CN112016664A - Method for realizing humanoid universal artificial intelligence machine - Google Patents

Abstract

The implementation proposed by the invention is similar to a human universal artificial intelligence implementation method. The method comprises two parts, namely how to realize the general artificial intelligence and how to enable the general artificial intelligence to better interact with human beings. The invention provides a method for establishing the cognitive common knowledge of a machine on the world by adopting multi-resolution feature extraction, adopting near storage, adopting associative activation, adopting segmented simulation and adopting an optimal path search method and continuously and iteratively using the method. Meanwhile, the invention also provides a method for enabling a machine to better understand the human knowledge and better reflect the human perception (including emotion) so as to better interact with the human. Through the method provided by the invention, the machine can gradually obtain simple to complex responses to the input information, and has similar motivation and emotional expression 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 humanoid universal artificial intelligence machine

Technical Field

The present invention relates to the field of artificial intelligence, and more particularly to how to build generic artificial intelligence that resembles human intelligence, skills and emotions.

Background

Current artificial intelligence is typically designed for a specific task, and there is no general artificial intelligence that can accomplish multiple uncertain tasks. The biggest obstacle to implementing the general artificial intelligence is how to establish a cognitive network similar to common human knowledge among various complex things. Only if the machine has common sense like a human, the machine may produce mental activities like a human. The result of the current deep learning is an exquisite characteristic mapping method which is greatly different from the learning process of human beings, so the achievement of the deep learning is difficult to generalize and use. At present, knowledge engineering, expert systems or knowledge maps all adopt a coding mode to organize human knowledge by a method which can be identified by a computer. However, these systems make it difficult for machines to learn and generalize autonomously, so that in the face of differentiated scenarios, machines cannot autonomously generate new strategies and methods. So these systems have so far been applicable only to a certain area and range and have not been able to generate human-like intelligence.

And general artificial intelligence similar to human intelligence, skills and feelings is established, so that the machine can better interact with human, for example, the robot and the human cooperate to carry out production and research and development, and social activities such as endowment, nursing, education and the like are organized together, which brings huge economic and social benefits to the human society.

The application of the invention mainly discloses a basic method for establishing the general artificial intelligence, in particular to a method for establishing the general artificial intelligence similar to human in all aspects, including thinking, actions, emotions and the like, so that the artificial intelligence can be better integrated into the social life of the human.

Disclosure of Invention

The present application contains two broad aspects. Firstly, how to establish the general artificial intelligence and secondly, how to make the general artificial intelligence similar to human beings so as to be capable of better interacting with the human beings.

The first aspect of the present invention application: how to establish general artificial intelligence, the main method comprises the following steps:

performing feature extraction on sensor data, wherein the content comprises the following steps:

the machine needs to perform multi-resolution extraction of the sensor data. After the sensor data is input into the machine processing unit, the machine selects the emphasis recognition area and the recognition resolution according to the expected target generated by the last thinking activity to extract the characteristic information in the emphasis recognition area and the recognition resolution.

There are many specific algorithms for extracting feature information in the industry, such as deep convolutional network. The problems that are not solved are "which data can be used as characteristic information" and "how to organize and utilize the characteristic information". The invention provides a local area common similarity method which is used as a basic method for selecting characteristic information by a machine. The invention provides methods of adjacent storage, associative activation, segmented simulation, optimal path search, iterative decision and the like as methods for organizing and utilizing the characteristic information.

In the present application, we propose "local consensus similarity" as a method of selecting the underlying feature information. The method comprises the following aspects:

first, we consider the evolution process of an organism to be developed in a direction of high efficiency in the use of computational power. Only in this way, under the condition that the exploration environment is more and more complex, the complexity of the algorithm is improved to process the complex problem, and the energy consumption is saved to the maximum extent, so that the survival probability of the organism is increased. If the corresponding reusable extraction algorithm can be formed for the local features which are widely existed, the energy consumption can be saved to the maximum extent obviously under the condition of keeping the feature extraction capability. This is a concrete embodiment of the direction of evolution. Since these algorithms can be multiplexed to the maximum extent possible, the energy efficiency ratio of the calculations can be improved.

The present application therefore proposes that the machine needs to extract the basic features, which are common data features widely existing in our world, such as the points, lines, textures, colors, etc. of the basis, and further combinations thereof include parallels, intersections, vertices, angles, basic shapes, edges, curvatures, hues, sizes, etc. The machine uses any algorithm to extract these basic feature information during the early stages of learning. Similarly, the method is popularized to other data, such as voice, smell and touch, by searching local common features among the data to serve as basic feature information, establishing corresponding basic feature information extraction algorithms, and widely multiplexing the algorithms.

After corresponding algorithms for extracting the basic feature information are provided, the machine can multiplex the algorithms, and the algorithms are continuously utilized to extract the corresponding basic feature information in our world. In order to make full use of the basic feature information, the machine can parameterize the basic feature information and use the basic feature information as a basic model for external information identification. The machine extracts basic characteristic information contained in the input data in a parameterized model comparison mode. The basic characteristic information further establishes a connection relation with a more complex concept through a memory and forgetting mechanism.

Second, for a specific concept, the machine needs to employ multiple resolutions to improve the generalization capability of the machine. We consider that the similarity between things must be established at a certain resolution. It makes no sense to speak about similarity apart from resolution. In the present application, we propose to build multi-resolution features under the same concept. This multi-resolution feature is a tower-like structure. Located overhead, are those fundamental features that are most prevalent in the same class of concepts. They are typically an integral feature of things, such as outline, texture, color, etc., which is a low resolution feature. While at the bottom of the column are typically various specific features of each specific thing, they are typically less repetitive among the same things.

In the present application, a memory and forgetting mechanism is used to build such a tower structure of basic characteristic information. In this tower structure, the basic feature combinations that can be repeated are memorized, and they can be static or dynamic, and can include all sensor information, such as image, language, taste, touch, temperature, etc. The information forms a local network with a relationship with each other through the relationship network and the association activation method provided by the invention. These local tightly connected networks are concepts. The extent to which the concepts are activated reflects the extent to which they are identified. If the activated degree of the concept network reaches the preset standard and is highlighted in the relationship network, the concept is considered to be successfully identified, and the information for initiating the activation belongs to the concept.

In the concept, all basic features have their own memory values. The memory value of the feature generally existing under the same concept is increased due to repeated reproduction, so that the feature has larger weight. And the memory value of the features with low repeatability is low, and the features are gradually forgotten from the related concepts. Therefore, through a memory and forgetting mechanism, the machine gradually establishes a tower model of characteristic information under a certain concept. The tower models are formed by combining basic characteristics with different memory values. In general, low resolution features, which repeat a high number of times among the same species, have high memory values and are usually located at the top of the tower. While those with less reproducible high resolution features are typically located in the column. And those that are incidental are typically located low on the tower. In the information recognition process, under the drive of associative activation, since a high memory value is more easily activated, a feature of low resolution, which usually possesses a high memory value, is usually more easily activated, similarly to the way in which a human recognition thing is first recognized from an overall feature. These preferentially activated high memory value features are typically common features among the same class. The activation memories they represent form a model of the beginning of the recognition of the input information by the machine. The machine uses these models to segment the input information features to produce the expected data intervals for further identification and the resolution for expected usage for further identification. With the deep recognition, new information continuously activates new memory information or adjusts the activation values of the memory information, the memory information with high activation values gradually transfers to the memory which is more matched with the input information, and finally, a specific concept is highlighted in the memory to reach the confidence coefficient of the machine recognition, so that the machine successfully recognizes the input information.

The whole process of identifying the input information by the machine is represented by that the machine determines the interval needing to be further identified and the resolution needing to be adopted by comparing the model with the input by taking the activated concepts as the model and utilizing a decision system, identifies more input information by a repeated iteration mode, and gradually reduces the 'highlight' range of the activation information, namely the attention mechanism of identifying the input information by the machine. The above recognition method may have a path dependency problem, so that the machine recognition may have a "first-in-first" phenomenon, which is similar to the human recognition process.

Another aspect of identification is that low resolution whole features, due to frequent use, often have connections with many other features. When the low-resolution global features are activated, a large number of other features can be activated through 'chain activation' proposed by the invention application, the transmission weight of the activation values is dispersed due to the large number of activation paths, and the activated features are large in number, so that the activation value of no specific node in the activated information is significantly higher than that of other concept nodes (nodes without highlights), and the machine needs to further use the features with higher resolution for secondary identification. In contrast, those high-resolution features which are less frequently used, once activated, do not have the advantage that the activation paths are numerous and the transmission weight of the activation values is dispersed because of the fewer connections between the high-resolution features and other concepts, so that the activation values of the high-resolution features are directly transmitted to the concepts connected with the high-resolution features, and the machine can quickly identify the concepts related to the high-resolution features (easily highlighted). This is similar to how humans quickly identify particular concepts by particular features.

The overall concept includes characteristic information that is a multi-resolution tower structure that is built according to the height of the memory value. Those features with high memory values are located at the top of the tower, while information with low memory values is located at the bottom of the tower. In the concept, those information that can be repeatedly presented, driven by the memory and forget mechanism, are usually high in memory value. For example, language information of a concept is widely existed in the life and is related to the concept, and the language information becomes the information characteristic with the highest memory value in a concept due to repeated reproduction, thereby becoming a common entry of the concept.

In the application, the characteristic extraction of the input data by a machine is carried out by adopting a layer-by-layer conceptual model comparison method. The concrete mode is as follows: the machine performs the basic feature extraction at the desired resolution for a particular data region, based on the results from the last mental activity. If the data area of interest and the expected object were not generated during the last mental activity, the area and resolution are selected for identification according to a preset program or randomly, and it is even possible to ignore the input information completely.

The method is a modeled data identification method, because the basic characteristics are parameterized and modeled in the early stage of learning of the machine. In subsequent use, basic characteristic information combinations related to concepts are further established for the parameterized model combinations through a memory and forgetting mechanism. These widely existing basic feature combinations are increasingly in tight connection with each other due to frequent use. When some of them are activated, other basic features may be activated by passing activation values between them. The combination of these basic features gradually behaves like an overall basic feature, so the combination of these basic features also gradually becomes a new basic feature. And in these combinations, the weight of each underlying feature is positively correlated with its corresponding memory value. So when these combinations are activated, they may exist in a variety of different degrees of activation. For example, between two generally dissimilar concepts, it is possible to pass activation values through certain low resolution features that bring both concepts within the scope of thinking, which is the origin of some humor, witty, metaphors or after-dialects.

Similarly, in the present application, the machine may extract basic features not only for image data, but also for information input by voice information and other sensor information (such as touch, temperature, gravity, taste, smell, etc.), and put the information into related concepts through a memory and forgetting mechanism. This information also requires the creation of a similar combination of multi-resolution tower features that are also identified using associative activation. The information can also be combined into more complex characteristic combinations through a memory and forgetting mechanism, and the characteristic combinations are also a multi-resolution basic characteristic information combination and comprise tower-shaped structures of basic characteristics with different memory values.

It should be noted that, in the present application, the features of the extracted input data include not only static features but also dynamic features. Dynamic basic features here refer to basic motion patterns such as wobble, circular, linear, curved, wavy, etc. similar basic dynamic features that are widely present in our world. Through a combination of these features, and optimization of the memory and forgetting mechanisms, the machine can further build more complex dynamic concepts such as dance, running, wandering, and fantasy concepts. These concepts are also tower-like structures that are a combination of basic dynamic features with different memory values. Typically those more ubiquitous underlying dynamic features have higher memory values and they are typically low resolution features. These features have a high probability of being activated due to their wide presence, but their resolution for a particular concept is reduced due to the wide presence in various concepts. The recognition process of a machine is usually based on low-resolution global features, and recognition is carried out by increasing recognition accuracy layer by layer for multiple times until the accuracy required by the machine for "gain" and "loss" estimation is met.

It should also be noted that the resolution of the dynamic features does not refer to the resolution for implementing the moving object itself, but to the temporal and spatial resolution for distinguishing the motion states. For example, for speech, the basic speech and speech rate part can be used as a static feature, and the change of audio frequency, tone and speech rate is a dynamic feature. The machine sliding samples the speech according to time windows of different lengths, which corresponds to different time resolutions. For another example, the motion mode means that the machine ignores the details of the composition of the moving object itself, and focuses on comparing the motion modes. For example, a person walks, slides, or runs to us, and on a rough level, we will not even notice the differences in these motion patterns, so we consider their motion patterns to be the same at this time. However, as we increase the spatial resolution, we find that the sliding person comes in a steady motion, 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 and slownesses, so we find that their motion patterns are not the same.

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 the window, and we find that the motion pattern is the same in all 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.

If we consider the memory as a three-dimensional space containing numerous fundamental feature nodes, then the relationship network is the context in this space. The relation network is a relation network between all information in the memory base and the memory values of the information, and one memory information is formed by association activation. Relational networks are "experiences" that machines obtain from a large number of everyday studies through memory and forgetting mechanisms. These experiences may be personal experiences that the machine itself has summarized through a memory and forgetting mechanism, or may be other people's experiences (knowledge) obtained through learning. The relationship network represents the cognition of the machine to the world and is the common sense established by the machine to the outside world. The relationship network is in a central location in the machine intelligence system.

The highlighting means that after the association activation is completed, if one or more basic feature combinations (including basic feature combinations of various information such as graphics, voice, text, taste, touch and the like) are activated once or more times, the activation value is far higher than the noise floor of the activation value and higher than the activation values of other information in the relationship network, and the activation value is highlighted. The machine takes the concept of containing this "highlight" as the recognition result. And uses it to combine and split the input feature information. The machine takes the high memory value information contained in the concept as a model, and uses a model comparison method to compare input feature combinations to be used as a basis for judging whether further identification is needed or whether a specific interval needs to be further identified or whether a specific resolution needs to be adopted for further identification.

There may be different methods of calculating the activation value noise floor when calculating the "saliency" in the relational network. For example, the machine may use the activation values of a large number of background feature map nodes in the scene as the activation value noise floor. The machine may also use the average of the activation values of the nodes that are currently activated as the noise floor. The machine can also use its own preset number as the noise floor of the activation value. The specific calculation method needs to be preferred in practice. These calculation methods are only related to basic mathematical statistics and are well known to practitioners in the art. These embodiments do not affect the framework claims of the method and steps of the present application.

When the highlighted nodes are absent in the activation information, the machine makes a decision according to the nodes with relatively high activation values, and the decision is made by means of 'intuition'. The "intuition" itself is not mysterious, it is based on a large number of information connections, but the lack of a prominent connection in the active information context.

The method is a method for extracting characteristics of sensor input data, which is provided by the invention.

In the application of the present invention, the method for storing the obtained feature information includes:

the machine firstly adjusts the position, angle and size of the extracted basic features according to the position, angle and size with the highest similarity with the original data by scaling and rotating, and places the basic features and the original data in an overlapping mode, so that the relative positions of the basic features in time and space can be reserved. The machine may store these base features, or may store these base features and the raw data in an overlapping manner. In the application of the invention, the extracted features are mainly used, the corresponding original data can be used as backup data, the backup data can be called again when needed, and the features are extracted again according to the same method and needs. Therefore, the two storage modes have no essential difference on the realization of the universal artificial intelligence of the machine.

We propose a way of information storage: those "input temporally adjacent relationships" are expressed with "storage locations spatially adjacent". The adjacency of information in the storage space may be the adjacency in physical location: that is, time-adjacent information is stored on adjacent memory cells. The information may also be logically adjacent in storage space: it is stored with logical locations contiguous and a specific physical storage unit location is represented by a mapping table between logical locations and physical locations. Another method is that each stored message has its own stored time coordinate, and the machine determines the neighboring messages by searching the neighboring time coordinates. There are of course other storage means, but they must be able to express temporally adjacent information.

The machine stores the information by adopting a memory screening mechanism: event-driven mechanisms and temporary repository mechanisms. After the machine extracts the basic features, the time and space information of the input information is reserved, and memory required to be stored is formed. These memories can reproduce, by calling, part of the external information at the time of occurrence, so that the information is called mirror memories. Since they are mirror images of the machine's memory of some of the information that occurs in the outer space.

In the mirror memory, every time an event occurs, the machine takes a snapshot of the mirror memory and saves the snapshot. The occurrence event refers to the change of the current input information and the last input information exceeding a preset threshold value through similarity comparison. This is referred to as an event occurring. It should be noted that the occurrence event not only refers to external information, but also refers to internal information of the machine, such as monitoring information of the machine itself, and a change in the demand information of the machine over a preset value also occurs, which is that the machine needs to update the memory again. The updated content includes basic features in the mirror memory (including relevant information such as external information, machine state, demand and emotion) and their memory values.

The initial memory value of the information being stored is positively correlated, but not necessarily linearly, with their corresponding activation value at the time of storage. A snapshot of the mirror memory 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 the memory occurs. In contrast, information in a memory frame may be forgotten over time.

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 bank, 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 bank for storage. The long-term memory bank can be a single database, and can also be a mark made on a stored data entry, and the mark reflects the memory and forgetting curve adopted by the memory entry. Those slowly changing memory and forgetting curves represent long-term memory library data.

In the application of the invention, the size of the capacity of the temporary memory library is limited by adopting a stack with limited capacity, and a quick memory and quick forgetting mode is adopted in the temporary memory library to screen materials to be put into a long-term memory library. Machines, when faced with large amounts of input information, those things, scenarios and processes that have been learned about, or those that are far from the point of interest, lack the motivation for the machine to analyze them in depth, so the machine may not recognize these data, or the activation values assigned to them are low. When the machine stores the 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.

The method is the method for storing the input data provided by the invention.

A method for a machine to utilize stored memory data, the method comprising:

when the machine searches for related experiences in memory, the adopted method is a associative activation method which comprises a near activation principle, a similar activation principle and a strong memory activation principle. Wherein "activation proximity" means that a particular message in memory is activated and only information in its vicinity is activated. "similar activation" refers to a specific feature in memory, and when receiving activation signals from other features, the receiving ability and the similarity between the features are positively correlated. "strong memory activation" means that the higher the memory value, the stronger the ability to receive activation signals from other features. Based on these 3 principles, the machine can achieve associative abilities similar to the human brain.

The chain activation refers to a process that a machine activates a plurality of memory information based on an input basic characteristic on the basis of a 'proximity activation' principle, a 'similar activation' principle and a 'strong memory activation' principle. When the basic features are input, the machine finds similar basic features through a similar activation principle and assigns activation values to the basic features according to motivation. At the same time, the memory of the adjacent part is activated according to the 'adjacent activation'. The activation of the adjacent memory is performed according to the principle of 'strong memory activation'. One possible implementation is that the activation value transfer coefficient is a positive correlation function of the memory values across the transfer line.

And after all the nodes receive the transmitted activation values and accumulate the initial activation values of the nodes, the total activation value is greater than the preset activation threshold value of the node, and then the node is activated. The 'close activation' principle, 'similar activation' principle and 'strong memory activation' principle are also adopted to carry out chain activation. This activation process is chained until no new activation occurs and the entire activation value transfer process stops, which is referred to as a chained activation process. In order to avoid the repeated mutual activation between the two basic features, the machine needs to limit the reverse activation value transmission to occur immediately after one activation value transmission between the two basic features.

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 activation value of the information input earlier, in the application of the invention, the activation value in the chain activation is decreased with 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 cause interference between information. If the activation value is not faded, the subsequent information input will be strongly interfered by the previous information. 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 has the advantage of both maintaining the contextual relevance of the information and balancing the weight of the contextual information. And the important information usually obtains the activation values assigned by a plurality of channels, so that the important information becomes a node with a high activation value. The activation values of the key information exist for a long time, and the key information exists in the activated information for a longer time and participates in the information identification and machine decision process for a longer time.

The activation value fade parameter needs to be preferred in practice. But this presents the problem of maintaining the active state of a message. When the machine is faced with a large amount of activated information, the information with high activation value is the focus of the machine. If the machine cannot complete information understanding at a later time, cannot find a response scheme that satisfies the machine evaluation system, and over time, the activation values fade, causing the machine to lose attention to the activated information and even forget what to do. The machine then needs to refresh the activation values for these points of interest again. One brushing method is as follows: the attention points are converted into virtual output, the virtual output is used as information input, and an information input process is carried out once to emphasize the attention points. This is why humans like to pronounce or to have a default in mind when thinking, sometimes, when not understanding or finding no idea. This virtual input, like the real input process, also uses the associative activation process to search for memory and update the memory value. Therefore, the method can lead the machine to intentionally increase the activation value of certain specific information, and can also lead certain specific information to repeatedly appear by using the method to increase the memory value of the specific 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 thinking process to process the new information, resulting in a loss of attention. Therefore, from an energy saving perspective, machines tend to be thinking-free, avoiding waste. At this point, the machine may actively send out buffered help words such as "take … or …" or otherwise send out information that it is thinking and not disturbing itself. There is also a possibility that the machine may be given a limited amount of thought time or may be overloaded with information and need to complete the information response as soon as possible, and the machine may also use output to input. In this way, the machine emphasizes the useful information, suppressing the interference information (the interference information is not entered again, its activation values fade out over time). 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, needs to refresh the attention information, or tells others to think by themselves or emphasize important points to eliminate the interference information according to a built-in program, or experience of the machine or a mixture of the two.

Since human communication is most frequently speech and text, in a concept local network, various features obtain activation values from various branches of a relational network, and all of them can transmit the activation values to the speech or text, so that the node (focus point) with the highest activation value is the speech or 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 text and input the text again.

The searching method using chain activation utilizes the implicit connection relation among languages, characters, 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 chain activation is a search method, which is not a necessary step in the present application per se, and may be replaced by other search methods that achieve similar purposes. When chain 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 according to the memory and forgetting mechanism in the memory bank to which the memory belongs.

The appearance of relationship context in relationship networks is that because of the memory and forgetting mechanism, those relationships that cannot be repeatedly activated are forgotten, and those relationships that can be repeatedly activated are strengthened. The underlying features that are connected by coarse relationship veins 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 concept is a little looser than the basic concept, namely the static extension concept and the dynamic concept extension concept (including the concept representing the relationship and the process characteristic graph), and the concept is a memory. In relational networks, static profiles (or concepts) are usually small parts that are widely used, dynamic profiles (including concepts representing relations) are connecting parts that are widely used, and process profiles are large frames that are organized in a certain temporal and spatial order of small parts (static objects) and connecting parts (dynamic features). The process features a large framework that we can use for reference. Dynamic profiles (including concepts representing relationships between things) are tools that can embody empirical generalization, while static profiles (or concepts) are objects that are replaced in the generalization.

It is noted that associative activation is a chained activation process that occurs simultaneously on multiple resolution bases. The machine can extract multiple resolution basic features of external input information at one time, and can also adopt a mode of multiple times of extraction (after each time of extraction, information processing is carried out, and then the section of next extraction and the used preset resolution are determined). Similar to the human brain, from an energy saving perspective, machines default to preferentially extracting global features of things, which are typically low resolution features (unless in the previous thinking process, the need to extract high resolution local features arises). For example, for an input object image, the low-resolution features are mainly basic features such as overall contours and textures of the object. These low resolution features, driven by associative activation, have the potential to activate many concepts, such as the concept where the most basic concept might be "object". It is also possible to activate the object with further increased resolution, possibly with information about the size, texture, weight, stiffness and whether it can be moved. And possible movement modes of the object can be identified through the extracted basic dynamic characteristics, and concepts related to the movement modes are further activated through the movement modes, so that the object is further identified. The activation process is usually an input tower of information (from low resolution to high resolution) that activates related concepts in a relational network (each concept itself is also tower of information and different concepts may share their own content). Different concepts obtain activation values from various paths through chain activation. Those concepts with high activation values preferably meet a machine preset confidence criterion, at which point the machine considers the input information to be recognized.

The phenomenon of first-come-to-first exists in association activation. For example, two machines having the same relationship network confront the same feature map and the same initial activation value, wherein one of the machines suddenly processes an input message regarding 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.

This phenomenon also occurs in the processing of input information. For example, when a feature is identified in the input information, similar features may be more easily activated due to similarity, and obtain a higher activation value. This will make it easier for us to identify similar features in the input information. This phenomenon makes it easy to identify global features in the input information that are made up of similar features. Such as a pattern of "dots" of the same color or shape, such as a line of "dots". Or a large pattern consisting of a few discrete small patterns. The association between these similar features makes it easier for the machine to identify the overall features made up of these similar patterns.

In addition, because the machine adopts a mode of pattern recognition, the input information is recognized through iteration. If the machine activates a certain conceptual model in other ways in the identification process, the machine can preferentially adopt the activated conceptual model as a basis when identifying input information, and establish the identification process by comparing information in the model with external information. The information of this model may be used by the machine to segment and classify the input information and thereby compare the similarity between the two. This is also a first-come-to-first subjective recognition bias.

In addition, because the machine adopts a mode of pattern recognition, the machine is easy to be preferentially activated by presetting models with high memory values of overall characteristics, so that the models are preferentially recognized in the iterative recognition process of the machine, such as human faces, or emotion recognition required in the life of some machines, and dangerous objects or dangerous situations.

These ways of recognizing the extrinsic information by the machine are similar to the way that humans are cognizant, and the root of this similarity in recognition patterns may be from reasons similar to the iterative recognition process by both the machine and the human.

The machine establishes a relation between external information and self requirements and emotions, and the content comprises:

human beings can achieve a goal, bring benefits (such as earning rewards) and avoid losses (such as survival needs), which is a gift brought to the human beings by evolution and is a motive force for the human beings to continuously develop. Similar instinctive motivation can be given to the machine, so that the machine can build self-development power.

In order to achieve the aim, the method provided by the invention comprises the following steps: in the memory frame, the machine stores not only external input information but also other types of information. Such as internal state data of the machine, needs and motivations of the machine, mood of the machine, and other types of data. The machine processes the information by the same processing method (including basic feature extraction) as that of the externally input information, by the same storage method (such as the method of storing the information input at the same time according to the adjacent storage principle, by the same memory and forgetting mechanism, by the same method of assigning the activation value and generating the initial memory value), and by the same information utilization method (such as the associative activation method).

The initial activation value assigned by the machine to the input information is also propagated to the demand and emotion data of the machine through the relationship network, resulting in activated demand and emotion data of the machine. The machine can use the data to further activate relevant experience and general knowledge, and adopt a method of seeking profit and avoiding harm to select the response of the machine to the input information. The demands of the machine and the mood data are therefore a very important type of "personification" data.

In the present application, the machine may be given various motivations, which are sources of power to drive the machine in response to input information. They are the control mechanisms behind machine behavior. We can give the machine needs and emotions as examples to explain how the machine decides its behavior according to these incentives. The motivation that the machine may be endowed with includes not only needs and emotions, but may also include other types of motivation. The difference and magnitude of these motivation types do not affect the claims of the present application. Since all types of motivational data are treated the same in the present application.

In the methods presented herein, machines employ symbols to represent the various underlying requirements that various humans impose on the machine. For example, the safety requirements of the machine itself, pursuit of pleasure, hope of obtaining human approval, hope of obtaining human respect, having congruence, agreeing with social ethical standards, further, for example, self-reward (sense of achievement) brought by the machine self-goal realization (goal achievement), unknown curiosity of the machine for exploration, and the like. These requirements can be represented by a symbol, and the symbol can be assigned a value to represent the state. 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.

In the method presented in this application, the machine employs symbols to represent the underlying emotions that various humans impart to the machine. The machine's emotions can be varied and each type of emotion can be represented by a symbol that can be assigned a value by the machine to indicate the state in which it is. The difference and magnitude of these mood types do not affect the claims of the present patent application. Since all emotions are treated in the same way in the present application.

In the method provided by the application, the relation between the emotion of the machine and the demand state of the machine can be related through a preset program. The parameters of the preset programs can be self-adjusted by the machine in the learning process of the machine according to the principle of 'tendency towards interest and avoidance from harm'.

In the method provided by the application, the emotional state of the machine and the explicit expression mode of the machine emotion can be connected through a preset program. The parameters of the preset programs can be self-adjusted by the machine in the learning process of the machine according to the principle of 'tendency towards interest and avoidance from harm'.

It is to be noted that the situation that the machine needs are fulfilled can be mapped to the mood of the machine and further mapped to the outer appearance of the mood of the machine. But the emotion of the machine itself can also be a requirement of the machine. The two may interact. But the underlying basis for determining the motivation of the machine is the satisfaction of the underlying requirements, whether preset. These underlying requirements, which may have established connections with a wide variety of specific needs during life, also generate revenue and loss connection values with various events. The machine uses these connection values to make decisions. The root cause of these decisions remains the underlying demand motivation for the machine. The underlying motivation for the machine is that humans give the machine a direction of development, which is the part that requires absolute care in design, especially in situations where the machine may possess knowledge and the ability to exercise that knowledge is likely to go far beyond humans.

In the world where people are located, relationships among things are complicated, and it is very difficult to establish various relationships among things manually, and it is also difficult to quantify and flexibly apply the relationships. In the invention application, the machine extracts the relationship between things by memorizing/forgetting and active learning.

When external data or internal data is input, the machine generates responses that are fed back externally and change internal states (e.g., low battery). The machine needs to establish the relationship between behavior and reward through three aspects and conduct autonomous activities through demand and motivation.

In a first aspect, an externalized display of the needs, motivations, emotions and emotions of the machine is established by a pre-programmed program. The preset program is mainly used for establishing the bottom layer requirements, motivations and emotions of the machine, such as the requirements of 'complying with the robot convention', 'complying with the human law', 'protecting the self', 'obtaining the recognition of the owner' and the like.

In the second aspect, the connection relation of the underlying requirements and other information is established through training. In the present application, for example, for "safety" requirements, the machine may be informed during training by preset symbols (such as language, actions or eye-gaze) that those environments are safe, that those environments are dangerous, or may be further informed of different levels. As well as 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 mutual features of the environment or process that are dangerous to the machine through training, memory and forgetting, and the danger-associated built-in requirement symbols. Then the next time the machine processes the input message, given the same initial activation value as the input message, the activation value of some features, due to its close relationship to 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 a memory frame to connect the risk factors and the risks through language, and the safety value and the risk value are used for telling the machine how to identify the safety factors and the risk factors so as to learn to protect the machine if the machine is protected. 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 some time after the event occurs, it may be sufficient to use the reward and penalty and tell it about the cause. As another example, during human growth, the first connection with interest may be "water", "milk", "food", etc., and later through experience summarization and learning, the human establishes a connection between things such as "test scores", "bank notes", etc., and profits. Later we may also establish a connection between revenue and something without entities such as "love", "time" and "life". These are the training processes that establish the link between the underlying motivation, demand and specific information.

In the third aspect, it is also possible to preset experience (such as directly modifying the brain nerve connections thereof, or giving virtual memory to the machine) to achieve the purpose of giving experience and common sense to the machine.

Similarly, the machine may associate the own body state evaluation value and the demand with the emotion and the external input information in order for the machine to understand the own body state evaluation value and the association 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 the same situation is repeated multiple times, 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.

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 acquired learning, 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. The trainer can continuously tell the machine, its emotional performance, which are rewarded and which are punished during the life of the machine 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.

Further, since emotions and other memories are stored together, they are in the same memory. When the machine needs a certain result, it mimics the memory that brings this result. For example, if a certain type of behavior brings about a certain result that can be repeated, the machine will mimic the memories that contain such behavior, and of course the emotions in those memories, so that it will adjust its own emotions for some purpose. This is a way of exploiting emotions. The mood of the machine is not only passively displayed, but it is also a means by which the machine can utilize. The machine just adjusts the emotion expression parameters preset by the machine in a way of trending interest and avoiding harm by evaluating emotion connection information related to various gains and losses, so as to realize benefit maximization.

Therefore, in the present invention, the machine can incorporate the reward and penalty information into its own mind by simply putting the reward and penalty in memory along with all external and internal information, without having to build various "rules" to tell the machine how to recognize the environment, what to do and how to express the emotion, which is also a task that is practically impossible.

It is pointed out that the machine intelligence established by the method proposed by the present patent application, whose thinking and emotion are visually controllable to human, is fully understandable by the reproduction decision process, which are linked by associative activation. Therefore, the machine is intelligent, and because the thinking process is visible, the condition that the black box thinking brings danger to human beings can be avoided. This is also a feature of the general artificial intelligence implementation method proposed in the present application.

The machine establishes an understanding and response decision for external or internal information, which comprises:

the reconstruction of the information is activated.

Language plays an important role in machine intelligence. Language is a set of symbols that humans create for better communication experience. Each symbol represents a particular thing, process, and scene. When a language is entered, the associated memory represented by the language is activated. These memories may have both information about the language itself and memory related to the manner in which the language is used (e.g., speech emphasis or text emphasis to emphasize important parts such as an irretentive voice or a jeopardy tone, etc.). The activated information constitutes an activation information stream. To balance the context of the language and the current semantic recognition, the activation value of the activated information may decay over time. The parameters of the decline are related to the motivation and state of the machine (such as demand and demand state, mood and emotional state).

The chained activation of the languages enables context-dependent recognition of all input information. The input information here includes both the environmental information and the activated memory information. The mutual activation and assignment of the information embody the context association. This correlation is more extensive than statistically generated semantic library content. It does not relate to only speech, but also to all sensory inputs and associated memory. The machine can implement the connection of language to static and dynamic images, feelings, demands and emotions, as well as to related languages and memories. When such a connection is incorporated into the understanding of the language input by the machine and is responsive to the understanding of the language, and associated experience, it is apparent that the machine actually understands what the input language really is.

The language input constitutes an input information stream and the corresponding activation memory also constitutes an activation information stream. When a machine understands language, it needs to reconstruct this active information stream, constituting an imaginary process. This is because one input message may activate multiple pieces of associative memory. The machine needs to integrate these multiple pieces of associative memory to make decisions.

The integration of the information by the machine is carried out by adopting a method of simulating and modeling in a segmentation way. The specific method comprises the following steps: among the activated information, those with high activation values are generally overall frame information, and they are generally some low-resolution overall features. The machine uses these low resolution global features as a model framework for the information flow. Then adding more input information or activated memory information into the model framework layer by layer to form the understanding of the machine to the input language information.

This framework may produce a variety of outputs. For example, a typical output is "context information". The machine combines the relevant information memorized by the machine into a tower-shaped structure according to the height of the memorized value, then adds the input information into the tower-shaped structure by simulating the similar information organization experience in the memory to form an integral 'environment information', thereby outlining an environment relevant to the language in the 'mind'. The machine then uses this "context information" to understand the information and make decisions.

It is noted that the viewing angles of the things in this created "context information" and the actual specific things in machine memory may not be the same. This is because the machine needs to process the angle and size of the actual things in memory according to the understanding of the machine after the machine performs different angle transformations such as rotation and scaling on the shape, so that the things can meet the requirements of the whole large frame.

The requirement of the whole large framework comes from the high memory value feature combination reserved after the machine carries out memory and forgetting mechanism optimization on similar scenes. These high memory value features, which are typically common features in this class of scenarios, establish a tight connection with other information of this class of scenario-related concepts through memory and forgetting optimizations. The overall recognition of information by a machine is essentially to use the basic features of the input to activate the relevant model, and then recognize the information by comparing the model with the activated model on a continuous basis. These identified model groups constitute the overall model framework as a whole. The machine adds more input information to this model framework, forming an imaginary process.

The environment information is a static reconstruction, and the environment information not only includes the map environment, but also includes all things in the environment.

Another typical output is a "dynamic manifold". Dynamic manifold is the process of building dynamic state of input information by a value machine. This process is usually built on top of "context information" (there may also be no context information). The process is that the machine forms an integral dynamic manifold by combining related dynamic characteristics into a tower structure according to the height of a memorized value and then adding input dynamic information into the tower structure according to the mode that the activation value in the memory is the highest, thereby outlining a dynamic process related to language in the mind. The machine then utilizes this "dynamic manifold" to understand information and make decisions. In dynamic manifolds, the objects that are concrete to make the motion may themselves be abstract, and may be replaced with points, lines, planes, volumes, or blurred low resolution contours. The purpose of the machine-built dynamic manifold is to analyze the dynamic process, so there is no need for clear initiation or characterization of the object of operation.

It is noted that the viewing angles of the objects in this created "dynamic manifold" and the actual specific objects in the machine memory may also be different. The machine processes the angle and the size of the actual dynamic track in memory according to the self understanding of the machine after the machine transforms different angles such as rotation and scaling of the motion track, so that the dynamic characteristics can meet the requirements of the whole large frame and the compromise is made.

The requirement of the whole large frame comes from the dynamic feature combination with high memory value reserved after the machine optimizes the memory and forgetting mechanism of similar dynamic composition. These high memory dynamic features, which are typically common features in this type of dynamic scenario, establish a tight connection with other information about concepts associated with this type of dynamic scenario through memory and forgetting optimizations. The overall recognition of dynamic information by a machine is essentially to use the basic dynamic features of the input to activate the relevant dynamic model, and then recognize the information by comparing the continuous and activated dynamic models.

It should be noted that the static model "environment information" and the dynamic model "dynamic manifold" are combined and performed synchronously, and the adjustment processes of the two are mutually supported. Here, two processes are separated for convenience of description. Such as: we hear "a dog is in the garden" we can build a low resolution image of a dog that we are familiar with. This figure is a high memory figure linked to the concept of "dog" in our memory, which is often a common feature of dogs we see, because of the high memory values obtained by repeated activation. These common features are a low resolution feature for resolving a particular dog. We can also build a low resolution image of the garden. This image creation process is also similar to the creation process of a dog image, which is a common feature of all garden-related scenes we see, because it gets high memory values due to being repeatedly activated. These common features are a low resolution feature for resolving specific gardens. If we hear "run to go" we may establish the notion that "dogs" run to go "in the garden", and because of lack of detail we may establish only the image of a low resolution dog with a higher activation value, the information of the garden and the animation of running to go. The dynamic feature of "run-to-run" may then pass the activation value back to the concept of "dog", further increasing the machine's confidence that the previous information "one dog is in the garden". Due to the relative memory of gardens and dogs. In this piece of language input, only those common information features that are widely present in "garden" and "dog" information are activated due to the lack of further detailed information. This is because these common features are activated multiple times to achieve high memory, and they are more easily activated according to the associative activation method. These activated messages, through reorganization, create a context (dog and garden) and dynamic manifold (run to run).

Although this viewing angle may not exactly match the garden and dog memories in all memories, the machine mimics its own general knowledge to create this scene. The basis of the intelligence of the machine is to build up common knowledge step by step through learning, memory and forgetting, and the common knowledge is embodied in the connection relationship in the relationship network.

These identified model groups constitute the overall model framework as a whole. The machine will continuously add higher resolution input information to the model framework, and the created imagination process will become more and more specific, and even finally locate a scene of a specific garden and a specific dog, and even locate information that the dog runs away in the garden at a certain time. Therefore, the information reconstruction process is a process with an ever-expanding activation range and an ever-detailed environment information and dynamic manifold. The machine need only determine a rough activation value criterion. Information above the activation value criteria is incorporated into the piecewise simulation to reconstruct the context information and dynamic manifold, and information below the activation value criteria is not involved in the process, and the machine can create one or more things, scenes and processes that may include context information and dynamic manifolds.

In the process of creating, if there is information that cannot be added, the machine has to determine whether the information is accurate, or adjust its model, or rebuild its model according to the specific information, which are part of the response of the machine to the input information, as described later.

Therefore, in the present application, the model is a conceptual combination formed by the basic feature combination with the highest activation value in the activated memory. The concept itself is an open local network, and the content contained in the concept itself is different under different activation value thresholds. The integration mode is to integrate the parts which can be overlapped in the multi-segment memory to form a frame. And the basic approach to integration is piecewise emulation. For example, for the reconstruction of an environment, activated scenes in multiple segments of memory are integrated, and these scenes may be the same scene that is activated or similar scenes that are activated. .

Similarly, by the same method, the machine can also integrate the information of the relevant memory, demand state, emotional state and the like activated by the input non-language information with the current state information of the machine. The materials used by the machine to create the "context information" and "dynamic manifold" are not just languages, but also include other forms of input information. Therefore, the same integration method is adopted by both the language information and the non-language information to integrate the activated information together, so as to form the comprehension of the information.

Generation of machine prediction capability.

The nature of machine prediction is a statistical behavior. The prediction of the machine is to estimate various possibilities and corresponding probabilities of the development of the object or various possibilities and corresponding probabilities of the behavior of another person based on the past experience or the similar past experience.

When information is entered, the machine does not need to exhaustively predict all possible outcomes, which is also an unfulfilled task. The machine need only evaluate the experience that has been activated in connection with the input of information, events that have occurred, or the like that have occurred, which may be a gain and loss to the machine. The machine-activated memories include information associated with input information (associative activation), memories similar to the input information (similarity principle), memories before and after the information (activation), and memories with profound memory, such as memories with great profit and loss (strong memory principle). This is equivalent to the fact that the machine utilizes its own common knowledge to limit the search range for searching the optimal response path, thereby converting an open problem into a problem of searching the optimal path within a certain range. The problem of searching for the optimal path within a certain range is the problem that artificial intelligence can solve at present.

Within this limited range, the development of things can be presumed empirically. The specific method is to assign an initial activation value to the input information and then associate the past memory with the related input information through an associative activation process. Each development result may bring different income/loss and emotional states to the user, and the information exists in the relationship network and is activated together through the association activation process. The activation values obtained for these gain/loss and emotional state are the possible gain/loss and emotional state values.

After information related to the income/loss, emotional state and the like brought by each possibility to the machine and the motivation, the machine can adopt any current artificial intelligence prediction method, such as Bayesian estimation, Monte Carlo search, decision tree, machine inference methods based on rules and the like, to search the most favorable path from various possible development paths.

Because the purpose of the machine is 'interest and harm avoidance', the basic starting point of the response of the machine to the input information is to make the response of the machine according to the past experience, and the occurrence probability of the things generating 'benefit' is increased as much as possible, especially the situation that a very high benefit value can be obtained. But to reduce the probability of "loss" occurrences, especially in situations where large loss values can be incurred. Therefore, the machine combines own response according to experience under the motivation of balancing advantages and disadvantages to achieve the aim of 'driving toward interest and avoiding harm'.

The decision of the machine is a path planning method based on the prediction capability of the machine. The goal of the path is to minimize losses, if at all. With predictive capabilities, the machine transforms decision making and response, a fully open problem, into a string of relatively closed problems of how to increase or decrease the probability of a range of things occurring. Since the knowledge is established in the previous steps, the conditions related to each event (which is the cause of the cause and effect relationship) can be obtained through the relationship network when the event occurs. The causality relationships with strong association are strongly connected in the relationship network because of repeated occurrence. The relationship network can express causal relationships layer by layer.

The goal of each step of decision making is to make the development direction of the things "favor and avoid harm". This is a process of interaction with the outside world. The interaction itself is a method to promote the development direction of things to be beneficial and harmful according to past experience. The probability of the event with high profit value is continuously improved through the information and the behavior obtained by interaction, and the probability of the event with high loss value is continuously reduced. This is an iterative process. But the manner in which each step is processed is the same. And the machine increases the probability of the occurrence of the events leading to high profit values layer by layer on the basis of the causal chain. This is similar to the chain activation process, activating those events that lead to high revenue paths step by step, while carefully avoiding those events that may lead to high loss values.

Since the probability of causal connection between paths is expressed by a relationship network, the response planning problem of the whole machine becomes the problem of finding the optimal path in the causal chain network, which is the problem that the current machine intelligence has solved. For example, the machine can determine the prior probability of an event (e.g., an event that results in a high value of gain or loss) by searching the memory. The causal strength (a posteriori probability) between a certain condition and the event can then be determined by means of a relationship network. The connection strength between different conditions in the relationship network can reflect whether the different conditions are independent or not. The machine can predict the probability of the event by only selecting a plurality of relatively independent conditions through a naive Bayes algorithm. The machine may determine its own response based on the calculated probability. These responses may take various forms, such as: the probability of the occurrence of the event is improved, or the probability of the occurrence of the event is reduced, or the probability of the occurrence of the event is not influenced. Depending on the value of the return and loss to the machine. And increasing or decreasing the probability of the occurrence of the event, and further planning to increase or decrease the probability of the occurrence of the condition related to the occurrence probability of the event. This process is essentially an iterative probabilistic path search problem.

For example, the following steps are carried out: if the response of the machine is further to determine possible gains and losses. First, the machine takes as a priori probabilities, in memory, probabilities of various possible outcomes, under conditions similar to the current situation. Various revenue values and loss value occurrence probabilities are then calculated based on the conditions associated with each outcome and the posterior probabilities between the outcomes. The machine then generates the next target and further determines the probability of each condition occurring. For example, the response of the machine at this time may be (a) searching for and counting the posterior probability between each condition and the corresponding occurrence of the profit and loss values. And then used to update the overall gain and loss assessment. This can be done by searching for the connection strength between the relationship networks. (b) Further updating the probability of a certain condition occurring at present. Such as by directly querying the information source as to the probability that a condition has occurred or is likely to occur, based on mimicking past experience. Or by obtaining in other ways the probability of whether a certain condition has occurred or is likely to occur. Depending on the behavioral mimic memory obtained during machine learning. (c) According to the principle of driving toward interest and avoiding harm, certain conditions closely related to income and loss are used for promoting the occurrence of the disease or avoiding the occurrence of the disease as a new target. Under the drive of a new target, the same evaluation process is adopted to respond. With such iterative responses, the ultimate goal remains to gain revenue and avoid losses.

Therefore, when information is input, the causal connection determined by the relationship network, through the principle of driving towards interest and avoiding harm, and through the relationship between the events and the gains and the losses established by the machines in the relationship network, the response of the machines which are seemingly completely open to the information input can be changed into a multi-level target. These goals are all served by increasing the probability of certain events occurring, or decreasing the probability of certain events occurring. Therefore, through the causal relationship of the relationship network, the machine can convert the interest and risk-averted target into a series of targets which are associated with each other in a specific situation. These goals constitute a realization path for the machine to maximize revenue and minimize losses.

In this process, the machine may respond by continuously searching for new information, or continuously passively obtaining new information, and updating the target path with the posterior probability between the new information and the result. The possible external feedback prediction of the machine after responding to the machine also comprises the activation of two types of motivational state memory. One is the need and emotional state of oneself in recurrent memory, which comes from various feelings and emotions about oneself in activated memory. One is the demand and emotional state of the person himself when viewing a similar scene from the observer angle, which comes from the activated memory observing the various feelings and emotions produced by the machine of others under similar scene. Therefore, when predicting the 'income' and 'loss', the machine simultaneously evaluates the 'income' and 'loss' brought by an event to the machine from the view angle of the machine and the view angle of other people.

The predictive capabilities of a machine include not only predicting the "gain" and "loss" that an event may bring. The method can also predict the possible response of the user or others driven by the 'income' and the 'loss', and the influence of the responses of others on the 'income' and the 'loss'. These are obtained by counting the motivational state values of the related needs and emotional states in the relationship network. Therefore, the evaluation results of the machine are dynamically changed with more input information. The decision and response process of the machine is a dynamic path planning process. It is jointly driven based on empirical response and probability calculations based on profit and loss.

By the method, the machine can decompose an abstract profit-and-harm-avoidance target into a large number of tasks for improving or reducing the probability of certain specific events layer by layer in a layer-by-layer iterative decomposition mode under a specific input condition. These tasks can be subdivided layer by layer into very specific target tasks, for example, up to the underlying drive capabilities of the machine. This process is the decision and response system of the machine.

Due to the complexity of the world, it is difficult for a machine to get exactly the same experience as the current situation. Therefore, the machine uses experience by selecting experience segments that can be used in real-world situations from a plurality of segments of experience, and using the experience segments as a basis to utilize input information and memory information through generalization. The process is a segmented simulation, and the essence of the process is a process of reorganization by using memory and input information, and is a creative process. In the segmentation simulation, the method for selecting the experience segment for the real situation by the machine is selected according to the activation value in the memory. Those have high similarity to the input situation, and have a memory of common elements with the input information, and the activation value is higher. Therefore, the machine selects the high memory value segments from a plurality of specific experiences, namely selects the experience segments which can be used in the real situation. The generalization principle of using the experiences is that elements with high activation values in memory and elements with high activation values in input information which belong to the same minimum concept can be mutually replaced on the basis of a dynamic process or a connection concept representing a relationship. The basis of generalization is usually a dynamic feature or a feature representing a relationship between objects. Since these features are unrelated to specific things, they are widely present in our lives, and exist between different specific things, so their memory values are also high, and they will often be selected as framework information. Therefore, it is usually the right approach to generalize, through the same minimal concept, what is activated by the same thing that is activated, by replacing the thing that is also higher in activation value with the thing that is higher in activation value, through such a framework bridge. The machine borrows the process frames, and forms a new process of shape and color by adding details in a generalization way.

The machine firstly finds the low-resolution high-memory values with generality as a framework process by increasing the threshold of the activation values participating in information reconstruction. These high memory values are widely connected and representative (so they can obtain high memory values). Then, the activation value threshold value participating in information reconstruction is gradually reduced, and more activated information and input information are filled in the frame. This process defines segment emulation from another perspective. The segmentation simulation is an iterative process, and each upper-layer link is expanded into a plurality of lower-layer links meeting the real conditions through the segmentation simulation. Then, in the process of section simulation, the same method is continuously adopted to expand each lower-layer link into a plurality of lower-layer links meeting the actual conditions again. This process is iterated until the machine is able to combine a response and use the connected revenue and loss values of the empirical segments to determine the overall possible revenue and loss values.

After the machine plans the response path, in order to further analyze possible results, the machine may further evaluate its own decision process by using a method of virtual output and virtual input. This process is to take the output of the plan as input information that a hypothesis has occurred. When organizing the assumed input information, the relevant information is still organized in a manner of establishing "environment information" and "dynamic manifold". This assumed input information is processed in the same manner as the actual input information. It further activates the relevant experience and goes through the analysis process again, thus excluding possible interfering information in the initial input. In this way, the machine performs a more in-depth analysis of the plan, and this additional number of analysis processes may be zero to many, depending on the size and probability of the machine evaluating potential gains and losses. The above-mentioned times of re-evaluation can be implemented by using preset programs, and the parameters of these preset programs can be adjusted by machine in learning according to the motivation of "driving toward and avoiding harm", according to the continuously updated values of interest and loss in experience and their probability changes that may occur under various conditions.

The machine executes a response process to external or internal information, and the content comprises:

the emulation capabilities are established. The mimic ability is the ability of a human to exist 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. Therefore, the machine needs to have the emulation built into the machine as a kind of underlying motive. The machine is willing to imitate the behaviors of other people (other machines), and the machine is continuously improved according to the evaluation of the machine or the external feedback information, so that the coordination and consistency abilities of various senses, limbs, languages and actions of the machine are continuously exercised, and the learning efficiency is improved. At different stages of machine learning, we can give the machine different strengths of motivation for simulation. For example, when the machine learns the language and the action output, the machine can be directly endowed with a strong imitation motivation, and at other stages, the machine can be endowed with a normal imitation motivation.

When the machine obtains external voice or action input, the voice or action can activate the relevant memory of the machine. These memories may be a similar pronunciation, or a basic action fragment. These memories further activate sensory, demand and emotional information, language or action memories associated with these memories. The machine, driven by the simulated motivation, will make similar speech output or motion output by adjusting the underlying driving parameters in the experience through the decision making system based on these activated memories. And the bottom layer driving means outputting bottom layer experience by voice or outputting bottom layer experience by action. They are muscle-driven commands corresponding to a particular voice or action, where the parameters are updated through acquired learning and continuously through feedback.

The machine establishes the preset capabilities. Humans may preset the machine with some of the most basic speech or motion (including expression and body language) capabilities. The optimization of the parameters can be realized by subsequent learning and training, the results of the parameters and behaviors are combined by memory, and are continuously adjusted by an emotion and demand system (influenced by self or external feedback), and finally, under the drive of a bottom layer motivation, the machine obtains the relationship between different parameters under the excitation of different external information through a memory and forgetting mechanism to form memory. These memories are all the knowledge and skills of a machine in the face of external information input. They include behavioral habits of language, motion, expression, limb movements, and the like.

Humans may also give the machine a preset conditioned reflex system. The role of these systems is that humans expect the machine to respond under certain input conditions. Such as evasive action of the machine in case of emergency, or specific output action of the machine under specific information input (for example, these conditioned reflex systems can achieve the purpose of self-checking the machine, or emergency shutdown, or adjusting the working state of the machine, etc.).

The machine establishes an execution process. After having the above various basic capabilities, the machine can perform the response specifically according to its own decision. Such as speech output, motion output (including expression and body speech output), or other forms of output (such as output data streams, images, etc.). The execute response step is a process of translating the plan into actual output.

If the machine selects the speech output in the various possible response steps, it is simple to implement by converting the image features to be output into speech through the in-concept translation, then organizing the language output sequence by using the relations between languages in the relational network (the grammar knowledge existing in the relational network), and calling the pronunciation experience. It should be noted that machines may choose from experience (either self or others) to express the dynamics of an entire sentence (e.g., using different movement patterns of tone, or variation in stress to express questions, jeers, distrust, emphasis, etc., which are commonly used in humans. Because the machine learns these expressions from human life, any expression of a human can be learned by the machine theoretically.

The problem becomes 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 sub-goals and final goals, the rest of which need to be randomly strained in practice.

The machine needs to respond to the sequence target to be output, and the sequence target is divided in time and space according to different time and space related to the targets, so that the execution efficiency of the machine is coordinated. 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 a change from the general scenario to the sub scenario.

The machine needs to expand the intermediate targets in each link layer by adopting a segmented simulation method in combination with the real environment again. The response planning proposed by the machine at the top level is usually composed of highly generalized process features and highly generalized static concepts (since these highly generalized processes can find many similar memories, the 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 the machine is still a process of analyzing various possible results and possible occurrence probabilities by using decision-making capability and selecting own response according to the principle of 'benefiting and avoiding'. The above process is continuously iterated, 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 are "on demand". If "not compliant," it is recreated. This process expands layer by layer, eventually creating a machine rich response.

In this process, the machine may be exposed to new information at any time, resulting in the need for the machine to process a variety of 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.

The machine breaks down other objects to more detailed objects while performing the emulation 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.

Thus, the machine can complete understanding and responding to the input information once by utilizing the capabilities. This process acts as a minimum period for the machine to interact with the environment. The machine is continuously repeated to use the process to achieve a larger target, which is represented by a continuous interaction process of the machine and the outside world and is represented by machine intelligence.

Machine self-experience summary, the content includes:

the experience of the machine, not only through the memory and forgetting mechanism to form a connection in the relational network, the machine can also actively strengthen such a connection. Such actively reinforced connections can take many forms: such as learning the experience of others through language. The machine activates the language to form a stream of information that, together with the language, forms the learned experience of the other person and stores this stream of information in memory as a virtual experience. This experience is stored in the memory as new input information, which is also part of the memory. For another example, the machine may proactively turn those memories into long-term experience by mimicking others' experiences with information that is closely related to the gain and loss, by repeating the memories.

The machine can also adopt a preset algorithm, the memories which can repeatedly appear and can greatly influence the income and the loss are memorized, the information of the corresponding events is virtually input, the virtual processing process is repeated, the related connection in the relational network is strengthened, and the experience is enhanced. In the process of this enhancement, the connections between common parts like those in experience may be gradually enhanced, so the experience becomes more and more concise and universal, and finally some rules for machine self-summarization are formed, which is new knowledge created by the machine itself.

In a second aspect of the present application: how to make the general artificial intelligence similar to human and thus better interact with human. The method mainly comprises the following steps:

1, a sensor group similar to a human sense organ is established.

In robots, such as robots performing security, production, and other tasks, they may optimize sensor composition for their working environment, such as using 360-degree view, using wheeled movements, and so on. However, robots that are responsible for human interaction, such as services, nursing, government, etc., need to have sensors similar to human so that human language and actions can be better understood, such as "one set at the front and one set at the back", which may make it difficult for a 360-degree-view robot to generate correct feelings, and for a robot similar to human to learn to obtain correct comprehension and feelings of the language. And a meaning such as "take a big step forward" may be difficult to understand for a wheeled robot.

The present application therefore proposes to use sensors similar to humans to train machines to understand much of the information relevant to the way humans perceive themselves. The specific method comprises the following steps:

binocular viewing angles, binocular positions, spacings, and viewing angle ranges are used to mimic human binoculars.

A binaural auditory sensor that mimics a human, including a mimic of position, spacing, auditory ability.

The gravity center is learned and adjusted by using a global sole pressure sensor array and input information of sole sensors.

The temperature sensor of the whole body is adopted to sense the outside temperature.

And a gravity sensor is adopted to sense the gravity direction.

The whole body tactile sensor is adopted to sense the whole body pressure and touch.

An olfactory sensor similar to a human is used to recognize the scent.

Taste sensation is recognized using a taste sensor similar to a human.

A human-like fatigue index is established to reflect the degree of fatigue of the machine.

Tension sensors are adopted for all bone joints of the machine, so that the machine can better determine the joint tension.

The four limbs of the machine are provided with acceleration sensors to sense the acceleration of the movement of the four limbs.

The machine needs to establish its own monitoring system for detecting its own attitude pattern.

The machine may also add corresponding sensor groups depending on the particular application.

The sensors, especially the whole body temperature, pressure and touch sensors, can be distributed according to the density distribution of the sensors of human beings, for example, the palm, the finger tip, the toe tip, the face can be densely distributed with the sensors, and other positions can correspondingly reduce the density of the sensors.

2, utilizing the sensor data through a relational network.

The data of the sensors are processed according to the method of the first aspect of the invention, stored in a memory together with other simultaneously input information, and given a memory value, and the connection relationship between the data and the data is established through a memory and forgetting mechanism. The memory value of this information is a function of the intensity of the values obtained by the sensors and of the needs and emotional state of the machine itself. The function can adopt a preset method, and the specific function form needs to be optimized through practice. Usually, the memory value of the machine is positively correlated (not necessarily linearly correlated) with the intensity of the sensor acquisition value, and the machine can adjust the mapping parameter of the preset sensor acquisition value to the memory value function according to the self demand and emotional state.

After all the sensor data are connected with the memory, the machine can continuously correct own behaviors according to experience, so that better income and loss evaluation is achieved. For example, when the four limbs move, the machine does not need to calculate the movement track and the acceleration carefully by using an algorithm, but continuously moves under the control of a preset primary and secondary bottom movement program through practice, senses the tension of joints of the whole body, senses the acceleration change of the moving limbs, senses the slight change of the temperature of the moving limbs, and senses the slight change of the pressure between the moving limbs and the air, thereby summarizing the experience of the movement modes under different environments.

These experiences are also a multi-resolution tower model. For example, the machine may be used to take a cup filled with water. Then at low resolution the machine first performs the experience of getting an object. The experience of taking an object is widely existed in the motion experience of the machine, all the repeated motion experiences obtain high memory values, and the machine firstly calls the experiences to initiate the bottom layer motion instruction of the hand. Then, the data transmitted back by the machine through the hand and the whole body joints are compared with the data of the hand and the whole body joints in the activated experience, and the difference is taken as an error to adjust the hand movement of the machine through negative feedback. The comparison data includes the tension of the joints of the whole body, the acceleration of the moving limb, the temperature change sensed by the moving limb, the slight change of the pressure between the moving limb and the air and the like. The difference in these data can activate previous experience with the machine that uses negative feedback to adjust its own motion floor commands. Similarly, the data may also activate those motor memories that cause failure. These failed motion memories are connected with a loss value memory. Therefore, the machine can adjust the motion state of the whole body according to the decision system, so that the machine obtains the sensor data of the whole body as close to the successful experience as possible and is far away from the failed experience. This process is an iterative process, and the machine improves success rate by modeling a series of process data for successful experience and experience to avoid failures. When the machine body is executed, the memory and experience related to the water cup can be activated because the machine body is used for holding the water cup. Then the concepts of "fragile", "need to be smooth", "potentially hot hand", etc. may also be activated. The machine then further adds this high resolution information to the movement to find a limb movement experience containing more similar details. The machine succeeds by mimicking the successful experience in these experiences, avoiding those failed experiences.

In the process, the machine may further add more high-resolution information, such as "this is an expensive crystal cup", and the like, and the machine may further subdivide and find relevant experiences. If the machine does not have experience with holding an expensive crystal cup, but the machine has experience of "holding things", experience of "holding valuables", experience of "the glass is fragile" (and this experience is activated by the similarity of the crystal cup and the glass), experience of "the glass is broken or damaged", experience of "the loss of things needs to be compensated", and experience of "the compensation of a large amount of money causes a large loss". These experiences may be activated, and the machine needs to calculate various gains and losses that may be brought to the machine itself, and the probability of these gains and losses occurring, so as to establish a series of sub-targets to increase the probability of a path that gains a gain, and decrease the probability of a path that loses, which is the specific decision process of the machine. In the example of "take the crystal cup", if the machine obtains the profit only to satisfy its curiosity, and the machine determines that it is out of hand with a certain probability, which may cause huge loss, the machine may make a final decision by profit and loss decision algorithms (which may be adjusted by learning) without touching the crystal cup. It is also possible to decide to take the crystal cup, but to adjust the motion parameters of the crystal cup more finely: for example, the initial activation value of the input data of the whole body sensor is increased, so that the sensor data is more sensitive to the self, the activated memory range is larger, the activated related memory is more, and the self motion parameters are carefully adjusted. Meanwhile, according to experience, peripheral conditions need to be avoided from interfering with own actions. The machine monitors its surroundings by increasing the initial activation value of the whole body sensor input data and preferentially processing the sensor data. The machine may also focus attention on input and memory information associated with the process of removing the crystal cup while suppressing activation of other input or memory information not associated with the target. This is achieved by performing a preliminary processing of the different input information and giving different initial activation values for further recognition. Those given very low initial activation values will almost quickly be forgotten in the temporary memory bank or will almost rarely trigger the associated associative activation, similar to the human attention mechanism.

Therefore, the overall process of the robot motion control provided by the invention is a process of experience utilization, decision creation, simulation execution and feedback adjustment. It does not only relate to sports but also to past related experiences, which are not limited to sports experiences only, but to all experiences and decision processes. The basis of this process is the world's knowledge and knowledge of the machine itself, so the motion of the machine is a complex intelligent problem, not just a motion control algorithm.

3, establishing similar demand motivation and emotion with human.

In the present application, we propose that in order for a machine and a human to be able to communicate better, the machine needs to establish demand motivations and emotional reactions similar to those of the human.

In the methods presented herein, machines employ symbols to represent the various underlying requirements that various humans impose on the machine. For example, the safety requirements of the machine itself, pursuit of pleasure, desire to obtain human approval, desire to obtain human respect, further, for example, self-reward (sense of achievement) brought by self-goal achievement (goal achievement) of the machine, for example, unknown curiosity of the machine for exploration, and the like. Such as giving the machine a moderate incentive to increase its own energy use efficiency (e.g. giving the machine some lazy nature). Humans can give machines almost all needs and motivation except for proliferation and violence. These requirements can be represented by a symbol, and the symbol can be assigned a value to represent the state. 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.

In the method presented in this application, the machine employs symbols to represent the underlying emotions that various humans impart to the machine. The emotions of the machine can be diversified, each type of emotion can be represented by a symbol, and the symbols can be assigned by the machine to represent the state, such as various emotions of excitement, anger, hurt, tension, anxiety, embarrassment, tiredness, coolness, confusion, aversion, pain, jealousy, fear, joy, romance, sadness, homonymy and satisfaction. The machine is given a difference and how much of the type of emotion it is, without affecting the claims of the present patent application. Since all emotions are treated in the same way in the present application.

In the method provided by the application, the relation between the emotion of the machine and the demand state of the machine can be related through a preset program. The parameters of the preset programs can be self-adjusted by the machine in the learning process of the machine according to the principle of 'tendency towards interest and avoidance from harm'. In the method provided by the application, the emotional state of the machine and the explicit expression mode of the machine emotion can be connected through a preset program. The parameters of the preset programs can be self-adjusted by the machine in the learning process of the machine according to the principle of 'tendency towards interest and avoidance from harm'.

Drawings

FIG. 1 is a basic functional block diagram for implementing general artificial intelligence as proposed in the present application.

Fig. 2 is a schematic diagram of a method for establishing a basic feature.

Fig. 3 is a schematic diagram of a machine decision process.

Detailed Description

The invention is further explained in the following with reference to the drawings. It should be understood that the present application text primarily addresses the main steps and interrelationships between steps that enable general artificial intelligence. Each of these main steps may be implemented using presently known structures and techniques. The emphasis herein is therefore placed upon illustrating the steps and interrelationships between the steps and not upon limiting the details of implementing each step using known techniques. The description of these embodiments is merely exemplary in nature and is in no way intended to limit the scope of the present disclosure. In the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the focus of the present application text. All other embodiments obtained by a person skilled in the art without making any inventive step are intended to be within the scope of protection of the present application.

FIG. 1 is a basic functional block diagram for implementing general artificial intelligence proposed by the present application.

S101 in fig. 1 is a sensor group, and the main function is to mimic the human perception function. The sensor group comprises a sensor for sensing external information and a sensor for sensing self information (such as self-sensing sensors for joint tension, touch, gravity direction, limb state, electric quantity monitoring and the like).

S102 in fig. 1 is a basic feature extraction module. We propose a method of establishing the basic features 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 patterns, 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, and are not described herein again.

S202 is to use a specific resolution window to find local similarity for each channel data. This step is to find common local features in the data window for each channel's data, and ignore the overall information. In step S202, the machine first uses a local window W1 to find local features that are ubiquitous in the data in the window by moving W1. For an image, local features refer to local similar patterns commonly existing in the pattern, including but not limited to the most basic features such as points, lines, planes, gradients, curvatures, and the like, and then local edges, local curvatures, textures, hues, ridges, vertices, angles, parallels, intersections, sizes, dynamic patterns, and the like, which are formed by combining the most basic features. For speech is similar audio, timbre, pitch and their dynamic patterns. The same applies to other sensor data, and the criterion for judgment is similarity.

It should be noted here that the windows of different resolutions may be temporal windows or spatial windows, or a mixture of both. In comparing data similarity within a window, a similarity comparison algorithm is used. In the similarity comparison algorithm, data preprocessing may be performed again, segmentation comparison may be performed on the data again, different windows correspond to different resolutions, and the similarity comparison algorithm at each resolution needs to be preferred through practice. This step is equivalent to our attempt to achieve the feature extraction capability that human beings have. The human feature extraction capability is established by trial and error in the course of evolution. Similarly, in the present application, the machine also needs to establish the similarity contrast algorithm at different resolutions by continuous trial and error with human assistance. Although these algorithms need to be optimized by practice, these algorithms themselves are very sophisticated algorithms that can be implemented by those skilled in the art based on well-known knowledge, and therefore will not be described in detail 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. Every time an existing local feature is found, the memory value of the local feature (basic feature) in the temporary memory library is increased according to a memory curve. The information in the temporary memory library complies with the memory and forgetting mechanism of the temporary memory library. The basic characteristics which survive in the temporary memory library can be put into the characteristic map library to be taken 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. S203 is to repeat the step of S202 using successively the local windows W2, W3, …, Wn, where W1< W2< W3< … < Wn (n is a natural number), to acquire the base feature.

In S204, a basic feature extraction algorithm model a is established by the machine. The algorithm model is an algorithm for finding local similarity: and comparing similarity algorithms. In S205, there is another algorithm model B for extracting basic features. 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.

In S205, the machine trains the multi-layer neural network using the selected information features as possible outputs. Since the information features at the bottom layer are not many, for example, the most essential features in the image, such as points, lines, planes, gradients, curvatures, etc., are mainly, and then the image features are combined by these features. So we can use a layer-by-layer training method. In S205, the machine first selects a data interval using the local window W1, and trains the neural network using the data within the interval. The output of the neural network selects the information features selected at a resolution close to the resolution of the W1 window.

In S206, the machine trains the algorithmic model again using the local windows W2, W3, …, Wn one after another, where W1< W2< W3< … < Wn (n is a natural number). In the optimization, after the window size is increased every time, a neural network layer from zero to L (L is a natural number) is added on the corresponding previous network model. When optimizing this added layer neural network, there are two options: 1, optimizing only an added zero-to-L (L is a natural number) layer neural network layer 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, in S207, 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 of the underlying features may be as large as the entire image. Such base features are typically a background feature map or a specific scene feature map of some images.

The extraction of dynamic features is to regard the objects in the spatial resolution window as a whole, which can be regarded as a particle, to extract the similarity of the motion trajectory. When the motion trajectories are determined, these trajectories can be viewed as static data. The selection of the motion features and the extraction algorithm of the motion features are similar to those of the static data. The rate of change is a motion feature extracted by time resolution (time window), which is sampled over time, and the rate of change is determined by comparing the similarity difference of motion trajectories between different samples. So the motion features have two resolutions, one is spatial and we use a spatial sampling window to implement the data within the window as one particle. One is time, and we sample through a time window and determine the rate of change of motion from the change in motion trajectory in these samples.

In fig. 1, S103A is an initial activation value assignment system, and S103B is a system for realizing the associative activation. After obtaining the basic features of the input, the machine needs to perform associative activation to find relevant information. In S103A, the machine assigns an initial activation value to the input information feature map in accordance with its own motivation by giving the extracted input information feature map. These initial activation values may be the same, which may simplify the initial value assignment system. Due to the processing of the input information by the machine, it may be interactive for multiple times. Machines typically preferentially identify global features, which are typically large class features, and typically have a low resolution. These global features are composed of high-memory-value base features that are included in a specific concept. Because these basic features are widely present in common in the various things, scenes, and processes represented by a particular concept, they are continually repeated as the machine processes such information, thereby achieving higher memory values. These common underlying features form a feature model of the concept. When a concept contains an activation value that reaches a preset confidence level, the machine uses the concept to represent the relevant input feature combination. It may also happen that no concept can reach the preset confidence level, and the machine preferentially uses the concept with the highest confidence level as a model to further judge the input information. This is the way the machine generates the desired information for further identifying the information according to the model. The response of the machine may be to seek further validation in accordance with other portions of the model. The machine thus generates an expected interval that further identifies the input data, and also sets one or more corresponding resolutions to extract the underlying features in the expected interval according to the expected possible information features. The above process may be iterated until the machine successfully identifies the input information. Or in the identification process, the expected model is switched due to more information input, and the identification is continued. Or the machine abandons the recognition because it is difficult to recognize the input information. Or the machine actively seeks more information, such as actively asking or looking for relevant information, by mimicking experience when it is difficult to recognize the input information. These decisions are made by the decision system of the machine.

In addition, during the process of identifying information, the machine can use past experience (activated memory) to search memories related to gains and losses in the activated memory range through associative activation, so that potential gains and losses possibly brought to the machine by each input information can be expected, and therefore, the initial activation value given to each piece of basic information in the process of identifying the machine can be different. It is also possible to assign higher activation values to those information that are more compact with potentially high-gain and high-loss connections, when a single assignment of an initial activation value is made. And different initial activation value assignment modes come from a decision system of the machine. The decision system is based on the common sense and the principle of interest and harm avoidance. The recognition system and the decision system are interactive and help the machine to select the path with the highest ratio of efficiency to energy.

In a specific implementation, the initial value assignment system may be a preset system, and program parameters of the preset system are influenced by the emotional state of the machine. When the machine predicts that serious profit and loss conditions can occur, and the machine needs to deeply analyze various potential profit and loss according to experience, the machine can adjust an initial value assignment program through a preset program, change the initial value assignment size of input information, or improve information which is more closely connected with potential high profit and high loss and is endowed with a higher activation value, so that the whole associative activation process can activate more memories, and a selection which has higher confidence coefficient and is more accordant with the expectation of the machine is made under the existing experience. The iterative identification process of the machine to the external information is the attention mechanism of the machine.

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 universal excitation module and the numbers in the codes are used for representing the initial strength endowed by the characteristics, the characteristics in the memory can receive the information on the bus by 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. Each zone may have multiple features activated for 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 translates into a memory value (not necessarily a linear relationship) according to a positive correlation. Thus, here, the food-related "positive demand" symbol (e.g., demand for deliciousness) 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 "need 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 "need for deliciousness" is established. 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. When the machine activates the memory of 'cake' and 'delicious requirement is satisfied', the emotional symbols may be activated together, so that the machine can feel 'joyful', 'satisfied', and the like.

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 conditions, subdivide the organized process into a large number of intermediate link targets layer by layer through piecewise imitation, and further realize the intermediate link targets step by step. Such as to complete the process of ordering "cake", finding "table" and meeting the needs of the user.

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. 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 graph for initiating activation, nearby memories are searched for initiating close activation, memories farther away with high memory values are searched for initiating strong memory activation, and similar features in other memories are searched for initiating similarity activation. 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: the memory space establishes connections between each feature and adjacent features, through which activation values can be passed out when activated, i.e. 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 it is the associative activation implementation algorithm based on 3 principles of "proximity activation", "similarity activation" and "strong memory activation", relates to the claims of the present application. The machine can use numerical value to represent the strength of a message in memory, code to represent the category of the activated electric signal, bus to represent the space of the activated electric signal and stereo coordinate distance to simulate the propagation loss, so the associative search speed can be much higher than the brain nerve activation mode.

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 basic features for each input of information. 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 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 and spatially rotating the stereo image in 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 basic features in the memory, the method comprises the step of searching for a feature map which can be matched after spatial rotation in the memory. Meanwhile, the machine stores the feature map of the current angle into memory and reserves the original angle of view. And then, the basic characteristic input with similar visual angles is carried out again, so that the search can be quickly carried out. 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 a 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. The root of the recognition of the rotation and scaling of the object by the machine is that the machine models the underlying features parametrically. And size and angle are one of the parameterized properties. When different angle and size basis features are parametrically activated, those combinations of similar angle and size basis features yield a large number of activation values under similar activation. The combined base features (including individual features and low resolution features formed by combining them) all deliver activation values to specific concepts, so that the specific concepts are activated and the corresponding concepts can be identified by the machine. And further activate other properties to which the concept itself relates from the corresponding concept, which is the process of re-developing associations from the concept.

S104 in fig. 1 is an environment information and dynamic manifold creation module. When input information is understood, the environment information and dynamic manifold creation module mainly has the functions of adopting a low-resolution abstract conceptual framework (the framework mainly comprises part of high memory value common characteristics), and then adding related high-resolution information into the framework layer by layer according to related experiences in the input information and memory, so as to form environment information (environment information) and dynamic manifold (dynamic process) for understanding the input information.

When the machine makes a decision, the decision path can be input according to a virtual process by adopting virtual input, and corresponding environment information and dynamic manifold are created by adopting the same method to analyze the profit and loss of the decision path for many times.

The underlying method of creating the corresponding context information and dynamic manifold is piecewise emulation. The essence of the segmented emulation is a process of reorganization using memory and input information, which is a creative process. The method uses information with high memory value (generally ubiquitous low-resolution characteristic information) in memory as a model frame, and continuously adds some input information into the model frame by adopting minimum homonymy inner replacement (which shows that a pair of input information and memory information has tighter connection relationship) between input information and memory information through an iterative identification process and between high activation values. This process is a generalization. In the generalization process, the low-resolution basic features including static features, dynamic features and relationship characterization concepts representing static or dynamic relationships between objects are key bridges. These features can be bridges because they are ubiquitous features in the same category and are widely used, and they have a connection relationship with many other information so that they can be activated and obtain a higher memory value once and again. They are formed by the preponderance and depreciation of memory and forgetting mechanisms. For example, in languages, they are the common expression organization modes such as common words and common sentence patterns.

The machine borrows the process frames and adds the input information to form a new process of shape and color. This process is called segment emulation. The segmentation simulation is an iterative process, and each upper-layer link is expanded into a plurality of lower-layer links meeting the real conditions through the segmentation simulation. Then, in the simulation process, the same method is continuously adopted to expand each lower-layer link into a plurality of lower-layer links meeting the actual conditions again. This process is iterated until the machine is satisfied to perform a response using the newly established process, or a gain and loss assessment is made.

S105A and S105B in FIG. 1 are modules that the machine uses the input basic information to make associative activations in memory and organize its possible responses by means of piecewise emulation. S105A contains the underlying requirements library (including requirements motivation and emotion) of the machine, which is a preset program that humans give to the machine. In S105A, the information including various things, scenes and processes, and their corresponding motivation and emotional information, which are created by the machine through memory and forgetting during the learning process, are also included. In S105B, the machine uses the associative activation to find the experience associated with the input information. These experiences include both experiences for understanding the input information and experiences due to responses to the input information. The machine builds different response paths through piecewise simulation and analyzes the profit-and-loss values under each path, and the magnitude of their probability of occurrence, through statistical algorithms. After information related to the income/loss, emotional state and the like brought by each possibility to the machine and the motivation, the machine can search the most favorable path for the machine in the activated information range by adopting any current artificial intelligence prediction method, such as Bayesian estimation, Monte Carlo search, decision tree, machine inference based on rules and the like.

The decision of the machine is a path planning method based on the prediction capability of the machine. The goal of the path is to maximize the benefit and minimize the loss. With predictive capabilities, the machine transforms decision making and response, a fully open problem, into a string of relatively closed problems of how to increase or decrease the probability of a range of things occurring.

The goal of each step of decision making is to make the development direction of the things "favor and avoid harm". This is a process of interaction with the outside world. The interaction itself is a means to promote the development direction of things to be beneficial and harmful according to past experience. The probability of the event with high profit value is continuously improved through the information and the behavior obtained by interaction, and the probability of the event with high loss value is continuously reduced. This is an iterative process. But the manner in which each step is processed is the same. And the machine increases the probability of the occurrence of the events leading to high profit values layer by layer on the basis of the causal chain. Since the probability of causal links between paths is expressed by a relationship network, the response planning problem for the entire machine becomes a problem of finding the optimal path in the causal chain network.

For example, the machine can determine the prior probability of an event (e.g., an event that results in a high value of gain or loss) by searching the memory. The causal strength (a posteriori probability) between a certain condition and the event can then be determined by means of a relationship network. The connection strength between different conditions in the relationship network can reflect whether the different conditions are independent or not. The machine can predict the probability of the event by only selecting a plurality of relatively independent conditions through a naive Bayes algorithm. The machine may determine its own response based on the calculated probability. This process is essentially an iterative optimal path search problem.

The S106A module is a module for dividing and subdividing sub-objects, and the S106B module realizes the functions of the respective sub-objects by using past experiences through segment simulation. This is the machine implementation.

S107, after the machine processes the previous input information, the generated response may be to further identify the input information. The machine may then be interested in certain specific areas of the input information and have the expected size of the thing. These expected sizes of things determine the resolution at which the machine will further recognize the information.

Fig. 3 is a decision process of a machine.

S310 is input information, S302 associates the activation process. The associative activation process may be initiated by input information or the machine during the decision process based on sub-goals generated by the decision process or new situations encountered.

S303, predicting the possible event by the machine through a segmented simulation method based on the activated memory. And then judging the possible income and loss caused by the predicted events and the probability of the occurrence of the income and loss according to the memories of the activated memories about the income and loss. This is a predicted behavior of the machine based on experience.

S304, S305 and S306 are machine combinations of various possible responses by piecewise emulation within the scope of the information being activated. And then the machine calculates the income and loss caused by various possible responses, searches the optimal response path according to the principle of trending interest and avoiding harm, and establishes each sub-target on the optimal response path.

Since the policy-building behavior of the machine is an iterative process. In the early stage of machine decision, the machine only establishes a rudimentary sub-target sequence which can bring benefits and avoid loss, and the strategy is a framework-class strategy. When the machine executes the strategy, each step of the strategy needs to be refined to the specific executable degree according to the actual situation, namely the strategy needs to be refined to the bottom layer driving command which can be directly executed by the machine.

The machine converts each sub-target on the optimal target path into a specific executable machine bottom layer drive command, and the adopted mode is still as follows: the sub-targets are used as new targets, existing information and memory information are combined, past experience is utilized through segmentation simulation, and an optimal path for achieving the sub-targets is found through combination of reality conditions. The principle of searching for the optimal path to reach the sub-targets is still the principle of interest and harm, and is still to estimate the possible income and loss and probability of various paths and then to find the path with the maximum benefit and the minimum loss as each sub-target on the new optimal response path.

The iterative process continues to be layered until each of the sub-targets on the optimal response path becomes an underlying drive command that the machine can directly execute. The machine then begins executing the underlying drive commands.

The machine will have new information input after executing the underlying drive commands. The new information may come from the outside or from the machine internal state or from a newly activated memory, which becomes the new input information that may activate the new memory by associating the activation process. The machine then needs to add new information or memory and re-evaluate the optimal response path again according to the above decision process. This is the case for S308, S309 and S310.

The above process may iterate repeatedly until the machine has completed all sub-objectives on the planned optimal response path (including sub-objectives on the optimal response path during ongoing adjustment), has reached the final objective required by the machine (this objective may not be the same as the objective established by the original machine, may be a similar objective, or has abandoned the original objective, or even has reached an objective that deviates completely from the original objective, etc.), and the entire response process is ended.

It is emphasized that there may be cases where multiple decision processes are interleaved in the decision process of fig. 3. That is, there are multiple decision making processes like that of fig. 3 that are performed at the same time period. There may or may not be a mutual response between them. But the overall strategy is: when a machine is performing a decision evaluation, the previous targets, and the previous sub-targets, are converted to inheritance targets. These inherited goals are one of the goals that the machine is to achieve, so the machine needs to take all into account simultaneously in the decision making process. However, when the targets are specifically realized, a part of targets can be realized first according to the limitation of real conditions (such as time, space and existing conditions), and the rest targets are taken as inheritance targets and continue to participate in the policy decision process of the machine at the subsequent time.

S108 is a step of updating the memory bank, which is performed throughout all steps, rather than a single step. In step S108, the machine first stores the memory information in the temporary memory library. 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. When the information reaches the preset standard in the temporary memory bank, the machine can convert the information into long-term memory. One method is to move the data in the temporary memory bank to the long-term memory bank, and the other method is to label the relevant information as long-term memory directly and maintain the information by using the memory and forgetting curve of the long-term memory bank.

In the present invention, we can use various forms of memory organization, such as: the time and space relation of information input is directly adopted and stored in sequence, and a three-dimensional coordinate is established 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. For another example, the features are numbered, and each number corresponds to the feature in a table form. 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. The machine may also put similar features together, but each feature has its own spatial coordinates in memory space. Therefore, the machine can quickly find all similar features and realize the proximity activation and the strong memory activation according to the space coordinate information of the features. The machine may also mimic brain neural tissue, establishing connective relationships between adjacent memories. Propagation and attenuation of the activating 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. The machine may also take a combination of the forms described above. However, whatever form of information storage organization is adopted, 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.

Claims (21)

1. A method for identifying input characteristic information by a machine is characterized by comprising the following steps:

the machine endows the input characteristic information with an initial activation value; the machine transmits an initial activation value obtained by inputting the characteristic information in a way of associated activation; the machine selects high activation value memory as an expected model of input characteristic information according to the propagation condition of the activation value; the machine uses the expected model to compare with the input characteristic information, and identifies and segments the input characteristic information; the machine adopts an attention mechanism and performs multiple times of identification on input information through an iterative identification process.

2. The method of claim 1, comprising:

in the process of identifying the input characteristic information by the machine, the attention mechanism means that the machine determines the data interval of the next identification and the adopted identification resolution according to the identification result; the machine can select one to more intervals in a single identification process, and can identify the intervals by adopting one to more identification resolutions.

3. The method of claim 1, comprising:

in the iterative identification process of the external input characteristic information by the machine, in each iteration, the initial activation values given to the external input characteristic information by the machine can be different; different initial activation values can also be given for different features in a single initial value assignment process.

4. The method of claim 1, comprising:

in the iterative identification process of the external input feature information by the machine, more input feature information is obtained through an increasing identification interval and identification resolution, and the range of an expected model of the input feature information is continuously reduced through the change of the activation value until the expected model of the input feature information reaches a preset confidence level.

5. An information storage method, comprising:

the machine considers that the information adjacent in input time has a connection relation with each other, so that the adjacent storage space is adopted to store the information; contiguous storage space means that there is a way to express that two storage units are contiguous; the expression may be physically adjacent memory spaces, or address coding may be performed on the memory spaces to express the adjacent relationship.

6. The method of claim 5, comprising:

the machine not only stores the input characteristic information data in the storage space, but also needs to store data expressing the memory and forgetting mechanism of each stored data; the data stored by the machine not only comprise externally input characteristic information, but also comprise information of internal sensors, and also comprise activated demand motivation and emotion related data; the machine stores demand motive and emotion data, the initial memory value of which is positively correlated with the activation value obtained by the demand motive and emotion symbols.

7. A machine decision process comprising:

the machine identifies memory information related to the input information through associative activation; the machine makes response decision according to the principle of benefiting and avoiding harm by combining the input information and the activated memory information; the machine predicts the probability of occurrence and possible income and loss under the current conditions according to the information of income and loss contained in the activated memory information and the probability of occurrence of the income and loss under different conditions in the memory by a statistical method; the machine adopts a segmented simulation method to combine various possible responses according to the probability of occurrence of possible events and the brought income and loss under the current conditions in the range of activated information, searches an optimal response path according to the principle of tendency toward interest and avoidance of harm, and establishes each sub-target on the optimal response path.

8. The method of claim 7, comprising:

the machine takes the sub-targets on the optimal response path as new input targets, adopts the associative activation method again, searches for the memory related to the sub-targets, and predicts the occurrence probability and potential income and loss of the sub-targets under the current conditions according to a statistical method; the machine adopts a sectional simulation method according to the principle of tendency toward interest and avoidance, reduces the probability of the occurrence of the events bringing loss as a new target and establishes each sub-target contained in the new target according to the probability of increasing the occurrence of the events bringing benefits; the machine iteratively executes a decision process, continuously subdivides the sub-targets, continuously uses a piecewise simulation method, continuously searches the optimal response path, establishes each sub-target on the optimal response path, and finally subdivides the sub-targets on the realization path until the machine can directly execute a bottom-layer driving command.

9. The method of claim 7, comprising:

the process of searching the optimal response path and iteratively decomposing the optimal response path layer by the machine is a dynamic change process; in the process, the machine can perform decomposition while performing, and re-perform the task of finding the optimal response path according to the latest obtained information, and further iteratively decomposing the optimal response path layer by layer.

10. The method of claim 7, comprising:

when the machine understands input information and establishes a local response path, a segmented simulation method is used for combining reality information and memory information into information combination forms such as 'environment information' and 'dynamic manifold' to establish a dynamic process; this dynamic process is used to represent input information or to represent a planned response path.

11. The method of claim 7, comprising:

the activation value of the activated information in the memory will decline with time, the parameters of the decline being related to the needs and the state of the needs, emotions and emotional states of the machine.

12. The method of claim 7, comprising:

when the machine encounters newly input information during the process of responding to the input information, the machine can convert the original target into the inherited target, and re-identify, understand and respond to the information in combination with the new information.

13. A method of implementing a human-like intelligent robot, comprising:

sensor groups similar to human perception organs are established, and the sensor data are utilized through a relation network to establish demand motivation and emotion similar to human.

14. The method of claim 13, comprising:

the sensor group of the machine comprises: simulating human binoculars by using binocular visual angles, binocular positions, intervals and visual angles; a binaural auditory sensor that mimics humans, including a mimic of position, spacing, auditory ability; a full sole pressure sensor array is adopted, and the gravity center is learned and adjusted through input information of a sole sensor; sensing the outside temperature by adopting a whole body temperature sensor; sensing the gravity direction by adopting a gravity sensor; a whole body touch sensor is adopted to sense the pressure and touch of the whole body; using a human-like olfactory sensor to recognize odors; using a taste sensor similar to a human to recognize taste; establishing a human-like fatigue index to reflect the fatigue degree of the machine; tension sensors are adopted for all bone joints of the machine, so that the machine can better determine the joint tension; the four limbs of the machine are provided with acceleration sensors for sensing the acceleration of the movement of the four limbs; the machine needs to establish a monitoring system for detecting the posture mode of the machine; the machine may also add corresponding sensor groups depending on the particular application.

15. A method of robot motion control, comprising:

the motion control of the machine is essentially a simulation process of multi-resolution experience by continuously adding high-resolution experience on the basis of low-resolution experience; the robot motion control process is a process of experience utilization, decision creation, simulation execution and feedback adjustment; it is not only a motion control algorithm problem, but also relates to past experiences which are not limited to motion experiences only, and may also relate to other experiences.

16. The method of claim 15, comprising:

the motion experience of the machine is a multi-resolution tower model; the machine continuously activates motion memory information on different resolutions in an iterative identification mode according to selection of different data intervals and different resolutions of external input information; the machine compares the multi-resolution data transmitted back by moving limbs and joints of the whole body with the motion-related multi-resolution data in experience, takes the difference value as an error, and adjusts the action of the machine through negative feedback; the comparison data comprises data such as the tension of joints of the whole body, the acceleration of the moving limb, the temperature change sensed by the moving limb, the slight change of the pressure between the moving limb and the air and the like; the difference in these data activates the machine's previous experience by which the machine uses negative feedback to adjust its own motion floor commands; these data include those experiences that were successful and possibly also those that activated motor memory that brought about failure; successful experience is connected with the profit value, and the failed exercise memories are connected with the loss value memory; therefore, the machine can adjust the motion state of the whole body according to the decision system, so that the machine obtains sensor data of the whole body as close to the successful experience as possible and is far away from the failed experience, and the success probability is improved.

17. The method of claim 15, comprising:

the motion process of the machine is a decision-making process, and not only relates to the motion per se, but also relates to the common knowledge established by the machine, the motivation and the emotion of the machine, and the specific external environment; it is integrated in the machine decision making process proposed in the present application, rather than existing in a separate motion control module; in the invention, the motion control of the machine and other decision processes of the machine are the same processing method, and a separate motion decision system is not needed.

18. The method of claim 1, comprising:

the machine can preset some models, and the overall characteristics of the models have high memory values, so that the models are easier to be activated preferentially, and are identified preferentially in the iterative identification process of the machine.

19. The method of claim 1, comprising:

one method of empirical generalization is to use the generalization principle of "the high activation value in memory and the high activation value in input information can be replaced with each other based on a dynamic process or a connection concept representing a relationship" which belongs to the same minimum concept.

20. The method of claim 7, comprising:

the machine presets the imitation motivation as a bottom requirement in a machine program; in different stages of machine learning, the machine can be endowed with simulation motivations with different strengths; for example, when the machine learns the language and the action output, the machine can be directly endowed with a strong imitation motivation, and at other stages, the machine can be endowed with a normal imitation motivation.

21. The method of claim 7, comprising:

when the machine identifies similar basic features, the machine also identifies the similar features after rotation and scaling; the basic features comprise a machine parameterized basic feature model and also comprise a basic feature combination formed by extracting the basic feature combination again, and the basic features are also machine-created basic feature models.

Images