - Google Patents

Abstract

The learning method provided by the invention simulates the associative activation process of human memory, recombines a plurality of response schemes by summarizing experience, generalizing experience, utilizing experience and input information, evaluates the schemes according to a mode of tending to avoid harm, and carries out response by means of dividing one response scheme into a plurality of intermediate links to search for simulative experience and the like. Through the method provided by the invention, the machine can gradually obtain simple to complex responses to the input information and has similar emotional expressions to human beings, which all show that the machine learning method provided by the invention has great difference with the existing machine learning method in the industry, and no similar method exists in the industry at present.

Description

Method for realizing general machine intelligence by simulating human memory

Technical Field

The present invention relates to the field of artificial intelligence, and more particularly to the field of establishing general artificial intelligence that resembles human intelligence.

Background

Current artificial intelligence is typically designed for a specific task, and there is no general artificial intelligence that can accomplish multiple uncertain tasks. Current artificial intelligence generally finds mappings from large amounts of labeled data, and they cannot infer causes, predict outcomes, and make selections and responses from the input information. Therefore, current machine intelligence and human intelligence vary widely. In the application of the invention, a multi-resolution object is established, multi-resolution connection relations among the objects are extracted in learning and life, and a multi-resolution relation network is established. After new information is input, the organization relation of the input information under multiple resolutions is called, similar organization relations in memory are searched, the generation reason of the information is presumed, the possible result is predicted, and reasonable selection and response are made. The invention further provides a method and steps for establishing artificial intelligence similar to human thinking, emotion and personality on the basis of the multi-resolution relation network.

Disclosure of Invention

In the present application, we first propose several basic assumptions about the way the human brain works, and then analyze the workflow of the brain through these assumptions. We then propose how to implement basic assumptions like the brain on the machine and how to mimic the workflow of the brain on the machine to implement general artificial intelligence.

First, we assume that the brain has feature extraction capability for input information at multiple resolutions, with the goal of relating the input information at different resolutions. Secondly, we consider the brain to have associative ability, which can predict the reason and possible result of the input information according to the past experience. Still further, we consider the brain to have the ability to generalize in order to apply past experience to different subjects. Still further, we consider that there is a need and emotional system in the brain, which is to create various possible responses under the stimulation of input information and to select the response that meets the expectations of the brain. Finally, we believe that the brain has the ability to mimic. The brain performs the simulation by simulating past multiple experiences and combining these experiences with realistic information through generalization capability, and continuously adjusts according to the actual situation during the output process.

Fig. 1 is a diagram of the main parts proposed in the present patent application for implementing general artificial intelligence. S1 is implementing multi-resolution data bit

The whole machine's thinking process is iterative. The way the machine processes new information each time is: the current target is converted into the 'inheritance target'. And performing multi-resolution feature extraction on the new information by the machine. In the relational network, through the associative ability, the experience related to the 'inheritance target' and 'new information' is searched. These experiences are generalized to the input information through generalization capability. By means of the past organization mode of the experiences, the generalized experience segments are combined into a possible response plan according to the time and space relation. Then based on past experience, the influence of the response plan on the user can be evaluated under the principle of 'driving toward the profit and avoiding the trouble'. If the evaluation can not be passed, the response plan is reestablished; if the answer plan passes, the answer plan is used as an output plan, and each link in the plan is expanded to a more specific intermediate link in a segmented simulation mode. This process is also iterative until it is deployed to the underlying experience that the machine can immediately execute. The process is a process of thinking while doing, in the process, once new information is input, the machine returns to the process of converting the current target into the inheritance target and performing multi-resolution feature extraction on the new information. Therefore, in the invention application, the machine only needs to use simple steps to repeatedly iterate, and the thinking process similar to human and the requirement and emotional response of human can be realized.

Drawings

Fig. 1 is a main component part of the present invention application.

Fig. 2 is a method of implementing multi-resolution feature extraction.

Fig. 3 is a functional block organization diagram.

Fig. 4 is a schematic diagram of another functional module organization.

Detailed Description

In the present invention we first explain why we have implemented general artificial intelligence in the above-described way and then in particular what way we have implemented these goals.

We first analyze a possible workflow for the brain. We hypothesized that the brain already has the ability to use multiple resolutions to extract features of the input information (assuming this is the ability to come within the gene). When external information is input, the preprocessing part of the information firstly extracts information features under different resolutions of 1-K1 (natural numbers), and each layer of resolution may have a plurality of corresponding features. The brain then begins searching for similar features in memory. The basic search method of brain is assumed to be to activate the feature to be searched, so that it emits a specific pattern of activation electrical signals, which can propagate in the memory space and will decay with the propagation distance. It is assumed that other neural tissue in the memory can receive the electrical signal of the specific pattern, but only the memory of similar features (neural tissue previously memorized by similar information) can better receive the activation electrical signal (because of pattern matching). It is obvious that these electrical signals can only activate the adjacent nerve tissues (because the input excitation is strong) and can also activate the similar farther nerve tissues (because the receiving ability of the other side is good), and the characteristics with deep memory (such as more neurons or synapses) can also be activated because of more receiving units. Given the nerve tissues that are activated, they also emit their own specific pattern of electrical signals if their activation value exceeds a preset threshold. Then, the electric signals sent by them can only activate the adjacent nerve tissues and the nerve tissues which are far away from them and have similar structures, and can also activate the memories which are not similar to them but have high memory values.

For example, the following steps are carried out: when a set of "table" features is input into our brain, the brain first converts them into information features at multiple resolutions (e.g., a solid figure at the coarsest resolution, then the shape of the table and the contour of the table leg at finer resolution, then the texture of the table, other details of the edge contour at finer resolution, etc.), and then the features at each resolution send out the activation electrical signals corresponding to these features in turn (assuming that there is a neural tissue, they can send out their activation electrical signals for the corresponding pattern under the excitation of different input information features). Obviously, these activating electrical signals can only activate our nearest memories (because of the close distance in the memory space, the attenuation is small, we call the activation of the proximity) and memories similar to the input information (because of the pattern matching, they have good receptivity, we call the activation of the similarity), and also those memories impressive to our (because of the many neurons or synapses of these memories, the receptivity is better). So by "table" we may recall a package of snacks we placed at the table yesterday, or a scenario where mom done manually at the table after recalling hours. To achieve the above-described associative capabilities, it implies that our memories are assumed to be stored in chronological order, and implies that the information stored in the memories per time period is assumed to be stored in a manner similar to reality. Thus, after the "table" signal activates the "I and mom do a manual at table" scenario after our hour, the memory of "when a ball suddenly breaks the window glass, although the message is far from the table, it has a high memory value. Since our brain stores our emotions while storing memory, those strong emotions are also deeply remembered, and also have more neurons or synapses to store them, we can also recall that our emotion changes from "happy" to "startled" (strong memory activation). We refer to the 3 activation modes described above (proximity activation, similar activation, and strong memory activation) as associative activation. Although we propose a hypothesis, we consider that this hypothesis is very similar to the associative function of our brain. The associative function is the main working mode of our brain.

Another working mode of the brain is "driving toward the profit and avoiding the harm". In the above example, the pleasure of "I and mom do manual work at the table" is because such activities satisfy our "needs", such as a safety need or an familiarity need, which brings pleasure to us. However, the danger is frightened because the emergency destroys the safety of people, so that the danger symbol is activated, and people are aware of the emergency and need to change the state of the people to deal with the emergency. When we know what happens by analogy with what happens through experience, the activation value of the danger symbol is reduced (the activation value of the danger symbol is increased and decreased, and is also determined by the connection strength between different input features and the danger symbol, and the connection strength is partly preset by the innate and partly realized by the associative activation process of the memory of our acquired). But at the moment, the 'window glass breakage' breaks down the requirement of property safety (at the moment, the safety requirement is generalized), so that the property safety is broken down, and the 'angry' emotion is brought to the people. This emotional transition from "frightened" to "angry" is also present in our memory. From this process, these emotional transitions are controlled by a shift in our mental state of need, an "inherent" ability (thinking that "baby" is "angry" because of no milk being consumed). On the other hand, the people can slowly learn how to adjust the emotion through a method of 'driving towards interest and avoiding harm' on the coming days, and the emotion is used as a means for expressing information. We refer to the above emotional response after assessment by demand as the "demand and emotional system". Although we propose a hypothesis, we consider that this hypothesis is very similar to our brain's mood-regulating function. These functions are the main modes of operation of our brain.

The brain has prediction ability on the basis of association and 'tendency and avoidance'. The prediction is the ability of the brain to be used at any moment, and the basis of the prediction is a relationship network between things and concepts established by the brain through a memory and forgetting mechanism. The relationship network mainly accomplishes 2 things: 1, increasing memory (using more neurons or synapses) for information requiring increased memory. 2, establishing the connection strength of each information to the needs and emotions. Once the relationship network establishes these two relationships, we can find similar memory by associatively activating the system when new information is entered. And by activating similar memories in sequence in both the time dimension and the similarity dimension, the causes and results of the similar memories can be found. The brain can generalize past experience (cause or outcome) to the input information using an analogy method to generate a prediction of the input information, so that the brain can make his/her choice based on the need and emotional system.

In the process of prediction, generalization is the most critical step. In the present application, we propose an empirical generalization method: in a plurality of similar processes, common parts in the similar processes are searched by reducing the resolution, and the common parts are taken as process characteristics. In reconstructing a particular process, real world objects may be substituted as long as they are the same as experienced objects at the resolution specified in the process features. This is a key step in the generalization of experience. This is also why we need to establish features at multiple resolutions of different things, scenes and processes.

With the associative activation system, generalization ability, demand and evaluation system, the brain has the ability to predict and select. On the basis of these capabilities, the brain performs its own selection by piecewise simulation. The segmented simulation is to plan the response of the brain to a concrete implementation: the main concepts (intermediate links) in the planned responses are expanded layer by layer through segmented simulation (more detailed intermediate links) to memorize the bottom-layer experience which can be specifically executed, and the responses are executed according to time and space division.

In the following, we implement general artificial intelligence by mimicking our proposed brain working mode.

1, establishing the capability of multi-resolution feature extraction.

We believe that there are not likely to be two things that are exactly the same in our world. When we say that two objects are homogeneous objects, it means that they are the same at the information resolution we use. Therefore, in the present application, we need to gradually use different resolutions to identify information from detail to abstraction. This is the process of creating features at different resolutions.

Before building multi-resolution features, one problem we first have to solve is which data combinations can be used as features. Things are complicated and complicated, and the corresponding characteristics of each kind of things need to be established, which is a task which cannot be completed. Therefore, in the present invention, we propose a method of "using local similarity as a feature". The reason for this approach is: we believe that, in the history of evolution, living beings have evolved in the direction that most conserves energy consumption when recognizing information. Since saving energy consumption means a higher chance of survival for the living being. Therefore, we introduce this idea into machine learning as well.

Combining the two aspects, we propose that the selection criteria of the information characteristics are: 1, these features are widely present in our world. Only then can we reuse these features in the information processing process, which saves most energy. 2, the same data has different data characteristics under different resolutions. So that we can compare the similarity of two at different resolutions.

1.1 selection of multi-resolution features.

We propose a method of establishing information features at multiple resolutions as shown in fig. 2. S201 is to divide input data into a plurality of channels by filters. For images, these channels include filters specific to the contours, textures, tones, dynamic modes, etc. of the graphics. For speech, these channels include filtering for audio components, pitch changes (a dynamic pattern), and other speech recognition aspects. These preprocessing methods can be the same as the image and voice preprocessing methods existing in the industry at present (for example, a better method for extracting multi-resolution features is adopted by adopting wavelet transform), and are not described herein again. The machine can also directly input the preprocessed data as multi-resolution data, and the local similarity among the data is searched for as a characteristic. It is also possible to subdivide the data into data intervals using different windows and then find local similarities between all the intervals of all the data as features. It is noted here that the windows of different resolutions may be temporal windows or spatial windows. S202, for data in each channel, windows with different sizes are used for searching local similarity. Data were selected using windows of different sizes, which mimic human attention intervals. Typically, data within a large window corresponds to using a low resolution, while data within a small window corresponds to using a high resolution. In S202, the specific steps may be as follows: the machine may use local windows W1, W2, W3, wherein W1 <. is a natural number) in series to compare all window data under all inputs to find a feature that can recurrently appear local similarity.

In comparing data similarity within a window, a similarity comparison algorithm may be used. Since the similarity comparison algorithm is a very mature algorithm, and can be realized by professionals in the industry based on the known knowledge, the description is omitted here. The machine places the found locally similar features in a temporary memory base. Each new local feature is put in, and an initial memory value is given to the new local feature. And increasing the memory value of the bottom layer characteristic in the temporary memory library according to the memory curve every time an existing local characteristic is found. The information in the temporary memory library complies with the memory and forgetting mechanism of the temporary memory library. And the bottom-layer characteristics which survive in the temporary memory library can be put into the characteristic map library to be used as the characteristics of long-term memory after reaching the threshold value of entering the long-term memory library. There may be multiple long-term memory banks that also follow their own memory and forgetting mechanism.

1.2 algorithm of extraction of multi-resolution features.

The machine needs not only to build a database of underlying features, but also to build a model that can extract these underlying features. In S203, the machine builds the underlying feature extraction algorithm. One possible algorithm is a similarity comparison algorithm a that finds local similarities. When new information is input, the machine uses a preprocessing (such as removing or compressing coefficients of part of the base after various coordinate basis transformations) method on the information, and then the machine uses a large window (low resolution) and a small window (high resolution) to extract data features in the window. The use of windows is equivalent to mimicking human attention so that we can obtain simultaneous extraction of the multi-resolution features and the location of the data features in the input. These positions are used in the mirror space where we use the features to reconstruct the input information. The storage mode of the mirror space is the basis for realizing the adjacent activation principle.

Another algorithm for extracting the underlying features is the neural network algorithm B. It is an algorithmic model based on a multi-layer neural network. After the model is trained, the calculation efficiency is higher than that of a similarity algorithm. The machine trains the multi-layer neural network using the selected information features as possible outputs. We can use a layer-by-layer training method. In S204, the machine trains the algorithmic model using the local windows W1, W2.., Wn, where W1.. is a natural number) one after another. In the optimization, one is to add a zero to L (L is a natural number) layer neural network layer on the corresponding previous network model after each window size increase. In S205, when optimizing this neural network with added layers, there are two options: 1, optimizing only the added zero to L (L is a natural number) layers of neural network layers each time; thus, the machine can superpose all network models to form an integral network with intermediate output. This is most computationally efficient. 2, the current network is copied to a new network each time, and then the new network with zero added to the L layer is optimized. Thus, the machine finally obtains n neural networks. One for each neural network model. When extracting features in information, a machine needs to select one or more neural networks according to the purpose of extracting information at this time. Therefore, the machine may obtain two kinds of neural networks for extracting information features. One is a single algorithm network with multiple output layers, which has the advantage of low computational resource requirements, but less feature extraction capability than the latter. Another is a plurality of single output neural networks. The method needs a large amount of calculation, but the feature extraction is better. It should be noted that the above method can be used for processing images and voice, and can also be used for processing information of any other sensor by adopting a similar method. It should also be noted that selecting different resolutions means selecting different windows and selecting different feature extraction algorithms. The size of the extracted features is also not the same. Some underlying features may be as large as the entire image.

Here, a method of training a multi-layer neural network is proposed: a multi-resolution training method. The multiresolution training method refers to decomposing the input information into different resolution layers. The partial resolution layer is then used to train the neural network. For example, the neural network is trained preferentially using information data with low resolution, and then the neural network is trained while gradually increasing the resolution. When the required accuracy is reached, the information of the other resolution layers can be discarded. Of course, the order in which the resolution layers are used may also be adjusted according to the purpose of recognition. The machine can also divide the input information characteristics under different resolutions into groups according to the resolutions, train the multi-layer neural network individually, and then weigh and average the outputs of the plurality of neural networks to be used as the total output.

The computer implementation of multi-resolution extraction and the computer implementation of similarity comparison are well-established algorithms for image processing at present, and are not in the claims of the present application, so they are not described herein again.

1.3 static feature map extraction.

Note that the static feature map is built based on resolution, which represents the machine's self-built classification of things by similarity. For example, two tables, which may belong to the same category at a coarse resolution, and which may be different categories at a fine resolution. The machine need only extract features of the input information at different resolutions and represent the input information as a whole. When the similarity comparison is carried out between the input information and the memorized information, the machine respectively carries out comparison under different resolutions. For example, the same two things are both formed by combining a plurality of resolution feature maps. When the similarity of the two is compared, the similarity of the two can be quantified only by comparing the two on different resolutions.

1.4 dynamic feature map extraction.

In a moving image, there are two kinds of similarities. One is the similarity of the images it contains to those in other processes. The machine only needs to carry out the feature map in the process and the feature maps in other processes according to the static feature map extraction method. They are also static feature maps in nature. However, in a dynamic process, there is another type of similarity, that is, similarity of motion patterns. The motion mode means that the machine ignores the construction details of the moving object itself, and focuses on comparing the motion modes of the moving object. Also, there are comparative resolution problems, such as a person walking or sliding or running, and we can not even notice the difference of these motion patterns at a rough level, so we think their motion patterns are the same at this time. However, as we increase the resolution, we find that the sliding person is moving smoothly, while the walking person and the running person have various motion characteristics including the relative motion of various parts of the human body and the overall motion of the human body as a whole, and also including varying speeds, so we find that their motion patterns are different.

To solve the problem, the invention provides a dynamic local similarity comparison method. In particular, windows of different sizes are used to track different portions of an object. Such as a person running, walking or sliding, we can use different windows to represent different resolutions. For example, when we use a large window and take the whole person as a whole, we track the motion pattern of this window, and we find that the motion pattern is the same in these three cases. However, when we use a smaller window to extract the motion patterns of the two hands, the two legs, the head, the waist, the buttocks and the like of the person, we distinguish the difference of the three motion patterns. Further, if we use more windows for the hand to focus on the motion pattern of the hand, we can get a finer resolution motion pattern. In addition to the spatial resolution, the machine needs to establish a different temporal resolution. For example, people on the street have endless streams, which is a movement pattern of people. But from the finer time resolution we can find the morning and evening work. The peak of the crowd flow at time. The change rate can be obtained by comparing the change of the motion trail under different time resolutions. And the rate of change is an important dynamic feature of motion over time. Therefore, the extraction of the motion pattern is based on a certain time resolution and a certain spatial resolution, and the machine searches for common dynamic features by processing a large amount of dynamic data.

When the machine finds a similar motion pattern, the machine puts data representing the motion pattern into a temporary memory base as a candidate of the dynamic characteristic diagram, and assigns a memory value to the candidate of the dynamic characteristic diagram. The machine uses windows of different sizes and iterates the above process on the data, so that the machine can obtain a large number of dynamic feature map candidates in the temporary memory base. Like the static feature map, the machine also uses a memory and forgetting mechanism to make the extracted dynamic feature map superior or inferior. The motion modes widely existing in various moving objects can be discovered once and again, so that the memory value is increased once and again, and finally the motion modes enter a long-term memory library to become long-term memory.

Similarly, we can do the same processing for other sensor information than the image. For example, for speech, we can use time windows with different sizes as resolution, take some specific language attribute (a certain feature) as object, then compare the variation pattern (motion pattern) of this observed object, and find the similarity of local variation pattern (such as rising tone, falling tone, vibrato, plosive, etc.). Similarly, for data such as touch, sensation and the like, a similar method can be adopted, and a dynamic feature map of the objects can be established only by taking a certain feature as an observation object according to different resolution scales on different dimensions of the data to find similarity between change modes of the observation object. It is noted that the dynamic profile is built based on dual spatial and temporal resolution, which represents the machine's self-built classification of dynamic processes based on similarity of dynamics. They do not relate to the static characteristics of the observed object.

In life, dynamic features are used in our lives with very high repeatability, since they are not related to the objects that implement them. For example, different objects may have similar dynamic characteristics. Such as running features of a puppy, which are repeated in memory one time, static and dynamic features of coarse resolution are repeated one time, thereby gradually increasing the memory value. These features include: coarse resolution features of puppies and coarse resolution features of puppy movements. In those memories, the specific characteristics of each dog, particularly to each breed of dog, are low and may even be forgotten gradually. The motion posture characteristics of different breeds of puppies may also differ, but the memory values of these differences are lower than the common motion characteristic memory values of puppies. The above differences come from: the motion of each puppy includes a rough puppy image and rough puppy motion characteristics. These rough puppy images and rough puppy motion characteristics are similar in all processes so they can be repeated over and over again, resulting in higher memory values. So we recognize the motion of the puppy with the highest memory value activated first, that is: two features, object, movement, then animal and running. If we never see a cat, we call a coarser relationship when we call memory recognition the first time we see a cat running: "animals" and "runs". This is because we do not recognize the cat, but we can judge it to be an "animal" from other features of the cat. In predicting a cat running, we would borrow memory about various "animals" and "runs", by finding the most similar memory, by generalizing to speculate. Therefore, at the core of generalization, under the more rough last concept, things in memory and things in reality are similar (same in attribute) in the experience we borrow, so that the principle of replacing the same in-concept attribute can be adopted to generalize the experience. The basis of generalization is to extract relationships between things at different resolutions using memory and forgetting mechanisms. Those things that do not look related are often the same kind of things at other resolutions. At this resolution, they are identical to the connection of the motion features, so they can be used for generalizing the experience. Moreover, because dynamic profiles are connected to a wide range of objects, their objects are often a broad class of things that can be easily exploited by machines using generalization mechanisms (intra-conceptual replacement). The dynamic features and multi-resolution relationships are 2 key tools of our generalized experience.

And 1.5, establishing a feature map library.

In the temporary memory base, we use a memory and forget mechanism to maintain these profiles. The method specifically comprises the following steps: every time a similar feature map candidate is found, the memory value of the feature map candidate is increased according to the memory curve. Meanwhile, all the memory values in the temporary memory library gradually decrease along with time according to the forgetting curve. If the learned value is decremented to zero, then this feature map candidate is removed from the temporary memory store. If the memory value of a certain characteristic diagram is increased to the preset standard, the characteristic diagram is moved into the long-term memory library to become long-term memory. Here, the memory value represents a time when the corresponding feature map can exist in the database. The larger the memory value, the longer the time it takes to survive. When the memory value is zero, the corresponding characteristic diagram is deleted from the memory base. The memory value is increased or decreased according to a memory curve and a forgetting curve. And different databases may have different memory and forgetting curves. The machine continuously uses the above process in daily life during the training process, finally obtains a large amount of characteristic maps, and the characteristic maps can be put into a quick search memory base. Similarly, we can do the same processing for other sensor information than the image. For example, for speech, we can distinguish the frequency components and relative intensities of different voices as static features, and find local similarity from the static features. Similar methods can be adopted for data such as touch, feeling and the like, and similarity comparison results at different resolutions can be established only by searching for similarities according to different resolution scales on different dimensions of the data, so that the characteristic diagram of the data is established.

And 2, realizing the associative activation capability.

The key to realizing association is to establish a memory network and adopt 3 activation principles: the proximal activation principle "," the similar activation principle ", and" the strong memory activation principle ".

2.1 establishing a memory network. When establishing memory, similarity, time and spatial relation among things need to be preserved, and the similarity, the time and the spatial relation are the basis of 'proximity activation', 'similarity activation' and 'strong memory activation', so that a method called mirror image space is adopted to store data. When the machine extracts multi-resolution information features from the input, the machine needs to use these features to create a mirror space. The machine firstly adjusts the position, angle and size of the bottom layer features according to the position, angle and size with the highest similarity with the original data by scaling and rotating the extracted features, and places the extracted features and the original data in an overlapping mode, so that the relative positions of the bottom layer features in time and space can be reserved, and a mirror image space is established.

The machine memorizes the information according to the input time sequence. So if the machine interval of interest switches back and forth between two spatial positions, then the adjacent space in memory is not the actual spatial adjacent position, but rather two constantly switching spatial positions in memory, as they are placed in adjacent temporal order. The "strong memory activation" of the machine is accomplished through a memory and forgetting mechanism. Each information input into the machine is converted into a multi-resolution information feature map. These profiles are stored in memory over time. At the same time, the multi-resolution characteristic maps can send out own activation signals in the memory space in sequence. It activates not only neighboring memories but also those similar to it. It is emphasized that each extracted feature will emit its own activation signal. That is, the same thing, scene, or process, at different resolutions, may signal activation of multiple corresponding resolution features. The feature that receives the activation signal and is activated, because it is activated once, increases its memory value according to the memory curve. Meanwhile, the memory values of all the characteristics in the memory interval are decreased progressively according to the forgetting curves of the respective memory banks. Thus, for those features that are repeatedly activated, the memory value will increase with the number of activation times, so that the received activation signal is stronger, thereby forming a positive feedback and increasing the memory of the user. So that the memory with high memory value is activated repeatedly to gradually increase the memory value or enough memory value is given at one time to make the information memorized. Each time new memory information is generated, the machine needs to determine the memory value given to the memory information when storing the memory. The principle that these memories are assigned memory values is: their memory values are positively, but not necessarily linearly, related to the activation values at which storage occurs. Their activation value, is the intensity of activation of them at the time of the occurrence of the storage.

The machine stores three types of data in the memory, and each type has own memory value. The first type is the information characteristics of external input, including the characteristics of all external sensor input information, which includes visual, auditory, olfactory, touch, taste, temperature, humidity, air pressure and other information, which are closely related to the specific environment, and they are stored according to the organization method of the original data, so that the stereo mirror space can be reconstructed; they maintain their memory values in accordance with a memory and forget mechanism. The second type is internal self information, including electric quantity, gravity direction, limb posture, operation condition of each functional module and the like, which is irrelevant to environment, and the memory values of the information are set according to a preset program. The third type is data of machine requirements and the state of the requirements, including data such as safety values, danger values, income values, loss values, target achievement values, domination values, self body state evaluation values and the like; status data from these needs and requirements is also included. Meanwhile, the machine also generates various emotions according to the situation that the self demand is satisfied. The relationship between these emotions and the situation in which the own needs are satisfied is set by a preset program. Meanwhile, the machine can also reversely utilize the relation among the internal condition, the external condition and the state of which the self requirement is met to adjust the preset program parameters generated by the emotion, so that the self emotion is utilized to influence the outside. To achieve this, we use the following method: different symbolic representations are established for the self demand type and the emotion type of the machine. When an event occurs in the mirror space of the machine, the machine needs to store the current mirror space in the memory. The machine stores all the profiles (including the profiles, the demand symbols and the emotional symbols) in memory together with their initial memory values (positively correlated, but not necessarily linearly correlated, with the activation values at the time of storage). We refer to the memory value obtained by the demand symbol, along with the demand symbol, as the demand state.

The demands of the machine can be various, and each kind of demand can be represented by a symbol. Such as safety and risk, revenue and loss, dominance and dominated, respect and neglect, etc. The difference and magnitude of the demand type do not affect the claims of the present invention. Since all requirements are treated the same way in the present application.

The machine's emotions can be varied and each type of emotion can be represented using a symbol. Such as excitement, anger, impairment, tension, anxiety, embarrassment, boredom, coolness, confusion, aversion, jealousy, fear, joy, romance, sadness, affection and satisfaction, etc. The difference and magnitude of the emotion types do not affect the claims of the present invention. Since all emotions are treated in the same way in the present application.

The initial activation value assigned by the machine to the input information is also propagated to the demand and emotional data of the machine through the relationship network, and the instinctive response of the machine to the information is generated. Machine requirements and mood data are a very important type of "personification" data. It is closely related to external input information and self-information in itself. Their relationship is: when external data or internal data is input, the machine will generate responses, which in turn will get external feedback and change the internal state (e.g. the amount of electricity decreases). In the present application, we assign to the machine a demand type similar to that of human beings and a demand obtainment value indicating the situation where the demand is satisfied. Meanwhile, for better communication with human beings, the meeting condition of the machine requirements and the emotion of the machine are connected through a preset program. The specific implementation method can be as follows: the human being, during training of the machine, tells the machine, through preset symbols (such as language, action or eye gaze), that those environments are safe, that those environments are dangerous, or may further tell the machine different levels. As with training a child, it is sufficient to tell it "very dangerous", "comparatively dangerous", and "somewhat dangerous", etc. Thus, the machine can gradually increase the connection strength (due to the increased number of repetitions) of the dangerous environment or process common features and the dangerous built-in requirement symbols through training, memory and forgetting. Then, when the machine processes the input information the next time, after giving the same initial activation value to the input information, the activation value of some features, because of the close connection relationship with the sign of danger, delivers a large activation value to the sign of danger. The machine is immediately aware of the danger and will immediately process this danger information based on its own experience (which may be a pre-set experience or a summarized experience). Of course, since a great deal of experience is already being passed on by humans, in training, we can also tell the machine directly how dangerous those particular things or processes are, which is a way to preset experience for the machine. The preset experience can enable the machine to establish the memory frame through language to connect the risk factors and the risks, and can also be realized by directly modifying the existing relational network of the machine (modifying the memory values of the risk symbols in the corresponding memory frame). The two values of safety and danger are to tell the machine how to identify safety and danger factors and thus learn if to protect itself. The profit and loss values tell the machine which behaviors we encourage and which behaviors are penalized, which is a reward and penalty system. As with training a child, we need only give a reward or penalty after it has made a particular action. Or when rewards and penalties occur, it is sufficient to tell it the reason. It is also possible to preset experience (for example, telling it in advance that there will be rewards for behaviors, that there will be penalties, or to modify its cerebral nerve connections directly for the purpose, human beings will bring pleasure (receive rewards) each time they reach a goal, which is a gift that we are brought by evolution, which is a motivation that our race can continue to develop, we can also give similar instincts to machines, which will build a motivation for them to develop themselves, so that when a goal is reached by a machine, the reward given by human beings can be given to the machine, or a value can be given by a preset program, thus motivating the machine to try constantly, the allocation and the being dominated tell it the range that it can dominate through gains and losses, which varies with different environments and different processes, which is a reward and penalty system, but it differs from the benefit system in that, the loss of interest system focuses on the outcome of the behavior and the dominance and dominance focus on the scope of the behavior. It uses the same training method as the benefit loss system. It is also possible to link the evaluation value and the demand of the own physical state of the machine with the emotion, the external input information, in order for the machine to understand the evaluation value of the own physical state and the link therebetween. For example, in rainy weather, the machine stores the memory if it finds its own power, or if other performance is degrading rapidly. If repeated as many times as possible, the machine will have a tighter connection between performance degradation and rain. These associations, when the subsequent machine selects its own response procedure, activate the rain feature, which is passed on to the loss value, which is signed larger, through the association activation procedure. And the loss value is one of the indicators that the machine uses to evaluate what response to select, the machine may be inclined to select a solution that excludes the loss value from rain. Therefore, in the present invention, we only need to put the reward and penalty into memory along with all external and internal information, and the machine can incorporate these into its own thought, without having to do many "rules" to tell the machine how to recognize the environment, what to do, and how to express the emotion (which is a task that is practically impossible to accomplish).

The emotion of a machine is an important way for a machine to communicate with a human. So in the present application we also take into account the mood of the machine. The emotional response of human beings is an inherent response to whether the needs of the human beings are met, but through the study of the future days, the human beings gradually learn to adjust the response, control the response and even hide the response. Similarly, the emotion of the machine and the requirement of the machine are connected through a preset program. For example, when a danger is recognized, the emotions of the machine are "worry", "fear", and "fear", which is how much the danger is. For example, each internal operation parameter of the machine is in a correct interval, and the machine is given emotions of 'comfort', 'relaxation', and the like. If some parameters are out of the correct interval (which is equivalent to the machine being ill), the machine's expression may be "awkward" and "worried". Therefore, with this method, we can give all the emotions that humans possess to the machine. The emotion itself is expressed by the facial expression of the machine and the body language. Similarly, these instinctive emotions of the machine are subject to adjustment by reward and penalty mechanisms. In the life of the machine, the trainer can continuously tell the machine, its emotional expression, which are rewarded and which are punished in different environments or processes. It can also be told directly what the appropriate mood is in a particular or in a process. Of course, its neural network connections may be directly modified to adjust its emotional response. In this way, therefore, the machine can adjust the mood to a similar degree to humans, and further, since the mood and other memory are stored together, in the same memory. When a machine needs a certain result, it mimics the memory that brings that result. For example, if a certain type of behavior is repeatedly brought to a certain result, the machine will imitate the memories containing the behavior and, of course, the emotions in these memories, so that it will adjust its own emotions for a certain purpose. This is a way of mood utilization.

It is to be noted that the machine intelligence established by the method proposed in the present patent application, whose thinking and emotion are visually controllable to human, is fully understandable, and they are connected by associative activation. Therefore, the machine intelligence does not bring danger to human beings, and the method is also a characteristic of the universal artificial intelligence implementation method provided by the invention application.

In the invention, the machine adopts a memory screening mechanism for storing the mirror image space: event-driven mechanisms and temporary repository mechanisms. In the mirror space, every time an event occurs, the machine takes a snapshot of the mirror space and saves the snapshot. The saved content includes features in the image space (including information, machine state, demand, and mood) and their remembered values. Their memory values are positively, but not necessarily linearly, related to the activation values at which storage occurs. A snapshot of the mirror space stores data, which we call a memory frame. They are like movie frames, and by playing back a plurality of frames in succession, we can reproduce the dynamic scene when memory occurs. In contrast, information in a memory frame may be forgotten over time. The occurrence of an event in the mirror image space means that the similarity of the feature combination in the mirror image space is changed beyond a preset value compared with the previous mirror image space, or the memory value in the mirror image space is changed beyond a preset value. The memory bank refers to a database for storing the memory frames. The temporary memory bank is one of the memory banks and aims to screen the information stored in the memory frame. In the temporary memory base, if a certain memory frame contains the characteristic that the memory value reaches the preset standard, the memory frame can be moved to the long-term memory base to be stored. In the application of the invention, the limited-capacity stack is adopted to limit the capacity of the temporary memory library, and the temporary memory library adopts a quick memory and quick forgetting mode to screen the materials to be put into the long-term memory library. Machines lack the motivation for deep analysis of things, scenes, and processes that have been learned by the machine when faced with large amounts of input information, or things, scenes, and processes that are far from the point of interest, so the machine may not recognize these data, or the activation values assigned to them are low. When the machine stores information into the temporary memory base in an event-driven mode, the memory value given by the machine to each information characteristic is positively correlated with the activation value when the storage occurs. Those memories with low memory value may be forgotten from the temporary memory bank quickly and will not enter the long-term memory bank. Therefore, only the information which we concern is put into a long-term memory base, and the trivial things which do not need to extract the connection relation every day are not memorized. In addition, because the capacity of the temporary memory pool is limited, the temporary memory pool also passively accelerates the forgetting speed because the stack capacity is close to saturation.

If we consider the memory as a volume containing innumerable information features, then the relationship network is the context in this volume. These veins appear because of memory and forgetting mechanisms, relationships that cannot be repeatedly activated are forgotten, and relationships that can be repeatedly activated are strengthened. The characteristic diagrams connected by the coarse relationship context constitute the concept. It links the image, voice, text or any other expression form of the same kind of information. Because these expressions appear frequently together and frequently translate into each other, the connections between them are tighter. The tightest local connection relationship forms the basic concept (including static feature map and its language, dynamic feature map and its language); the method is slightly looser than the basic concept, namely static extension concept and dynamic concept extension concept (including concept representing relationship and process characteristic diagram), and is looser than the concept, namely memory. In relational networks, static profiles (or concepts) are usually small parts, dynamic profiles (including concepts representing relations) are connections, and process profiles are large frames, which are organized in a temporal and spatial order of a number of small parts (static objects) and connections (dynamic features). The process features a large framework that we can use for reference. While dynamic profiles (including concepts representing relationships) are tools that can implement empirical generalization, static profiles (or concepts) are objects that are replaced in the generalization. The generalization process is a process of performing simulation by reorganizing real objects and objects in memory after replacing the real objects and the objects in memory through information flow with high activation values.

In order to improve the search efficiency, we can separate the relationship network from the memory and build a single network. One possible approach is to: the characteristic maps in each memory frame are firstly established into connecting lines, and the connecting values of the characteristic maps are functions of the memory values of the characteristic maps at two ends of each connecting line. The join values emitted by each feature map are then normalized. This results in the two characteristic maps not being symmetrical in their connection value with each other. Then, the similarity feature maps between the memory frames are connected according to the degree of similarity, and the connection value is the similarity. After the steps are carried out, the obtained network is the cognitive network extracted from the memory base. The cognitive network can be put into a fast search library (one of the memory libraries) separately for applications that require fast instinctive responses, such as automatic driving applications, or for applications that require only simple intelligence (such as production lines). The memory and forgetting in the relation network adopts a mechanism of memorizing and forgetting connection values: the connection value increases according to the memory curve every time the relation is used. And all connected values are decremented over time according to the forgetting curve. Yet another approach is to place the memory containing frequently invoked concepts and process features, maintaining the organization of the memory library, into a single fast search library. In these memories, the specific details may have been forgotten, and the remaining ones are strong memory values. These strong memory value memories can be activated by association to quickly invoke the relevant concept and process features. This speeds up the memory search efficiency of the machine. This method can also be used in applications where a fast response is required, such as in automotive applications, or in applications where only a simple intelligence is required (such as in a production line).

It should be noted that the establishment of the individual relationship networks can take many forms, but as long as such relationship networks are based on the basic assumptions made in the present application, they are a variant of the relationship networks in the present application, and there is no essential difference from the relationship networks in the present application, so they are still in the claims of the present application.

2.2 implementation of associative ability.

After the memory space exists, the machine can realize the associative capability through 'proximity activation', 'similarity activation' and 'strong memory activation'. Any algorithm that can realize "proximity activation", "similarity activation" and "strong memory activation" can be applied to the present application. Here we propose several methods to implement the above activation principle (but not limited to these methods):

the method comprises the following steps: using memory values (real numbers) to represent the number of neurons or synapses; using the activation value to represent the intensity of an activation electrical signal emitted by the feature; using a particular code to represent different mode activation signals emitted by different features; propagating the activation value using a bus instead of the entire memory space; the three-dimensional stereo coordinate point positions are used to represent the positions of different feature information in the memory space, and the spatial distance (the spatial distance between the activation source and the reception feature) is used to calculate the attenuation amount. When the input characteristics distribute the activating electric signals corresponding to the codes to the bus through the general excitation module and the numbers in the codes are used for representing the initial strength endowed by the input characteristics, the characteristics in the memory can receive the information on the bus through periodically reading the information on the bus and calculate the required attenuation. If there is activation information similar to itself, for example, it may belong to a large class, or a sub-class, etc., then there is a different reception capability. If the activation value obtained after the received activation signal passes through the receiving channel of the feature itself exceeds the activation threshold value preset by the feature itself, the feature takes the received activation value as an initial value and activates itself. There may often be situations where multiple input features simultaneously activate a small memory zone, such as a "table" having multiple features of different resolutions that, in turn, may activate multiple small zones across a bus of memory zones. There may be multiple features activated for each bay that relate to the "table". When the characteristic diagrams concentrated in the cells are activated again, the activation values are given to each other through adjacent activation. Their activation values may "stand out" in the value memory space. Under the mutual adjacent activation effect, a certain cell can activate the memory of a delicious cake on the table at the time. This is because the cake is given a very high activation value to the food-related "positive demand" symbol by means of a taste sensor-related preset program. When memory storage occurs, the activation value of the food-related "positive demand" symbol is converted to a memory value (not necessarily a linear relationship) according to a positive correlation. Therefore, here, the food-related "positive need" symbol (e.g., a need for a savory taste) is a strong memory. It is near the 'dining table' memory, and because its memory value is high, it also can obtain very high activation value according to the 'strong memory activation' principle. When it is activated, the memory "cake" in close proximity to it (since both may be stored in memory at the same time) may also be activated. In addition, and the "cake" and the "requirement for deliciousness" are often activated together, in the memory they have more and more memory, so that at any time, but after one has been activated, the other is also often activated, we establish the correct connection between the "cake" and the "requirement for deliciousness" is fulfilled. In addition, in the application of the invention, the emotion of the machine is realized by presetting a preset program between the condition that the requirement of the machine is met and the emotion of the machine. The preset program can send out higher directional activation values to emotion symbols such as 'joy', 'satisfy', and the like under the input excitation of 'requirement for deliciousness is satisfied'. The emotional symbols of "pleasure" and "satisfaction" of the machine then obtain a higher activation value. When storage occurs, these activation values are also converted into memory values (not necessarily linear) in a positive correlation, so that in these memories, the mood is also memorized. After the machine activates the memory of 'cake' and 'the requirement for deliciousness' the emotional symbols may be activated together, so that the machine can feel 'pleasant' and 'satisfy' emotions.

When the machine needs to seek "pleasure", "satisfaction", etc. (such as giving the machine such instinctive needs), the machine looks for memory related to "pleasure", "satisfaction", which may activate memory of "cake", "table", etc. The memories can become a response target, the machine can possibly obtain experiences of 'cakes' and 'tables' through the association of the targets, further generalize the experiences through generalization capability, organize various process characteristics after generalization through imitating past experiences under the existing condition, subdivide the organized process into a large number of intermediate link targets layer by layer through segmentation imitation, and further realize the intermediate link targets. Such as to complete the process of ordering "cake", finding "table" and fulfilling the needs of the person.

The above process is a process of distributed computing. This method can also be changed to a 2-layer structure. For example, each small segment is memorized with a computing module connected with the bus as a portal for exchanging information with the bus, and the computing module is used for identifying an activation signal outside the jurisdiction and then determining whether the activation signal is transmitted into the jurisdiction or not. And is also responsible for transmitting the activation in the district to the bus again. This is done to reduce the number of computing modules. Of course, this structure can also iterate itself, employing similar multi-layer structures to further reduce the computational modules.

The method 2 comprises the following steps: method 2 is a centralized computing method. It adopts a special calculation module to search the memory (memory search module). Every time an input information characteristic at multiple resolutions is found, the machine directly activates the most recent memories in the current time and assigns corresponding activation values according to their memory values. This completes the proximity activation and the strong memory activation. Related similar features are directly searched in memory, and after the features are found, activation values are directly given to the features according to the similarity. The similarity can adopt a field comparison method or a pre-coding layer-by-layer classification method.

The same method can be used by the memory search module when the activated characteristic diagram sends out the activation electric signal again. By searching the feature map for initiating activation, nearby memories are searched for initiating close activation, those memories farther away with high memory values are searched for initiating strong memory activation, and similarity activation is initiated by searching for similar features in other memories. And each activated module emits an activation electrical signal with its own coding and intensity information. This process may be iterated over and over.

The method 3 comprises the following steps: method 3 is a mixed mode. After the machine completes similarity activation search through the memory search module, further activation can be carried out in a local network of each memory segment. Proximity activation and forced memory activation are achieved through a network of connections established between features in the memory. One implementation of such a local network is: each feature in the memory space establishes a connecting nerve with the adjacent feature, and when the feature is activated, the activation value can be transmitted through the connecting lines, namely the adjacent activation. And the transfer coefficient between the two characteristics is positively correlated with the memory values of the two characteristics, which is the strong memory activation.

All of the above 3 methods can realize the association capability in the memory network. There are many ways to implement "proximity activation", "similarity activation" and "strong memory activation", and various specific ways can be established on the knowledge in the art. The 3 implementations listed in the present application are not, therefore, to be considered as limiting in scope, but rather as demonstrating the principles underlying therein. Any other mode, as long as the associative activation implementation algorithm is established on the basis of 3 principles of 'proximity activation', 'similarity activation' and 'strong memory activation', relates to the claims of the present application. In the machine, the numerical value can be used for representing the strength of information in memory, the code is used for representing the category of an activated electric signal, the bus is used for representing the propagation space of the activated electric signal, and the three-dimensional coordinate distance is used for simulating propagation loss, so that the associative search speed of the machine can be far higher than the neural activation working mode of the brain.

In comparing the similarity of the input feature map to the feature maps in the relationship network, the machine may need to deal with the problems of size scaling and angle matching. A processing method comprises the following steps: (1) the machine memorizes the characteristic diagrams of various angles. The feature map in memory is a simplified map created by extracting the underlying features for each input message. They are common features that retain similarities under the relationship extraction mechanism. Although they are similar to each other, they may have different viewing angles. The machine memorizes the feature maps of the same thing in life but different angles to form different feature maps, but the feature maps can belong to the same concept through learning. (2) The machine uses all the angle views, overlaps the common parts of these profiles, imitates their original data, and combines them to form a stereo profile. (3) And embedding a view changing program for scaling the size and the space of the stereo image into the machine. This step is a well established technique in the industry and will not be described in detail here. (4) When the machine searches for similar underlying features in memory, the method comprises the step of searching for a feature map which can be matched after spatial rotation in memory. Meanwhile, the machine stores the feature map of the current angle into memory and reserves the original angle of view. And the subsequent input of the bottom-layer characteristics similar to the visual angle can be quickly searched. Therefore, in this method, the machine uses a method of combining different visual angle memories and spatial angle rotation to search for similar characteristic maps, which brings about the phenomenon that the familiar visual angle is identified more quickly. Of course, the machine may also use only the method of performing similarity comparison after spatial angle rotation.

And 3, realizing generalization capability.

The generalization capability is based on the multi-resolution concept, so the machine needs to establish the concept at different resolutions first.

3.1 establishment of the underlying concept. Our ancestors invented languages and used these languages to represent categories built by comparing similarities, such as stone, tree, fig, rabbit, lion, etc., that are closely related to life. It is also used to refer to those dynamic classifications established by comparison of similarities, such as running, jumping, tapping, grinding, planing, throwing, and streaming, which are closely related to life. After the languages exist, the languages can be organized in a certain organization mode to express ideas, which is a conventional process. These language symbols, together with the things and actions they represent, constitute the underlying concept.

The specific method of machine building concepts is in the same way as humans. For example, when an image feature map is input into a machine, we synchronize the language that represents the image feature map. The two pieces of information of interest are memorized as adjacent information, and they have a proximity relation. The machine can build a very close relationship between the image feature map and the corresponding language feature map in the relationship network after repeating for many times, and the image feature map and the corresponding language feature map can be activated to generate association with each other and can activate the information of languages or other forms existing in different memory intervals through similarity, so that the association of the concept is realized.

The method of machine building the concept can also be an artificial implanted memory. The language symbols of the concept and other forms of content contained in the concept (which are experiences summarized by human beings) are put together at adjacent positions in the memory and are endowed with higher memory values, so that the language symbols and other forms of information can be activated with each other to generate associations, and the concepts are formed.

Because of the similarity of the similar image feature maps existing in different memories, there may be no high similarity of the language in different memories. When different memories are connected in series through images and languages, language symbols (such as voice or characters) are frequently used (so that the memory value is high) and highly similar to each other (so that the transmission coefficient of the activation value between the memories is large), and the language symbols in information contained in the same concept (such as various apple images, various apple voices and various apple characters) are likely to have the highest memory value (due to frequent use and high similarity). When searching for concepts in memory, machines often first find a token and use the token to represent the concept. When we use languages to express ideas, it is the method (speech or text symbols) that is used to sequentially activate other forms of information (such as images, tastes, senses, sounds, etc.) that are closely related to these languages, and it is the flow of these other forms of information that lets us understand the information represented by the language. It is important to note that the image and other perceptual form information streams created by the linguistic information streams can be made part of our memory as if they were actually occurring. This is because the speech information stream activates the image and other perceptual form information streams, as do the external input information. Therefore, both will bring new memory and memory value, and will also affect the memory value in the original memory.

And (5) establishing an extension concept.

In the application of language, we must combine the frequently used information, use a symbol to represent, and form consensus among the groups. Thus, when information is exchanged, the symbol can be used to simply represent the string of information combination. This is to create a new concept on the basis of the basic concept.

We can attribute different underlying concepts to one concept because these different underlying concepts contain some common attribute. The common attribute is established by gradual induction in life. For example, in our ancestry memory, "hunting" may activate "spears" and may also activate "stone axes". The "spear" and "stone axe" establish a closer association in memory through "hunting". Our ancestors created the language notation "weapon" to refer to all tools associated with "hunting" activities, perhaps to express this association more conveniently. If a partition wall falls down to a 'pistol' of our ancestor, it is told to be a 'weapon'. Although he has no experience of using the thing, by the property of the weapon, our ancestors borrow the experience related to the weapon and throw the 'pistol' as a stone to the animal when hunting. Because handguns and stones are the same thing in the weapon after the resolution is reduced, the experience of using them can be generalized. This is the generalization principle of "the same property can be substituted under the same concept" presented by us.

For example, we can refer to people who have dinner as customers and various amounts of money which the customers add to us as tip fees, which reduces the resolution of things and only retains their common attributes, so they are similar to each other and are summarized as a concept. Similarly, we also classify apples into red fuji apples, american snake apples and tobacco apple. This is to increase the resolution of the things to distinguish the differences. Such new concepts can be built because these things share common features. Which are features that are common to some action, scenario, or process.

The extension of the dynamic concept is also to create new dynamic feature classifications by increasing or decreasing the resolution. For example, "running" and "dancing" are collectively referred to as "sports," and "running" is classified into "fast running," "slow running," and "long running," and the like. These are also new classes established by different attributes of the dynamic profiles and new tokens are created to represent these classes.

The concept of expansion, therefore, is a class of tokens in our memory that exist by establishing a more intimate association with other tokens whose content is represented by common features in the content represented by the other tokens. For humans, a large number of extended concepts have been summarized in human historical development. In our lives today, most concepts are obtained through learning (the summary result of predecessors is directly obtained), and few concepts are established through establishing close relation between similar things and a certain language symbol in memory. The same method can be used for learning the machine. For example, many concepts of humans are learned by interpretation. For example, when a scene or feeling of "happiness" appears, we are informed that this is "happiness". As another example, we may learn the interpretation of the concept "solar system" through a dictionary. Likewise, machine learning may also do so. We can directly give the extended concepts and their explanations to the machine:

the method comprises the following steps: directly let the machine learn. For example, let the machine learn the meaning included in a concept through words and voice. Putting this information into memory, it is sufficient for the machine to establish connections of similar activations, proximity activations and forced memory activations between these concepts and their interpretation by repeated learning.

The method 2 comprises the following steps: a section of 'fake' machine memory is directly established, and connection relations among similar activation, adjacent activation and strong memory activation are artificially given to the machine in the memory (for example, high memory values are given to relevant information, memory positions are put together, the machine can be helped to find similar characteristics more quickly according to coding of the information, and the like).

In this way, when inputting these extended concept tokens, the machine can associate the basic concepts it contains, and then other forms of information (such as images, sounds, smells, touch, emotions, feelings, etc.), and use these forms of information to add to the information stream formed by the tokens, and the machine can search for similar information streams by association, thereby using similar information streams in the past to infer the cause and possible outcome of the information stream. And the causality is put into the demand and emotion assessment system to determine the response, which is intelligent.

And 3.2, realizing generalization.

If the same in-memory information stream as the real input information stream cannot be found, generalization is required. The idea of generalization is to estimate the cause and result of a real input information flow from similar information flows in memory. In the present application, generalized tools are primarily concepts that represent motion characteristics. Because the dynamic feature is a dynamic motion mode, the main body of the dynamic feature is a generalized main body. Machines may use particles or stereo graphics to represent abstract bodies of motion. Just because the motion body is a generalized body, the machine can bring similar concepts into the motion characteristics, thereby realizing the generalization capability of experience. While such concepts of uniformity may be intuitive and dissimilar, they may be generalized to the same category by reducing resolution.

The concept of representing relationships between things is also a dynamic feature. It considers the objects at both ends of the relation as a virtual whole. Therefore, in the present application, by assigning a dynamic feature to a concept representing a relationship, a machine can correctly use the concept representing a relationship by using the dynamic feature. For example, the relationship represented by the language "though.. a.," though., "though.," etc. may be represented using a dynamic feature of a turn. The parallel concept of "one side.", "both. The concept of a relationship "included in.

The specific establishment method of the relation dynamic characteristics comprises the following steps: 1, machines find their common points, which are usually concepts representing dynamic patterns or relationships, by using a memory and forgetting mechanism for a large number of languages, because of their independence from specific objects, they can be widely used. The words are gradually organized into commonly used words, commonly used sentence patterns, grammars and the like. The method is similar to the language organization method in the current artificial intelligence and is a mechanical simulation method. 2, in the present application, the machine needs to further understand the meaning of these concepts. The machine understanding method is to memorize the specific static characteristic diagram and dynamic characteristic diagram associated with each use of the concepts and then store the concepts through a memory and forget mechanism. Because specific objects always change during the process of describing relationships (they only exist in similarities at coarse resolution, because only those features at coarse resolution are common features through memory and forgetting mechanisms, they are very broad objects, being able to bring a large number of different objects into the same action or relationship), while invariant is the dynamic features that represent relationships. A relational application such as "one-sided.. another-sided" is often used in the dynamic nature of two objects moving side-by-side. So, after accumulation, the machine can represent the words of the expression relationship of "one side.

Another aspect of dynamic expansion is: in our lives, there are many processes, which are generalized motion patterns composed of multiple entity concepts or extended abstract concepts, and we refer to these process features. The process characteristic is an expanded dynamic characteristic, and is characterized in that: 1, a plurality of observation objects, which are not necessarily a whole. 2, the entire motion pattern has no clearly repeated trajectory. Such as the processes of returning home, going on a business trip, washing hands, cooking and the like, which are a plurality of physical concepts or extended abstract concepts, form a generalized movement pattern. The pattern is so called because these concepts are continuously repeatable in our lives. Since repetition is possible, there are common features in describing the process of representing these concepts, and otherwise, we cannot represent them with one concept.

Process features are typically dynamic processes involving large spaces and long times. The specific details for implementing it are closely related to the environment, so it is difficult to find similarities from it. However, these links are usually represented by language symbols. Therefore, when searching for a process feature, we can first search for the repetition of the token for each link. For example, the machine forms a tower-shaped conceptual relationship which is gradually expanded by memorizing language symbols corresponding to each link when the machine goes to an airport each time. For example, the following steps are carried out: the top layer of the concept is 'going to the airport', the next layer is 'ready to go', 'on the way' and 'to arrive', and the next layer is 'ready to luggage', 'finding a car', 'telling a friend', 'sitting in a car', 'on the way', 'arriving at the airport garage', 'going out of the garage' and 'arriving at the airport entrance'. The next layer is "ready for clothes", "ready for toiletries", "ready for money", and "materials relevant to the preparation work". This process can be continually refined. Initially, the distinction of each link may be arbitrary. But we get a tower-shaped conceptual organization after each visit to the airport. The tower-shaped concept organization is subjected to a memory and forgetting mechanism, and finally, only a small amount of necessary concepts which frequently appear can be kept in the memory on each resolution level. They are the process features at the corresponding resolution. These process features are a series of concepts organized with temporal and spatial order. Especially at the bottom level, it is usually only possible to leave a static and a dynamic profile every time it goes to the airport. These profiles are few in number, but they are lacking. These are static or dynamic profiles representing key links, such as "security check" or "check-in". The upper level concepts associated with the key links are also absent (they may be fewer in number). And by sequentially pushing upwards, only the concept of 'going to the airport' is finally realized. Therefore, the establishment process is characterized by being realized by a memory and forgetting mechanism from positive selection (a link which is deliberately memorized by the experience of other people for self reference) and negative selection (an upper link corresponding to all things at a time).

These remaining tower concepts and underlying feature maps are the objects we have simulated each time we go to the airport. By only putting specific things in the real environment into the process characteristic according to an analogy method, the target planning capacity of each stage of going to the airport from any place can be built. In the concrete implementation, the abstract concepts need to be expanded layer by using segmentation simulation, and more links meeting the actual situation are added. Thus we have established the ability of machines to move down airports in a variety of different environments. This capability is an empirical generalization. The whole process is a process of repeated iteration of simulation and generalization.

After the above basis is established, the machine can be generalized in a manner understood by the computer. The generalization process is described below by taking as an example the instruction that the machine receives "go to airport". After the machine receives the instruction of going to the airport, the information (possibly action images and feelings) related to the language symbol of going to the airport is activated through association, and the information (possibly airport images) related to the language symbol of going to the airport is also activated. If the machine ever goes to an airport, the two activation points will directly activate a large series of other memories through 'proximity activation', 'similar activation' and 'strong memory activation', and the memories may contain static images, dynamic images, voice and words, and also feelings and emotions. They are all time and space related in memory. The machine only needs to take the string of information as an intermediate link target according to the time and space relation, and gradually imitates and realizes the target, so that the purpose of going to an airport can be achieved. If the machine has no experience going to the airport, then the relevant images of "going to train station", "going to shop", "going to travel" and also the relevant image of "airport" may be activated. The machine, by comparison, finds that the activation value of "train station" is the highest among these images, except for "airport". This is because there are multiple activation channels that deliver activation values to the "train station". Experience such as "go." activates the presence of similarities in rough resolution to go to "train stations," such as "airports" and "train stations. For example, by learning the experience of others, "both airports and train stations are places where vehicles are to be taken," this approach would also pass activation values from "airports" to "train stations. Therefore, on the basis of establishing the extended static and dynamic concepts, the machine can link a series of bottom static or dynamic non-language information (including sound) through the connection relation between the concepts to the input information, and the information can adopt different resolutions to compare the similarity so as to realize the similarity activation, the proximity activation and the strong memory activation. This activated non-verbal information will then act again as a bridge through the linguistic concept, activating other memories that are not similar, or adjacent, or whose memory value is not very high. When all the associated activations are completed, the machine selects the information with the highest activation values of 1-N (natural numbers) and organizes the information according to the time sequence or the space sequence of the information, and the information is the generalization experience which the machine needs to imitate. The generalization capability does not require the machine to deliberately build. The machine only needs to establish correct concepts and relationships between the concepts, and through associative activation, the generalized experience that can be utilized can emerge automatically. In the present application, two keys that can achieve this are: 1, action features and relationship concepts need to be extracted separately, unhooking them from specific objects. 2, establishing a reasonable relationship network to realize correct association capability, wherein the association capability must be quantifiable.

And 4, realizing the requirement and the evaluation capability.

When inputting information, the machine first finds one or more most relevant memories in the memory, which are a series of information flows related to the input information, and this is achieved through associative activation. These memories are responses of past machines to similar input information, or responses of past machines to a plurality of information locally similar to the input information. The originator of these responses may be the machine itself, or something else. The machine takes as the destination of the information source the response associated with the input information that occurs the most frequently between itself and the information source. If there is not frequent interaction between the machine and the information source, the machine considers the response most used by others as the purpose of the information source to send out the information. This is reasonable because the information source sends out information for the purpose of getting a response. The information source has preset possible responses based on its own experience. These predetermined responses are established based on the interaction of the information source with the machine or the interaction experience of the information source from others. This can be done using preset experience. When a machine understands the purpose of the information source, it also understands the input information.

After the machine understands the purpose of the information source, the machine needs to establish a corresponding response. The method for the machine to establish the response is as follows: 4.1, the machine finds one or more segments that can be used as a reference to build a memory of responses. The specific method comprises the following steps: the machine converts the input language information into non-language information streams by means of associative activation, and uses these information streams and other input non-language information as total input information.

4.2, the machine uses the converted information flow as new virtual input information, and uses the association activation for the information again. After the completion of the associated activation, information about the cause is information about the cause at a time before the similar information is input, and information about the result is information about the result after the similar information is input.

4.3, the machine may find multiple pieces of memory about itself, or may find multiple pieces of memory about others. The real need of the machine is to find the change of state of the demand and the change of state of the emotion stored with these memories. If the memory of the other person is concerned, the machine needs to replace the activity of the other person with the activity of the machine as a new virtual input again, then the memory related to the virtual input cause and the memory related to the result are searched again through associative activation, and the change process of the demand state and the change process of the emotional state in the memories are searched.

4.4, the machine takes one or more sections of memory tissues with the highest total activation value in the step 4.3 as one or more sections of responses to serve as the objects of simulation. Because these activated messages have their own time and relationship relationships (which are all synchronously stored in memory), they may be organized into one or more responses.

4.5, analyzing by the machine whether the change process of the demand state and the change process of the emotional state of the user in the one or more pieces of response-related memories to be imitated are positive or negative. These one or more responses are selected according to the evaluation of the principle of "driving toward and avoiding harm". Since these memories contain changes in "demand values" and changes in "mood values", a simple statistical algorithm is required to achieve the "welfare and aversion" selection. These algorithms may be preset.

4.6, the machine combines one or more segments of responses that pass the "Trend" assessment into one large process. Temporal and spatial organization information may exist in the responses, and the organization proceeds according to the temporal and spatial information. If there is no explicit temporal or spatial order information in the responses, the machine needs to input the responses again as a new virtual information to find more memory to find their temporal or spatial order by associative activation. This process iterates until the order of the responses can be determined (it is also possible to find the order of the responses by memory to be arbitrary).

4.7, the machine selects response under the driving of the motive machine.

The source of power to drive the machine is the motive of the machine, which may be summarized as "hedging". The part of the 'benefit' and 'harm' is preset; some are established in acquired learning because they are related to the needs of the machine itself. Compared with human beings, such as 'water', 'milk', 'food' is 'interest' preset in nature at the beginning, then the relation between 'test score', 'bank note' and our innate needs is obtained by learning, then we also find that the operation object can be 'love' and 'time', and even we also pursue dominance in the population, which is an extension of underlying motivation 'goal achievement' existing in our genes. In a similar way, we can also give the machine the motivation we wish it to have. Because in our relational network, when all memory frames are stored, the required symbols of the machines at that time and the corresponding memory values are stored at the same time. These memory values are positively correlated with the state values of the demand symbols at that time. For example, if a machine is responsible after some action. Because the responsibility is a loss (the experience can be preset, expressed by the language of the trainer, or realized by directly modifying the relationship network), and the degree of responsibility (such as a word representing the degree in the language) brings different loss values to the machine. The stronger the accountability, the higher the memory value that the machine assigns to the symbols lost in this memory. Then in this memory, since the memory value of the loss symbol is higher, all other feature maps with higher memory values in this memory frame have stronger connection with the loss symbol. If similar actions send out an object or receive an object in a similar environment and similar responsible behaviors occur again, the characteristic diagram causing loss and the loss symbol in the memory frame are repeated, and the memory values of the characteristic diagram and the loss symbol are increased according to the memory curve in the memory frame, so that the relationship between the characteristic diagram causing loss and the loss symbol is increased. Through repeated times, the relationship between the characteristic diagram actually bringing loss and the loss symbol is selected according to a memory and forgetting mechanism. It is unclear from the outset as to what is being blamed, and it is clear to the latter as to what has been blamed for itself as a consequence of the blame. This process is similar to the learning process of human children.

Similarly, the return value, safety value, risk value, goal achievement value, dominance value, etc. of the machine are similar. They are all through the machine's experience in the past, constantly linking behavior and behavior results together. The method of linking together is to put them into the same memory frame. Even if the machine is not fed back in time when the action takes place. The trainer may also at a later stage point out the behavior itself and give feedback, thus linking the behavior and the result in a single memory frame. The trainer does not even need to indicate which behavior is good or bad, and the machine can gradually establish the connection relationship between the correct behavior and the required value by receiving correct feedback every time and memorizing and forgetting. Such as those that must receive a reward or penalty, which are simultaneously memorized each time they occur. With each repetition, their memory increases and eventually the connection between them becomes tighter and tighter than the others.

The machine evaluation system is a preset program. This process determines whether a virtual output is to be converted into a real output based on the state of satisfaction of the machine's demand for revenue and loss values, security and risk values, goal achievement values, dominance values, etc. These types of requirements are imposed on the machine by humans. Of course, we can give machines more human beings to expect their own goals, such as "follow robot convention", "follow human law", "be rich in congruence", "be told", "be behaving elegance", and so on. These goals can be achieved by setting the demand symbol in memory and adjusting the behavior of the machine through trainer feedback. It should be noted that these goals can be increased or decreased as desired by humans. And the addition or subtraction of these objects does not affect the claims of the present application.

For better communication with humans. The invention provides a method for converting the actual satisfied state of the machine requirement into the emotion of the machine by using a preset program and taking the actual satisfied state of the machine requirement as the input of an emotion system. The purpose of this is personification, which simulates the emotional response of a human being in different states of satisfying the requirements. Only then does the machine communicate better with humans. Meanwhile, the following method is adopted to realize that the machine can achieve the purpose by utilizing the emotion of the machine: 1, the machine synchronously stores own emotion every time the machine stores memory. 2, the trainer needs to make feedback on the mood of the machine. Through the feedback of the trainer, the machine determines how the mood should be adjusted. 3, the machine can modify the parameters of the preset program by itself and output the emotion according to the experience of the machine. With the above 3 points, the machine can link emotion and feedback. Such emotions are both a way of expression and a means that can be utilized. Since specific emotions are linked to specific external feedback. In the search for specific feedback, the machine may be remembered to be a mimic of what the machine expects to reproduce a particular result. It should be noted that the variety and intensity of the emotion can be increased or decreased according to the human desire. And the addition or subtraction of these objects does not affect the claims of the present application.

The evaluation values established by the machine also need to be combined with the internal state values of the machine (such as whether the machine is in short of power, whether the machine is in a system failure, etc.) to make a judgment, and the judgment result is passed or not passed. The machine's evaluation system is a pre-set program. The method is a link for endowing the machine with individuation, and different choices are equivalent to different characters. The machine can also reserve some parameters which can be adjusted by the machine, and different results are brought by trying different choices, so that an evaluation system which best meets the requirements of the machine is built step by step. This step can be achieved by known techniques and is not described in detail here.

If the machine establishes a response, it cannot pass the evaluation system. The machine needs to re-establish the response and remove the behavior of the last evaluation that brought about the negative consequences of significant loss, danger, etc. These behaviors are those resulting from the combination of static and dynamic profiles that contribute to the penalty. Removing negative behavior is also a more complex machine thinking process. In the process, the machine needs to convert all current targets into inherited targets, and the computing power is left free for removing the computation of a temporary target such as a negative surface behavior. The machine then needs to look for all the memory about this negative behavior, finding out from it the experience of how to exclude it. After removing the actions that bring negative results, the machine re-establishes a new response. The established process is still the preferred dynamic feature map, the concept replaces the static feature map, and then the combination mode of the static feature map and the static feature map is determined by means of similar memory. If the machine iterates multiple times, it is still unable to establish a response that can pass the evaluation. There is a possibility that there is an error in the previous step or that the machine encounters an unsolved problem. At this point the machine enters processing for the "information not available for processing" flow. That is, the "information cannot be processed" itself is a result of processing the information. The machine builds a response to the "information cannot be processed" based on its own experience. These responses may be left alone, may again confirm the information with the information source, or again use a higher resolution to identify the information, etc. These are also all reasonable responses similar to human behavior.

In the above process, the machine needs to use the associative activation process repeatedly. It is important to note here that even if the activation value transfer coefficient between feature maps is linear due to the presence of the activation threshold, the activation value accumulation function of the feature maps is linear, but due to the presence of the activation threshold, the same feature map and the same initial activation value, whether during a single associative activation or during multiple associative activations, are different, but because the activation order is chosen differently, the final activation value distribution is different. This is due to the non-linearity brought about by the presence of the activation threshold. The information loss caused by different transmission paths is different. The preference of the selection of the order is activated, which corresponds to the difference of the machine personality, so that different thinking results are produced under the same input information, and the phenomenon is consistent with the human being.

In addition, the relationship strength and the latest memorized value (or connection value) in the relationship network are correlated. The machine will have a first-come-first phenomenon. For example, two machines with the same relationship network confront the same feature map and the same initial activation value, wherein one machine suddenly processes an input message about the feature map, the machine updates the relevant part of the relationship network after processing the additional message. One of the relationship lines may increase according to a memory curve. This increased memory value does not subside in a short time. Therefore, when facing the same feature map and the same initial activation value, the machine processing the extra information will propagate more activation values along the just enhanced relationship line, thereby leading to the phenomenon of first-come-first.

In addition, in order to reasonably process the information input sequence and ensure that the activation value brought by the information input later is not shielded by the information input earlier, in the invention application, the activation value in the association activation is decreased along with the time. Because if the activation value in the relationship network does not fade over time, the change in activation value by the following information is not significant enough, which may lead to inter-information interference. If the activation value is not faded, the subsequent information input is strongly interfered by the previous information, so that the user cannot correctly find the attention point. But if we completely empty the memory value of the previous information, we lose the connection relation which may exist between the previous information and the next information. Therefore, in the present invention, we propose to use a progressive fading method to achieve the balance between the isolation and concatenation of the front and back segment information. This regression parameter needs to be preferred in practice. But this presents the problem of maintaining the active state of a message. If the machine cannot find out a response scheme satisfying the machine evaluation system in the course of thinking, the activation values fade out as time passes, so that the machine forgets the relevant information and forgets what the machine wants to do. At this point the machine needs to refresh the active value in memory again. One brushing method is as follows: the information with the highest activation value is converted into virtual output, the virtual output is used as information input, and the process is repeated to emphasize the concerns, namely why the people like to have a praise in thinking and a thought is not understood or cannot find the thought in some cases, or the people like to have a praise in mind. The virtual input can search for memory and update memory values as the real input process. Therefore, this method can be used by machines to intentionally add memory to certain information. This is to use a reading or memory enhancement method. In addition, in this case, if new input information occurs, the machine has to interrupt the thought process to process the new information. Therefore, from an energy saving perspective, machines tend to be thinking-free, avoiding waste. At this point, the machine may actively send out buffered auxiliary words such as "take … or …" to send out output information indicating that it is thinking and not disturbing. There is also a possibility that the machine may be given a limited amount of thought time or too much information, and the machine needs to complete the information response as soon as possible, and the machine may also use the output to input. In such a manner, the device emphasizes useful information and suppresses interference information. These modes are commonly used by humans, and in the present application we also introduce it into the machine's mind. The machine can determine whether the current thinking time exceeds the normal time according to a built-in program, or the experience of the machine, or the mixture of the two, and the attention information needs to be refreshed, or other people are told to think by the machine, or emphasis is given to the emphasis, and the interference information is eliminated.

Since the most frequent human communication is voice and text, in a concept local network, when other feature maps obtain activation values from each branch of the relationship network and transmit the activation values to voice or text, the common focus is on the voice and text of the concept. Therefore, in the method of filtering or emphasizing self-information of the machine, the virtual output is usually voice, because the method is the most common output mode. The machines output them with minimal energy consumption. This, of course, is closely related to the growth process of a person. For example, a person who learns from a book may convert information into characters and input the characters again.

The search method using associative activation utilizes the implicit connection relation among languages, texts, images, environments, memories and other sensor input information to mutually transmit activation values, so that the related characteristic diagram, concept and memory are supported and highlighted by each other. The difference between the method and the traditional 'context' for identifying information is that the traditional identification method needs manual work to establish a 'context' relational library in advance. In the present application, we propose the basic assumption of "similarity, there is an implicit connection between information in the same environment". On the basis of the assumption, the relations of the shapes and the colors are simplified, so that the machines can build a relation network by themselves. It does not only contain semantics but also common sense. It is to be noted here that the associative activation is a search method, which is not a necessary step in the present application per se, and can be replaced by other search methods for achieving a similar purpose. When the associative activation is used, the machine can regard the characteristic diagram with the activation value exceeding the preset value in each memory as being used once, and maintain the memory value of the characteristic diagram according to the memory and forgetting mechanism in the memory bank to which the memory belongs.

And 5, establishing response execution capacity.

Mimicking is the ability of a human to be present in a gene. For example, for a child to say an calandering, if we say "you are back" with his (her) callout every time he (she) comes home. After a few times, when he (she) comes home again, he (she) will actively say "you come back". This indicates that he (she) has begun to imitate others for learning without understanding the meaning of the information. Similarly, we let machine learning use the same approach. The machine also mimics the experience of others or oneself to understand and respond to the input information.

The step of performing a response is a translation process. If the machine selects the voice output in the selection of various possible response steps, the machine is simple, and only the image feature diagram to be output is converted into voice, and then a concept replacement method is adopted to combine the dynamic feature diagram (including the concept representing the relationship) and the static concept by utilizing the relationship network and the memory to organize the language output sequence and call the pronunciation experience for implementation. It should be noted that the machine might employ some form of experience (either self-contained or experienced by others) to express the dynamic nature of an entire sentence (e.g., using different movement patterns of tone, or accent to express questions, jeers, distrust, emphasis, etc., in a human-used manner). Because the machine learns the expressions from human life, the expressions of human beings and the machine can learn theoretically.

The problem is much more complicated if the machine chooses to output motion, or a mixture of speech and motion. This corresponds to the tissue moving about. In response planning of a machine, there may be only major steps and final goals, the rest requiring random strain in practice.

5.1, the machine needs to target the sequence of image feature maps to be output (this is the intermediate and final targets), according to which different times and spaces are involved. Machines need to divide them in time and space in order to coordinate their own execution efficiency. The approach taken is by selecting as a group closely related targets in time and closely related targets in space. Because the dynamic characteristic diagram and the static characteristic diagram are combined to form an information combination, and the environment space of the related memory is provided with time and space information, the classification method can be adopted in the step. (this step corresponds to the change from the general scenario to the minute scenario).

5.2, the machine needs to expand the intermediate targets in each link layer by adopting a segmented simulation method by combining the real environment again. Response plans developed by machines at the top level are usually composed only using highly generalized process features and highly generalized static concepts (since these highly generalized processes can find many similar memories, responses built by them are highly generalized). Below the total output response, such as "business trip," the "go airport" is an intermediate link target. But this goal is still very abstract and the machine cannot perform emulation.

Therefore, the machine needs to be divided according to time and space, and links needing to be executed in the current time and space are taken as the current targets. And temporarily putting other time and space targets to one side as inheritance targets. After the machine targets the middle link, the machine still needs to further subdivide the time and space (write the lower level script again). This is a process of increasing temporal and spatial resolution. The process of converting one target into a plurality of intermediate link targets by a machine is still a process of creating various possible responses, evaluating the responses by using an evaluation system and selecting own responses according to the principle of 'driving and avoiding'. The above process is iterative, and the process of dividing each target into a plurality of intermediate targets is a completely similar processing flow. Until the underlying experience of the machine is resolved. The underlying experience is that for language it is the muscles that are mobilized to make syllables. For an action, it is decomposed into the issuing of drive commands to the relevant "muscles". This is a tower-like decomposition structure. The machine starts from the top level target and decomposes one target into a plurality of intermediate link targets. This process is to create virtual intermediate process targets that are retained if they "meet" requirements. If "not compliant," it is recreated. The process is developed layer by layer, and finally colorful responses of the machine are established.

5.3, in the process, the machine may encounter new information at any time, resulting in the need for the machine to process various information, and these original goals become legacy motives. This is equivalent to the situation that new situations are encountered continuously in the process of organizing activities, and the problems need to be solved immediately, and the activities cannot be organized. The director then calls off other activities to resolve the problem that was encountered in front. After resolution, the activity continues. In another case, the director suddenly receives a new task during the process, and then decides to suspend the activity after the director balances the interest and the disadvantage, and processes the new task preferentially.

5.4, the machine decomposes other objects to more detailed objects while performing simulation tasks that can be performed. The machine is thought at the same time. This is because the reality is very different and it is impossible for the machine to know the external situation in advance and make a plan. It is the process of an environment and machine interaction to accomplish an objective.

By this, the machine has completed understanding and responding to a single input of information. This process, as a minimum period of machine and environment interaction, is continually reused to accomplish larger goals.

And 6, updating the memory bank.

The updating memory base is in all steps, and is not a single step and is the realization of a relation extraction mechanism. In step S1, the underlying features are established mainly by using a memory and forgetting mechanism. And the machine increases the memory value of the local feature according to the memory curve when finding a similar local feature through the local view and if the similar local feature exists in the feature map library. If there is no similar local feature in the feature map library, it is stored in the feature map and given an initial memory value. The memory values in all feature map libraries are gradually decreased according to the forgetting curve along with the time or training time (increasing with the number of training samples). Finally, the common simple features which are widely existed in various things can have high memory values and become the bottom layer feature map.

In step S2, each time an underlying feature or feature map is found, if there is already a similar underlying feature or feature map in the temporary memory library, feature map library or memory, its memory value is increased according to the memory curve. They also follow the forgetting mechanism. In step S2, the machine first stores the environment space in a temporary memory. When the machine stores the environment spaces in the memory base, the characteristic maps in the environment spaces and the memory values of the characteristic maps are simultaneously stored, and the initial memory values of the characteristic maps are positively correlated with the activation values when the storage of the characteristic maps occurs. In steps S3, S4, S5 and S6, the memory values of the feature map in the memory library comply with a memory and forgetting mechanism. And when one relation in the memory is used once, the memory value of the characteristic diagram related to the relation is increased according to the memory curve, and simultaneously all the characteristic diagrams forget the memory value according to the forgetting curve of the memory bank where the characteristic diagram is located.

In the present invention, we can use various forms of memory organization, such as:

and 6.1, directly adopting the time and space relation of information input, sequentially storing, and establishing a three-dimensional coordinate to express the distance between information. The timeline for this coordinate can follow an event-driven mechanism: the timeline is incremented by one unit each time an event-driven, memory is stored.

6.2, establishing serial numbers for the characteristics, and corresponding each serial number to the characteristics in a form of a table. In the memory space, a code is used instead of the feature (or the feature itself is used but with the code attached). The codes can be classified layer by layer according to the similarity, and the machine can quickly find similar characteristics only according to the classification information of the codes.

6.3, put similar features together, but each feature has its own stereo coordinates in memory space. Therefore, the machine can quickly find all similar features and realize proximity activation and strong memory activation according to the space coordinate information of the features.

6.4, modeling the brain nerve tissue, a connective relationship is established between adjacent memories. Propagation and attenuation of the activation electrical signal is mimicked by this connection relationship. Meanwhile, each characteristic receiving and activating electric signal also simulates cerebral nerves, the characteristic receiving capacity with high memory value is strong, and the characteristic receiving capacity is positively correlated with the matching degree of the activating electric signal and the characteristic receiving capacity.

6.5, and can also be a combination of the forms.

Whatever form of information storage organization is used, it is a specific embodiment of the method proposed in the present application, as long as the purpose of the organization is to implement the associative activation process.

The memory value in the memory banks can be refreshed in different ways, for example, a separate memory value refreshing module is used in each memory bank, or a memory value refreshing module is used in the whole machine, or memory refreshing is realized by a program or hardware for realizing the associative activation process, and as long as the purpose of the memory refreshing is to realize the memory and forgetting mechanism for the memory banks similar to the memory bank in the present application, the memory refreshing module is a specific implementation manner of the method proposed in the present application.

7, implementation schematic diagram.

In the application of the invention, the most computation amount is the two processes of multi-resolution feature extraction and associative activation. It is proposed in the present application that a separate hardware algorithm can be used to improve the machine operation efficiency. FIG. 3 is a block diagram of a general machine intelligence implementation. The core idea of the method shown in fig. 3 is to use a single module to implement the associative activation process. After the module obtains the input information, the module searches and memorizes the information to find the information adjacent to the input information, the similarity information and the strong memory value information. Then, the activation value is directly given to the information in the memory according to the corresponding algorithm. Then, the information in the memory is directly endowed with the activation value according to the corresponding algorithm again by searching the adjacent information, the similarity information and the strong memory value information related to the activated information. This process iterates until the associative activation process stops because an activation preset exists for each message. The method adopts a preset algorithm from the perspective of 'god of Shangdu', and directly adopts an external algorithm to complete associative activation according to the spatial distance, the memory value and the similarity of a memory space. The algorithm for fading the memory value and the activation value in the memory over time can be refreshed by using software or hardware in a memory bank, and can also be realized by using a memory association activation module. Wherein S600 is a build machine feature extraction module. The module selects static features and dynamic features of data under different resolutions by comparing local similarity, and establishes comparison similarity or trains a neural network, or any other existing algorithm to extract features of data. Wherein the modules S601 and S602 are modules for extracting multi-resolution information features from external input information, which relate to different resolutions. Machines may need to perform feature extraction on input data at multiple resolutions. In S601, different features of the data may be extracted by preprocessing the same sensor data into multiple data. In S602, data features at different resolutions may be extracted by using different preprocessing algorithms again at different resolutions. To improve efficiency, S601 and S602 may employ separate hardware to perform the multi-resolution feature extraction function. After completing the input information extraction, the machine may include two modules in S603. One of them is for the associative activation of a dedicated module, which may be a dedicated piece of search hardware. The purpose of this is to solidify the search memory and assign activation value algorithm, improving efficiency by using specialized hardware. The other is a module for combining memory information and reality information, which is equivalent to software for realizing data reorganization. In the step, the experience is generalized mainly by searching dynamic processes from related memories and then by the generalization ability of action characteristics. S604 is the entire memory bank (including the fast search bank, which is created to improve the search efficiency, and which contains the common memory information, and also contains the temporary memory bank, the long-term memory bank, and possibly other memory banks). The memory bank corresponds to a storage space, but it carries a life cycle (memory value) of each information. The memory banks may employ a dedicated memory value refresh module to maintain the memory values. S605 is a demand evaluation system which makes a logical judgment using the demand value obtained in the process of S603. S605 may be a software implementation. S606 is a piecewise simulation process (a process of iteratively developing concepts), which requires constant calls to S603 and S604, which may be a software implementation. S607 is a logical decision, which may be a software implementation. S608 is a new memorized storage process, which can be implemented by software or by using special hardware. The new memory includes internal and external input information of the machine, demand information of the machine, and emotional information of the machine. They are first stored in a temporary memory bank. S609 is a state of completing one information response cycle. In the embodiment of fig. 3, it is characterized in that the multi-resolution feature extraction can be implemented using separate hardware and the associative activation process can be implemented using separate hardware.

FIG. 4 is a schematic diagram of another module for implementing general machine intelligence. The core idea of the method shown in fig. 4 is to distributively integrate the algorithm implementing the associative activation process in the memory base module. In fig. 4, S704 is a memory bank that can mimic brain memory function, and implement proximity activation, strong memory activation, and similarity activation functions. The brain-simulated type memory device receives the excitation electric signals transmitted by the characteristics in a brain-simulated mode, realizes the transmission and attenuation of the excitation electric signals in the memory according to the distance of a memory space, and simultaneously simulates the brain to realize strong memory activation. The memory module itself can also integrate the search algorithm to realize the similarity activation, and there can be many methods to realize the similarity activation. They need to be embodied according to different memory database data organization methods. The rest of fig. 4 is the same as in fig. 3.

Of course, the manner of implementing the associative activation function may also be implemented in a manner similar to the centralized in fig. 3 and the distributed establishment hybrid in fig. 4.

Claims (14)

1. A method for finding a memory associated with an input message in a memory, comprising:

when information is input, methods of 'near activation', 'strong memory activation' and 'similarity activation' are adopted to search for memory related to input information.

2. A method of storing information in a memory, comprising:

when the machine carries out multi-resolution feature extraction on input information and stores the multi-resolution features, the information features are stored in a memory base according to the original similarity relationship, time and space relationship among reserved information; machines use values or symbols to indicate the time at which such information can exist in a memory base, which are called memory values; the memory values may be stored with their corresponding characteristics or may be stored separately.

3. The method of claim 2, wherein the organizing of the memory information comprises:

directly adopting the time and space relation of information input, storing the information in sequence, and establishing a three-dimensional coordinate to express the distance between the information; the timeline for this coordinate can follow an event-driven mechanism: the timeline is incremented by one unit each time an event-driven, memory is stored.

4. The method of claim 2, wherein the organizing of the memory information comprises:

establishing codes for input characteristics, wherein each code corresponds to the characteristics in a form of a table; in the memory space, the features are replaced with codes (or the features themselves are used, but with codes attached); the codes can be classified layer by layer according to the similarity, and the machine can quickly find similar characteristics only according to the classification information of the codes.

5. The method of claim 2, wherein the organizing of the memory information comprises:

similar features are put together, but each feature has its own stereo coordinates in memory space.

6. The method of claim 2, wherein the organizing of the memory information comprises:

establishing a connection relationship between adjacent information in the memory, wherein the connection relationship simulates the propagation and attenuation of an activation electric signal; meanwhile, the ability of each feature to receive the activation electric signal is positively correlated with the self memory value and the similarity degree of the activation source and the feature.

7. An intelligent implementation method for a general machine is characterized by comprising the following steps:

the machine comprises a multi-resolution feature extraction module and a joint activation algorithm module for input information, and the modules can be realized by adopting hardware.

8. The method of claim 7, wherein the associating comprises:

the associative activation function of the machine is realized by the associative activation algorithm module through modifying the activation value of the feature in the memory.

9. The method of claim 2, wherein the method for enhancing the search of memory for data comprises:

the machine extracts the common relation network in memory to form a single common memory base which can improve the searching efficiency.

10. An organization of memory information, comprising:

the machine stores the memory, there are three kinds of data, the first kind is the information characteristic of the external input; the second type is internal self information; the third category is data of machine needs and states in which the needs are, emotions and states in which the emotions are.

11. The method of claim 10, further comprising:

the machine stores these three types of data in chronological order, and when stored, the initial memory value to which the data is assigned is positively correlated with the activation value of the data at the time of storage.

12. A method of training a multi-layer neural network, comprising:

the machine first extracts information features of the input information at multiple resolutions and then trains the neural network using the features at partial resolutions.

13. The method of claim 12, further comprising:

the machine can train the multi-layer neural network separately according to the grouping of the resolution aiming at the input information characteristics under different resolutions, and then the output of the multi-layer neural network is weighted and averaged to be used as the total output.

14. A memory recall and store method, comprising:

the machine converts the information streams of the linguistic input into non-linguistic information streams and stores these information streams as input information streams in such a way that the input information streams are stored.

Images