A method for building strong artificial intelligence

Abstract

The method for realizing strong artificial intelligence proposed in the present invention application adopts the mapping of input data to local common features at the perception layer, and the mapping from local common features to concepts at the cognitive layer. The establishment of concepts is realized through relationship networks, memory and forgetting mechanisms. The basis of relationship networks is multi-resolution features and associative activation. The decision-making of the machine is achieved through the connection between the machine needs and emotional motivations established in the relationship network, as well as the motivational state and specific things, the principle of seeking benefits and avoiding harm, and the iterative use of prediction, decision-making and response methods. Through the method proposed in the present invention application, the machine can gradually obtain responses to input information from simple to complex, and have motivations and emotional expressions similar to humans. These all show that there is a huge difference between the machine learning method proposed in the present invention application and the machine learning methods currently available in the industry. There is currently no similar method in the industry.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to how to establish strong artificial intelligence. Background Art

[0002] At present, artificial intelligence is usually designed for specific tasks, and there is no general artificial intelligence that can complete multiple uncertain tasks. The biggest obstacle to achieving general artificial intelligence is how to establish a cognitive network similar to human common sense between complex things. Only when machines have common sense similar to humans can they produce thinking activities similar to humans. The result of current deep learning is a sophisticated feature mapping method, which is quite different from the human learning process, so the results of deep learning are difficult to generalize and generalize. Current knowledge engineering, expert systems or knowledge graphs all use coding to organize human knowledge in a way that computers can recognize. However, these systems make it difficult for machines to learn and summarize autonomously, so when faced with differentiated scenarios, machines cannot autonomously generate new strategies and methods. So far, these systems can only be applied to a specific field and specific range, and cannot produce intelligence similar to that of humans.

[0003] This invention is based on the patent invention application of the same applicant with application number 202010400739.7, entitled “A method for imitating human memory to realize general machine intelligence”, and further explains how to realize general artificial intelligence and further elaborates on the details of the implementation method. Summary of the invention

[0004] In the patent invention application with application number 202010400739.7, entitled "A method for imitating human memory to achieve general machine intelligence", a method of establishing a relationship network between things through memory is revealed. In this invention application, we further deepen how to establish strong artificial intelligence (general artificial intelligence) by establishing a relationship network in memory.

[0005] In this invention application, we propose a conceptual diagram of the composition of a machine. Figure 1In the present invention, the S101 module is the sensor module of the machine. In order to make the machine produce a similar cognitive method as humans, the S101 module needs to use one or more general sensors: visual sensors, auditory sensors, taste and olfactory sensors, tactile sensors, gravity direction sensors and posture information sensors, etc., and sensors for specific applications can also be added (for example, autonomous driving can add infrared sensors, lidar sensors, etc.). The machine also needs to use sensors to monitor its own status, which are also part of the machine's perception information. S101 is a module mainly composed of sensor hardware and sensor corresponding software. Its purpose is to perceive information outside the machine and the machine itself through sensors. The differences and numbers of these sensor types do not affect the claims of the present invention application. Because in the present invention application, all sensor data are processed in the same way.

[0006] exist Figure 1 In the S102 module, the machine simplifies the sensor input information. The machine simplifies the input information, mainly refers to the machine extracting the underlying features of the input information. It can use any existing feature extraction method, including but not limited to convolutional neural network, image segmentation, contour extraction, downsampling feature extraction, etc. Any existing machine image recognition algorithm can be used in the S102 module.

[0007] However, the difference between the S102 module and the current mainstream machine algorithm is that: 1. The S102 module is not intended to identify specific things. In the currently popular neural network, the machine processes the input data layer by layer, and then optimizes the data processing parameters through error back propagation, with the goal of achieving minimum error under large sample statistics. The algorithm implements the mapping from data space to label space. In the present invention application, the goal of the S102 module is to extract local common features in the input data. In the patent invention application with application number 202010400739.7, entitled "A method for imitating human memory to achieve general machine intelligence", we proposed a method of using sampling windows of different sizes and different resolutions to repeatedly extract input data, and establish connections by placing these data adjacent to each other in memory. And strengthen the ubiquitous connections through memory and forgetting mechanisms, and weaken those occasional connections. Therefore, in the present invention application, the purpose of the S102 module is to find those widely existing local common features, rather than to find the mapping relationship between specific sample space and label space. In the present invention application, the same input sample space may contain a large number of "local common features", which are extracted at different resolutions. It should be pointed out that the combination mode of local features of things is also a local feature. Local features have nothing to do with size, but refer to part of the information of things extracted at different resolutions. The local features of some images can be as large as the image itself, but the resolution is low, and only contain part of the information of the original image, such as the composition mode of other local features of the original image. For example, in an image, local features may include underlying geometric features such as contours, straight lines, curves, textures, vertices, verticals, parallels, curvatures, etc. at different resolutions, and also include features such as color and brightness at different resolutions. It may also include motion patterns at different resolutions, and also include the combination topological mode of underlying geometric features. The current popular deep convolutional neural network is to find the mapping relationship between the same input sample space and a small number of specific labels. In the S102 module, the deep convolutional neural network can be used as an application algorithm to realize the mapping between input data and local common features. This algorithm itself does not belong to the claims of the present invention application, but finding multi-resolution local common features from the input data and using these multi-resolution local common features to establish a connection relationship between things belongs to the scope of rights required by the present invention application.

[0008] The S102 module also includes the extraction of multi-resolution dynamic features. Similar to image feature extraction, the S102 module also extracts local common dynamic features. The local common dynamic features here refer to basic motion patterns, such as swings, circles, straight lines, curves, waves, and other similar basic dynamic features that are widely present in our world. Therefore, it does not establish a mapping between a large number of motion sample spaces and specific labels that represent dynamics (such as label spaces such as dance, running, parades, and carnivals), but rather a mapping between the input sample space and those similar basic dynamic features that are widely present in our world.

[0009] In particular, dynamic features are the basis of knowledge generalization. Human analogical application (generalization) of knowledge must be an association based on some similarity. This similarity can be static similarity (such as similar appearance or similar abstract features) or dynamic similarity (such as similar movement or similar change of abstract features). The dynamic features themselves can be represented by abstract particles or volumes, so the motion features can serve as a bridge between the experience generalization of different things.

[0010] In the S102 module, the machine also uses a similar method to process other sensor input data, including extracting static multi-resolution features and dynamic multi-resolution features. For example, for speech, the basic speech and speaking speed can be used as a static feature, while changes in audio, pitch, and speaking speed are dynamic features. The machine slides and samples the speech according to time windows of different lengths, which is equivalent to different time resolutions. The machine needs to extract static and dynamic features at different time resolutions and different detail resolutions. In the strong artificial intelligence implementation method we proposed, the use of multi-resolution to extract dynamic features of things is a crucial part. The multi-resolution extraction method of dynamic features has been described in the patent invention application with application number 202010400739.7, entitled "A method for imitating human memory to achieve general machine intelligence", which will not be repeated here.

[0011] The S102 module is a perception layer feature extraction. It is crucial to extract static and dynamic features from input data at multiple resolutions. Because the connection between things is different at different resolutions. The similarity between things is also different at different resolutions. Therefore, the machine needs to establish a relationship network between things at different resolutions. Two things that are unfamiliar in daily cognition may have similar attributes at different resolutions. And these attributes are the bridge for the generalization of related knowledge.

[0012] The S102 module is a perception layer local feature extraction, but it does not extract all multi-resolution local features every time. Instead, it determines which resolutions to use and the key intervals to extract based on the machine's search target for sensor data. The machine's search target for sensor data comes from the expected target generated by the machine in previous activities.

[0013] S102 is the information processing layer of the machine perception layer, and is a software layer that simplifies the sensor input data. The input of S102 is the data collected by the sensor and the parameters sent from the machine cognition layer. These parameters are the resolution range that the cognition layer tells the perception layer to use and the range that needs to be extracted. These parameters are determined by the machine after processing the previous information and are determined according to the target size and attributes that need to be further identified. They are part of the machine's response to the input information.

[0014] In S102, various specific algorithms such as currently popular convolutional neural networks, recurrent neural networks, and image filtering processing can be used, but their output targets are local common features, rather than being directly mapped to a specific classification space. The local common features are mapped to a specific classification space by the cognitive layer.

[0015] In the present invention, the assumption that multi-resolution feature extraction and temporally adjacent input information are connected to each other is a key step in establishing the cognitive layer. In the S103 module, the information connection relationship (relationship network) in the memory is optimized through associative activation, memory and forgetting mechanisms, which is to establish the cognitive layer.

[0016] In this invention application, we put forward a basic assumption: "There is a connection relationship between the information input by the sensor groups that are adjacent in time." This is the key assumption for establishing the cognitive layer we proposed.

[0017] The machine stores input information in the order of input. This information includes external sensor data, internal sensor data, motivation data such as demand and emotion data. External sensor data is the machine's perception of external information. Internal sensors are the machine's own monitoring information. Demand is a kind of motivation and motivation state preset in the machine. Emotion is also a kind of motivation and motivation state preset in the machine. Demand and emotion are part of the preset motivation of the machine.

[0018] In this invention application, we can give various motivations to the machine, which are the driving force that drives the machine to respond to input information. They are the control mechanism behind the machine's behavior. In this invention application, we take giving needs and emotions to the machine as an example to illustrate how the machine determines its own behavior based on these motivations. The motivations that can be given to the machine include not only needs and emotions, but also other types of motivations. The differences and quantities of these motivation types do not affect the claims of this invention application. Because in this invention application, all types of motivation data are processed in the same way.

[0019] In the method proposed in this application, the machine uses symbols to represent various underlying needs that humans give to the machine. For example, the machine's own security needs, the pursuit of happiness, the hope of gaining human recognition, the hope of gaining human respect, and the self-reward (sense of accomplishment) brought by the machine's self-goal realization (goal achievement), such as the machine's curiosity to explore the unknown, etc. These needs can all be represented by a symbol, and this symbol can be assigned a value to indicate the state. The difference and number of demand types do not affect the claims of the present invention application. Because in the present invention application, all needs are handled in the same way.

[0020] In the method proposed in this application, the machine uses symbols to represent various underlying emotions given to the machine by humans. The machine can have a variety of emotions, and each type of emotion can be represented by a symbol, which can be assigned by the machine to indicate the state. The differences and number of these emotion types do not affect the claims of the present invention application. Because in the present invention application, all emotions are processed in the same way.

[0021] In the method proposed in this application, the relationship between the machine's emotions and the machine's demand state can be linked through a preset program. The parameters of these preset programs can be self-adjusted by the machine in its own learning process according to the principle of "seeking benefits and avoiding harm".

[0022] In the method proposed in this application, the emotional state of the machine and the explicit expression of the machine's emotions can also be linked through preset programs. The parameters of these preset programs can be self-adjusted by the machine in its own learning process according to the principle of "seeking benefits and avoiding harm".

[0023] In module S103, we need to establish connections between external information at different resolutions, memory information, internal sensor information of the machine, and the motivation-related symbols and states of the machine. This connection relationship needs to correctly reflect the common sense of the world we live in. This connection relationship is called a relationship network in the present application.

[0024] In the world we live in, the relationships between things are complicated. It is very difficult to artificially establish various relationships between things, and it is also difficult to quantify and flexibly apply them. In the present application, we extract the relationship between things through memory. First, we use the concept of mirror space to store the extracted multi-resolution information. Mirror space means that we store information according to the time order of input information, and for information input at the same time (such as images), we store information according to the corresponding original space organization method. The specific method is: after the machine extracts multi-resolution information features from the input, the machine needs to use these features to establish a mirror space. The machine first adjusts the position, angle and size of the underlying features by scaling and rotating the extracted features according to the position, angle and size with the highest similarity to the original data, and overlaps them with the original data, so that the relative positions of these underlying features in time and space can be retained, and a mirror space can be established. What the machine stores in memory is multi-resolution feature data organized in a way that is most similar to the original data, which we call a mirror space.

[0025] The information stored in the memory has its own memory value. The new memory stored in the memory includes the memory value of the demand symbol and the memory value of the emotion symbol. They are related to the activation value of the corresponding information (symbol) when the storage occurs. They are usually positively correlated and can be linear or nonlinear.

[0026] In order to establish the relational network in memory, in this invention application, we put forward a basic assumption: "There is a connection relationship between the information input by the sensor group adjacent in time". This is the key assumption for us to establish the relational network. At the same time, we put forward three other assumptions for the optimization of the relational network: "proximity relationship" assumption, "similarity relationship" assumption and "memory strength relationship" assumption.

[0027] "Proximity" hypothesis: We assume that in memory, memory information that is adjacent in time is connected to each other. Adjacent refers to the temporal proximity in memory storage. "Similarity" hypothesis: In memory, similar memory information is also connected to each other. "Memory strength" hypothesis: In memory, those memories with high memory value are more likely to be activated.

[0028] When a piece of information in memory is activated, it will use the "proximity activation" principle, "similarity activation" principle and "strong memory activation" principle to activate other information.

[0029] "Proximal activation" means that after specific information in memory is activated, it will activate the information that is adjacent to it (that is, information that has a proximity relationship).

[0030] "Similar activation" refers to a specific feature in the memory. When receiving the activation signal sent by other features, the receiving ability is positively correlated with the similarity between each other. This is a directional receiving ability. So after a similar memory is activated, it will send out its own activation signal and may easily further activate other similar memories. This is because similar memories have a strong ability to receive each other's activation signals. In the present application, a simple activation value transfer relationship can be that the transfer coefficient is proportional to the similarity. Of course, other transfer functions can also be used, but the principle must be that the transfer coefficient is positively correlated with the similarity.

[0031] "Strong memory activation" means that the higher the memory value, the stronger the ability to receive activation signals from other features. Therefore, those deeply remembered information are more likely to be activated. In the present application, each memory information is assigned a memory value to indicate the time it can exist in the memory. Those memories with high memory values ​​may exist for a long time, and they have a strong ability to receive activation signals from other features. This is to imitate the number of synapses in the human brain to represent the strength of memory, and it is assumed that memories with more synapses are more likely to obtain more activation energy from the surrounding environment and be activated.

[0032] The above three activation methods are collectively referred to as associative activation.

[0033] In the present invention application, we use the memory and forgetting mechanism to maintain the memory value of information in the memory bank. The memory and forgetting mechanism is a widely used relationship extraction mechanism in the present invention application. In the present invention application, each time the information in the memory is activated, it is considered to be used once, so the memory value is increased according to the memory curve of the memory bank in which it is located. At the same time, all memories decrease the memory value according to the forgetting curve of the memory bank in which they are located. The memory function refers to that certain data increases with the increase of the number of repetitions. The specific increase method can be represented by a function, and this function is the memory function. It should be pointed out that different memory functions can be used for different types of data. The forgetting function refers to that certain data decreases with the increase of time. The specific decrease method can be represented by a function, and this function is the forgetting function. It should be pointed out that different forgetting functions can be used for different types of data. The memory and forgetting mechanism refers to the use of memory functions and forgetting functions for memory information.

[0034] If we regard memory as a three-dimensional space containing countless information, then the relationship network is the context in this space. The emergence of these contexts is due to the memory and forgetting mechanism. Those relationships that cannot be repeatedly activated are forgotten, while those relationships that can be repeatedly activated are strengthened. The multi-resolution information connected by the thick relationship context constitutes the concept. It connects images, voices, texts or any other forms of expression of the same type of information. Since these forms of expression appear together frequently and frequently convert to each other, the connection between them is closer. Since humans use language very frequently after birth, usually in a concept, the number of activations of language may be the most, and the memory value of language is also the highest. At the same time, since language is usually connected to all the attributes of a concept, it is a bridge for mutual activation between various attributes. So it seems that language is the center of our conceptual thinking. However, in the patent application of this invention, we believe that language needs to be reconstructed through language to obtain its true corresponding meaning, so the activated information flow after language reconstruction is the information that truly bears our thinking.

[0035] The connections in the relational network are linked by "proximity activation", "similar activation" and "strong memory activation". The tightest local connection relationship constitutes the concept (including static feature maps and their language at multiple resolutions, dynamic feature maps and their language); experience is a little looser than the concept. Experience is those cognitive relationships that often recur. Because they can recur, the memory connections between them can be gradually increased. The experience that is universal in human cognition is common sense. Memory is looser than experience.

[0036] In the present invention application, we propose a memory organization form that can simply express the relationship network: information is stored in the order of input time. Those "relationships that are adjacent in input time" are expressed by "storage locations that are adjacent in space". Information that is adjacent in storage space can be physically adjacent: that is, information that is adjacent in time is stored in adjacent storage units. Information that is adjacent in storage space can also be logically adjacent: that is, it is stored in a logically adjacent manner, and the specific physical storage unit location is represented by a mapping table between the logical location and the physical location. Another method can also be that each stored information has its own storage time coordinate, and the machine determines the adjacent information by searching for adjacent time coordinates. Of course, there can be other storage methods, but they must all be able to express information that is adjacent in time.

[0037] In this invention application, we propose a method of searching for similar information: one method is to use a dedicated similarity comparison calculation unit to handle the task of finding and inputting similar memories from memory. The similarity comparison calculation unit can be implemented using hardware or software. It can be a separate module or a module integrated into the entire operation unit. Similarity comparison is a very mature algorithm and will not be described in detail here.

[0038] When activations are close to each other, the information activation value transfer coefficient is large.

[0039] During similar activation, similar information transfers activation values ​​through similar activation, and the transfer coefficient is positively correlated with the similarity.

[0040] When strong memory is activated, information with high memory value is more likely to be activated by obtaining large activation value from adjacent activation. Therefore, strong memory activation is a specific case of adjacent activation.

[0041] In the above-mentioned "proximity activation", "similarity activation" and "strong memory activation", the transfer function of the activation value needs to be determined through practice, but the basic principle is that the activation value transfer coefficient of "proximity activation" is inversely correlated with the time distance between the two information storages. The time interval can be considered as a medium for the attenuation of activation value propagation. The activation value transfer coefficient of "similarity activation" is positively correlated with the similarity between the two information. "Strong memory activation" is when the activation is proximal, and the activation value transfer coefficient not only considers the attenuation of the time interval, but also needs to consider the memory value of the party receiving the information. The memory value of the party receiving the information can be considered to reflect the size of the ability to receive the activation value.

[0042] It should be pointed out that the same object may have a large number of features at different resolutions to represent the attributes of the object. The similarity between two objects at different resolutions is different. After a feature (attribute) at a certain resolution of an object is activated, this feature may activate features at other resolutions of the object through proximity activation, and may also activate other features similar to the object at the current resolution through similarity, thereby further activating other attributes of other objects through proximity activation. Therefore, similarity is based on the specified resolution. This is why we need to extract multi-resolution features from the input information at the same time.

[0043] In the present invention, the specific activation value transfer function does not affect the basic working principle of the machine. However, the optimal activation value transfer function for each type of activation needs to be determined through practice, but they must all comply with the limitations of the above principles.

[0044] It is also important to point out that the activation value obtained by information in the relationship network will fade over time. The decay curve needs to be optimized by practice. The length of the decay time needs to balance the connection between the previous and next activation information and the activation state of the new input information. If the decay time is too long, the activation state brought by the new input information is easily covered by the original activation state, and the connection relationship between the new information and other information cannot be clearly expressed. If the decay time is too short, the connection relationship between the previous and next activation information is easily ignored.

[0045] In the relational network, static concepts are analogous to small parts widely used in machines, while those dynamic feature graphs (including concepts representing relationships) are analogous to widely used connectors. The large framework representing a class of processes is composed of multiple small parts (static objects) and connectors (dynamic features) organized in a certain time and space order. They can often be used for experience imitation. Among them, dynamic feature graphs (including concepts representing relationships) can often be used as tools for experience generalization, because they play the role of abstract relationships of changes in abstract things and are common attributes across specific things. The essence of the generalization process is the process of analogizing and applying existing experience through common attributes. Common attributes may include certain features at multiple resolutions (such as certain similar features at low resolution, or one or more similar features at high resolution), and may also include similar attributes connected by cognitive layers as bridges (such as language, analog dynamic patterns). However, these generalized relationships and bridges can be reflected in the associative activation process of the relational network, and are generalized and applied through prediction, decision-making and execution systems, so we will not specifically explain how generalization is achieved here.

[0046] In the memory space, external information, internal information, needs and need states, emotions and emotional states, as well as specific motivations and corresponding motivation states that can be added for other applications of the machine, are directly stored in the memory in a time-adjacent order. They all maintain their memory values ​​according to the memory and forgetting mechanisms, and they all perform associative activation according to the "proximity activation" principle, the "similar activation" principle, and the "strong memory activation" principle. Therefore, in the present application, the processing methods of these information are similar. Their types, quantities, and data formats do not affect the claims proposed in the present application.

[0047] In the present invention application, another feature of memory storage is to store the motivation and motivation state data of the machine (such as demand and demand state data, emotion and emotion state data) in the memory, so that this information, like other information, establishes an activation value transmission path between other information. That is to say, when a memory information is activated, it may transmit the activation value to many demand and emotion symbols in the related memory through chain activation. The size of the transfer coefficient is the direct or indirect connection strength in the relationship network optimized by the memory and forgetting mechanism. The essence of the relationship network is a causal network, which reflects the causal relationship between two pieces of information. Information that is adjacent in time usually has a causal relationship, and those similar connections connect the causal relationships that exist at different times in series, forming a large causal network.

[0048] exist Figure 1 In the S104 module, we need to build an application layer that utilizes the relationship network. The application layer includes:

[0049] 1. Initial activation value assignment system.

[0050] The input data of the sensor must first be simplified in step S102. Then the machine uses these simplified information features and uses the initial activation value assignment program to assign initial activation values ​​to these input information. The initial activation value assignment program is a preset program, and its input parameters include input information, the machine's needs and emotions and other motivational states at this time. Its output includes the initial activation value assigned to the input information.

[0051] The machine can also repeatedly assign activation values ​​to the input data after simply processing the data. This step is one of the response results of the prediction, decision-making and execution system output.

[0052] 2. Lenovo activation.

[0053] After the machine assigns the input information an initial activation value, it uses the "proximity activation" principle, the "similar activation" principle, and the "strong memory activation" principle to perform chain activation. The machine uses chain activation to achieve the associative function, so we also call it associative activation. The machine uses associative activation to find (a) experiences related to the input information. (b) motivations and motivational states (including needs and need states, emotions and emotional states) that are connected to these related information. (c) other types of information that are related to the input information in memory.

[0054] The information activated by the machine through association is usually a memory with a higher memory value. These memories are usually obtained through the memory and forgetting mechanism because they can be repeatedly activated. And the reason why these memories can be repeatedly activated is that such information relationships can be reproduced continuously in our lives. They are the "experience" gained by the machine. Therefore, the memory and forgetting mechanism is an optimization mechanism for the relationship network and the basis for the generation of intelligence.

[0055] The memory activated by the machine based on the input information may include all types of memories in memory. For example, when voice input is received, the machine may activate images, texts, feelings, emotions, other related voices, or a past memory that is closely related to these languages ​​based on similar voices in memory. The specific activation content depends on: (a) the relationship network obtained by the machine through learning experience and learning parameter settings. (b) the machine's chain activation parameter settings (this is equivalent to setting the machine's associative method). (c) The initial activation value assigned by the machine to the input information. The larger the initial activation value, the more content the machine can activate.

[0056] 3. Activate information reconstruction.

[0057] Through associative activation, the machine acquires experience related to the input information. These experiences need to be further processed before the machine can refer to these experiences to respond to the input information.

[0058] Environmental reconstruction: After entering an environment, the machine identifies specific objects, scenes, and processes by extracting the underlying features of images, language, and other sensor inputs. It also overlaps the features of similar objects, scenes, and processes found in memory with similar parts in reality, so that the machine can infer the temporarily invisible parts of objects, scenes, and processes.

[0059] Since the size of an object is one of the low-resolution features, the machine also uses the size of the specific object in the field of view to compare with the normal size of the object in the feature map to help the machine establish the depth of field in the environment. The machine usually adjusts the size of the reconstructed environment according to the size of its own target (usually the image related to the concept with the highest activation value).

[0060] Reconstructing a three-dimensional environment through multiple local information (including the approximate framework and different details of a type of object at different resolutions) is a mature technology and will not be elaborated here.

[0061] Self-reconstruction: By overlapping similar parts of multiple memories about self-information (including memories of self-information at different resolutions, some memories are about the overall framework, and some memories are about details), a three-dimensional image of self-information is constructed. This is self-image reconstruction. Since the machine's memories about its own information include some visual, some tactile, some emotional, and some demand, when their common parts are overlapped together, they constitute a reconstruction of the overall perception of the self. This information is integrated into our overall image by overlapping similar parts. We create a self-image in our minds, and we seem to be able to see the movements of the self-image. This self-image is part of our self-consciousness. It is the basis for the machine to distinguish between self and non-self.

[0062] After establishing the external mirror space and the machine's own mirror image, the machine's consciousness is a way of distinguishing itself from the outside world and determining its interaction with the outside world in a way of "seeking benefits and avoiding harm". Therefore, the essence of consciousness is a way of behavior. It is precisely because of the distinction between self and non-self that the machine has consciousness and establishes a connection between itself and external things. The essence of this connection is the relationship between the "benefits" and "harms" of external things and itself. And this relationship is gradually established in the learning process of the machine.

[0063] After the input information activates many related memories in memory, the machine uses these memories to build a mirror space and the machine's own mirror, and uses the mirror space and the machine's own mirror as a way to combine these experiences. The machine views these memories from a third-person perspective. These memories contain information about the machine's emotions, needs, and results. The machine uses this information, in a way that seeks benefits and avoids harm, and draws on these experiences to plan its own response to the input information, and may evaluate the possible results of these responses many times.

[0064] 4. The underlying motivation of the machine.

[0065] The power source of a machine is its motivation, which can be summarized as "seeking benefits and avoiding harm". Some of the benefits and harms are preset, while others are established through learning, because they are related to the needs of the machine itself. Analogous to humans, for example, at the beginning, "water", "milk", and "food" are the innate "benefits". Later, through learning, we learned the connection between "test scores" and "money" and our innate needs. Later, we also discovered that the objects of operation can also be intangible things such as "love" and "time". We even pursue dominance in the group. This is an extension of the pursuit of benefits and avoidance of harm in the underlying motivation in our genes.

[0066] Using a similar method, we can also give machines the motivations that humans want them to have. Because in the relationship network, when all memories are stored, the machine's demand symbols and corresponding memory values ​​are stored at the same time. These memory values ​​are positively correlated with the state value of the demand symbols at the time. For example, if the machine receives a blame after a certain behavior. Since blame is a kind of loss (this cognition can be preset, expressed through the trainer's language, or directly modified to achieve), and the degree of blame (such as the words that indicate the degree in the language) brings different loss values ​​to the machine. The stronger the blame, the higher the memory value the machine gives to the loss symbol in this memory. Then in this memory, since the memory value of the loss symbol is relatively high, all other feature graphs with relatively high memory values ​​in this memory frame have a strong connection with the loss symbol. If a similar blamed behavior occurs again in a similar environment, with a similar action issuing object or receiving object, then the feature graph that brings loss and the loss symbol itself in this memory frame will increase in memory value according to the memory curve because they are repeated, thereby increasing the relationship between the feature graph that brings loss and the loss symbol. Through repeated repetition, the relationship between the feature graphs that actually caused the loss and the loss symbol was selected according to the memory and forgetting mechanism. The machine did not know why it was scolded at the beginning, but later it could clearly understand what caused it to be scolded. This process is similar to the learning process of human children.

[0067] Even if the machine does not receive timely feedback when the behavior occurs, the trainer may later point out the behavior itself and give feedback, thus connecting the behavior and the result in a single memory frame. The trainer does not even need to specify which behavior is good or bad. The machine only needs to receive the correct feedback each time, and through memory and forgetting, it can gradually establish the connection between the correct behavior and the required value.

[0068] Therefore, we only need to give the machine basic underlying motivations, and the machine will establish relationships between various specific things and underlying motivations based on external feedback. And these relationships are the basis for the machine to make decisions. The underlying motivations of the machine are relatively simple and can be preset. For example, give the machine the motivation to learn and abide by human laws, give the machine the motivation to seek human recognition, give the machine the motivation to avoid danger, give the machine the motivation to protect the owner's safety, etc.

[0069] Machine motivation and motivational state are closely related to emotions. In this invention application, we use motivation and motivational state to determine emotions through preset programs, and also use preset programs to realize the relationship between emotions and emotional manifestations. That is to say, using preset programs, the emotions of the machine are expressed through actions (expressions and body language) and language (language output methods). But at the same time, emotions or emotional expressions can also be adjusted through the motivation of "seeking benefits and avoiding harm". Through learning and feedback, the machine has learned what emotions can bring gains and losses in what environment, and in turn adjusts emotions or emotional expressions. Precisely because of the close relationship between emotions and machine motivation (including motivational state), the motivation of the machine can also be uniformly expressed by emotional needs. Machines can make choices and responses by pursuing emotional needs.

[0070] It should be pointed out that when the machine stores any input information, it will also store the machine's motivation and motivation state (such as needs and need states, emotions and emotional states, etc.). In memory, motivation is represented by a symbol, and its memory value is positively correlated with the activation value when it is stored. Therefore, motivations that obtain high activation values ​​may become long-term memories because of the high memory values ​​obtained when they are stored. These long-term memories may be activated again and again, thereby affecting the machine's decision-making and behavior for a long time. And those information related to motivation states with high activation values ​​(such as strong emotions, large gains or losses, etc.), because they are adjacent to motivation states with high activation values ​​in time, they may be reversely activated by the high activation values ​​of motivation symbols with high activation values, thereby increasing their own activation values ​​when stored, and thus increasing their own memory values ​​obtained when stored.

[0071] 5. Reconstruction of language information.

[0072] Language plays an important role in machine intelligence. Language is a set of symbols created by humans to better communicate experience. Each symbol represents some specific things, processes and scenes. When language is input, the related memories represented by the language are activated. These memories may contain information about the language itself, as well as memories about the use of the language (such as voice emphasis or text emphasis to emphasize key points, such as a tone of distrust or a mocking tone, etc.). These activated information constitute an activated information flow. In order to balance the context of language and current semantic recognition, the activation value of the activated information will decay over time. The decay parameters are related to the motivation and state of the machine (such as needs and need states, emotions and emotional states).

[0073] The chain activation of language realizes the contextual association recognition of all input information. The input information here includes both environmental information and activated memory information. The mutual activation and assignment of these information reflects the contextual association. This association is more extensive than the content of the statistically generated semantic library. It not only involves language, but also all sensory inputs and related memories. Therefore, through the multi-resolution information formed by S102 and the common sense network formed by S103, the machine can realize the connection between language and static and dynamic images, feelings, needs and emotions, and also realize the connection between language and related language and memory. When this connection is incorporated into the machine's understanding of language input, and responds based on the understanding of language and relevant experience, it reflects that the machine truly understands the true meaning of the input language.

[0074] Language input constitutes an input information flow, and the corresponding activated memory also constitutes an activated information flow. When the machine understands language, it needs to reconstruct this activated information flow. Reconstruction is to reconstruct the environmental information in it, by overlapping the same parts of the activated environmental related information (such as images, sounds, and feelings), to form an imaginary process.

[0075] Similarly, the machine needs to integrate the information about feelings, emotions, vision, movements, body states, etc. activated by the input language with the machine's own existence. This information usually activates similar experiences of the machine itself. Therefore, the machine can experience the feelings, vision, emotions, movements, or body states brought by these languages.

[0076] 6. The emergence of machine prediction capabilities

[0077] The essence of prediction is a statistical behavior. Machine prediction is to infer the various possibilities and corresponding probabilities of the development of things, or the various possibilities and corresponding probabilities of other people's behaviors based on past experience or similar experience.

[0078] When information is input, the machine does not need to exhaustively predict all possible outcomes, which is an impossible task. The machine only needs to evaluate the possible outcomes of those activated experiences related to the input information. This is equivalent to using common sense to limit the range of possible outcomes. Within this limited range, the machine can use any current artificial intelligence prediction method, such as Monte Carlo search, decision tree, Bayesian estimation, rule-based machine reasoning methods, etc., to infer the probability of the current development of things to every possible extent.

[0079] In past experience, each possible outcome is linked to the demand state and emotional state through a relationship network, and these states represent the "benefits" and "harms" that the outcome brings to the machine. Therefore, by combining the possible probability with the relevant demand and emotional state, the machine can infer the size, type, and probability of the "benefits" and "harms" that each outcome of the development of things may bring to itself. Therefore, when the machine predicts the "benefits" and "harms" that the possible outcomes will bring to itself, it needs to take the probability of occurrence into account at the same time. Therefore, the machine must evaluate both the "benefits" and "harms" and their corresponding probabilities, and make decisions based on the two.

[0080] 7. Machine decision-making and response system

[0081] After the information is input, the machine limits the relevant decision-making scope according to the activation state of the memory through the associative chain activation method. Therefore, the machine's evaluation and response to the input information are based on the input information and the scope limited by the activated memory to search, evaluate and respond. This is equivalent to using common sense to limit the scope that needs to be searched, evaluated and responded.

[0082] Within this range, there may be one or more memories related to the input information. The machine can imitate these past experiences through segmented imitation to establish a possible response process. Then, within this limited range, the machine can use any current artificial intelligence prediction method, such as Monte Carlo search, decision tree, Bayesian estimation, rule-based machine reasoning methods, etc., to choose its own decision and response based on the predicted probability of "benefit" and "harm".

[0083] Because the purpose of the machine is to "seek benefits and avoid harm", the basic starting point of the machine's response to input information is to make its own response based on past experience, so as to increase the probability of things that produce "benefits", especially those that can obtain high profit values. And reduce the probability of things that produce "harm", especially those that can bring huge losses. Therefore, driven by the motivation of weighing the pros and cons, the machine combines its own responses based on experience to achieve the goal of "seeking benefits and avoiding harm".

[0084] The decision-making of the machine is a path planning method based on the machine's predictive ability. The purpose of the path is to maximize benefits and minimize losses. With predictive ability, the machine transforms a completely open question such as decision-making and response into a series of relatively closed questions on how to increase or decrease the probability of things happening within a certain range. Since common sense has been established in the previous steps, when each event occurs (this is the effect in the causal relationship), the related conditions (this is the cause in the causal relationship) can be obtained through the relationship network. Those causal relationships with strong correlations have strong connections in the relationship network because they occur again and again. Therefore, the relationship network can express the causal relationship layer by layer.

[0085] The goal of every decision is to make things develop in a direction that "seeks benefits and avoids harm". This may be a process of interaction with the outside world. And interaction itself is a way to promote the development direction of things to "seek benefits and avoid harm" based on past experience. Through the information and behaviors obtained through interaction, the probability of events with high profit values ​​is continuously increased, and the probability of events with high loss values ​​is continuously reduced. This is an iterative process. But the processing method is the same at each step. Based on the causal chain, the machine gradually increases the probability of events that lead to high profit values. This is similar to a chain activation process, which activates events on the path to high profit step by step, and carefully avoids events that may lead to high loss values.

[0086] Since the probability of causal connections between paths is expressed by the relationship network, the response planning problem of the entire machine becomes the problem of finding the optimal path in the causal chain network, which is exactly the problem that current machine intelligence has solved. For example, the machine can determine the prior probability of an event (such as an event that brings high profit value or loss value) by searching memory. Then the causal strength (posterior probability) between a certain condition and the event can be determined through the relationship network. The connection strength between different conditions in the relationship network can reflect whether the different conditions are independent. The machine only needs to select some relatively independent conditions and predict the probability of the event through the naive Bayes algorithm. The machine can decide its response based on the calculated probability. These responses can take various forms, such as: increasing the probability of this event, or reducing the probability of this event, or not affecting the probability of this event. This depends on the profit and loss value brought to the machine by this event. Increasing or reducing the probability of this event can be further planned as increasing or reducing the probability of the condition related to the probability of this event. This process is essentially an iterative probability path search problem.

[0087] For example, if the machine's response is to further determine the possible gains and losses. First, the machine uses the probabilities of various possible results in memory under similar circumstances to the current situation as prior probabilities. Then, the probability of occurrence of various gain values ​​and loss values ​​is calculated based on the posterior probability between the conditions related to each result and the result. Then, the machine generates the next goal to further determine the probability of occurrence of each condition. For example, the machine's response at this time can be (a) searching and counting the posterior probability between each condition and the corresponding gain value and loss value. Then it is used to update the overall gain and loss assessment. This can be done by searching the connection strength between the relationship network. (b) Further update the probability of occurrence of a certain condition. For example, based on imitating past experience, directly ask the information source about the probability of whether a certain condition has occurred or may occur. Or obtain the probability of whether a certain condition has occurred or may occur through other means. This depends on the behavioral imitation memory obtained in the machine learning process. (c) According to the principle of seeking benefits and avoiding harm, certain conditions closely related to gains and losses are promoted or avoided as new goals. Driven by the new goal, the same evaluation process is used to respond. Through such iterative responses, the ultimate goal remains to gain gains and avoid losses.

[0088] Therefore, when information is input, we can transform the seemingly completely open machine's response to information input into a multi-level goal through the causal relationship determined by the relationship network, the principle of seeking benefits and avoiding harm, and the relationship between events and "benefits" and "harms" established by the machine in the relationship network. These goals are all for increasing the probability of certain events, or reducing the probability of certain events. Therefore, through the causal relationship of the relationship network, the machine can transform the goal of seeking benefits and avoiding harm into a series of interrelated goals in specific situations. These goals constitute the path for the machine to maximize benefits and minimize losses.

[0089] In this process, after the machine responds, it may continuously search for new information, or continuously passively obtain new information, and use the posterior probability between the new information and the result to update the target path. The prediction of possible external feedback after the machine responds to itself also includes the activation of two types of motivational state memories. One is the demand and emotional state of the machine in the recurrent memory, which comes from the various feelings and emotions about itself in the activated memory. The other is the demand and emotional state of the machine when observing similar situations from the perspective of an observer, which comes from the various feelings and emotions generated by the machine when observing others in similar situations in the activated memory. Therefore, when the machine predicts "benefits" and "harms", it simultaneously evaluates the "benefits" and "harms" of an event to itself from both its own perspective and the perspective of others.

[0090] The predictive ability of a machine does not only include predicting the "benefits" and "harms" that an event may bring. It also needs to predict the responses that one or others may take under the drive of "benefits" and "harms", as well as the impact of others' responses on one's own "benefits" and "harms". These are obtained by statistically analyzing motivational state values ​​such as related needs and emotional states in the relationship network. Therefore, the evaluation results of the machine change dynamically with more input information. The decision-making and response process of the machine is a dynamic path planning process. It is jointly driven by experience responses and probability calculations based on gains and losses.

[0091] Through the above method, the machine can decompose an abstract goal of seeking benefits and avoiding harm into a large number of tasks that increase or reduce the probability of certain specific events under specific input conditions through iterative decomposition. These tasks can be subdivided into very specific target tasks, such as decomposing them all the way to the bottom-level driving capabilities of the machine. This process is the decision-making and response system of the machine.

[0092] 8. Imitation ability

[0093] Imitation ability is a human gene. For example, for a babbling child, if we greet him / her and say "you are back" every time he / she comes home, after a few times, he / she will take the initiative to say "you are back" when he / she comes home again. This shows that he / she has begun to imitate others to learn without understanding the meaning of the information. Similarly, we let machine learning adopt the same method. Therefore, the machine needs to put imitation into the machine as a bottom-level motivation. Make the machine willing to imitate the behavior of others (other machines), and continuously improve according to its own evaluation or external feedback information, so as to continuously exercise its ability to coordinate various senses, limbs, language and actions, thereby improving learning efficiency. At different stages of machine learning, we can give the machine different strengths of imitation motivation. For example, when the machine learns language and action output, we can directly give the machine a strong imitation motivation, while at other stages, we can give it a normal imitation motivation.

[0094] When the machine receives voice or action input from the outside world, these languages ​​or actions will activate the machine's own related memories. These memories may be a similar pronunciation, or a basic action fragment. These memories will further activate sensory information, demand and emotional information, language or action memory related to these memories. Driven by the motivation of imitation, the machine will make similar voice output or action output based on these activated memories and adjust the underlying driving parameters in the experience through the decision-making system. The underlying drive refers to the underlying experience of voice output or the underlying experience of action output. They are muscle drive commands corresponding to specific voices or actions, in which the parameters are learned and continuously updated through feedback.

[0095] Humans can preset some of the most basic speech or action (including facial expressions and body language) capabilities for machines. Their parameter optimization can be achieved through subsequent learning and training, so that these parameters and the results of the behavior can be associated through memory, and continuously adjusted through the emotional and demand systems (affected by self or external feedback). Ultimately, driven by the underlying motivation, the machine obtains the relationship between different parameters under different external information incentives through the memory and forgetting mechanism, forming memory. These memories are the knowledge and skills of the machine when facing external information input. They include behavioral habits such as language, action, facial expression, and body movements.

[0096] Humans can also give machines preset conditioned reflex systems. The function of these systems is to respond to specific input situations as humans expect the machine to respond. For example, the machine's evasive action in a critical situation, or the machine's specific output action under specific information input (for example, these conditioned reflex systems can be used for machine self-checking, emergency shutdown, or adjustment of the machine's working status, etc.).

[0097] 9. Execution process

[0098] With the above basic capabilities, the machine can respond according to its own decisions, such as language output, action output (including expression and body language output) or other forms of output (such as output data stream, image, etc.). The response execution step is a process of translating the plan into actual output.

[0099] If the machine chooses voice output among various possible response steps, this is relatively simple. It only needs to convert the image feature map to be output into voice, and then use the relationship between languages ​​in the relational network (grammatical knowledge existing in the relational network) to organize it into a language output sequence, and call on pronunciation experience to implement it.

[0100] It should be pointed out that the machine may choose some dynamic features to express the entire sentence based on experience (its own or others' experience) (such as using different movement patterns of tone, audio pitch or stress changes to express doubt, ridicule, distrust, emphasis, and other common human ways. These methods are usually low-resolution features of a sentence or an entire speech). Because the machine learns these expressions from human life, in theory, any human expression can be learned by the machine.

[0101] If the machine chooses to output actions, or a combination of voice and actions, the problem becomes much more complicated. This is equivalent to organizing an event. The machine's response plan may only contain the main steps and the final goal, and the rest needs to be improvised in practice.

[0102] The machine needs to respond to the sequence targets to be output, and divide them in time and space according to the different time and space involved in these targets, so as to coordinate its execution efficiency. The method adopted is to select targets that are closely connected in time and targets that are closely connected in space as groups. Because the information combination formed by the combination of dynamic feature graphs and static feature graphs has the relevant memory environment space with time and space information, this step can adopt the classification method. This step is equivalent to rewriting from the general script to the sub-script.

[0103] The machine needs to combine the intermediate goals in each link with the real environment again, and use the segmented imitation method to expand them layer by layer. The response plan proposed by the machine at the top level is usually composed of highly generalized process features and highly generalized static concepts (because these highly generalized processes can find multiple similar memories, so the responses established by borrowing from them are also highly generalized). For example, under the total output response of "business trip", "going to the airport" is an intermediate link goal. But this goal is still very abstract, and the machine cannot perform imitation.

[0104] Therefore, the machine needs to divide according to time and space, and take the links that need to be executed in the current time and space as the current goal. And take the goals of other time and space as the inherited goals and put them aside temporarily. After the machine takes the intermediate link as the goal, the machine still needs to further subdivide time and space (write the hierarchical script again). This is a process of increasing time and space resolution. The process of the machine converting a goal into multiple intermediate link goals is still to use decision-making ability, analyze various possible results and possible probabilities, and choose its own response according to the principle of "seeking benefits and avoiding harm". The above process is constantly iterating, and the process of dividing each goal into multiple intermediate goals is a completely similar processing flow. It must be decomposed until the bottom experience of the machine. The bottom experience for language is to mobilize muscles to produce syllables. For actions, it is to decompose to issue driving commands to related "muscles". This is a tower-shaped decomposition structure. The machine starts from the top-level goal and decomposes a goal into multiple intermediate link goals. This process is to create virtual intermediate process goals. If these intermediate process goals "meet the requirements", they will be retained. If they "do not meet the requirements", they will be recreated. This process unfolds layer by layer, and finally establishes a rich and colorful response of the machine.

[0105] In this process, the machine may encounter new information at any time, causing the machine to process various information, and these original goals become inherited motivations. This is equivalent to constantly encountering new situations in the process of organizing activities, which need to be solved immediately, otherwise the activity cannot be organized. So the director stops other activities and solves the problem at hand first. After solving it, the activity continues. Another situation is that in this process, the director suddenly receives a new task, so after weighing the pros and cons, the director decides to suspend the activity first and give priority to the new task.

[0106] The machine is performing the imitation task that can be performed while breaking down other goals into more detailed goals. So the machine is thinking while doing. This is because the real situation varies greatly, and the machine cannot know the external situation in advance and make plans. So this is a process of achieving a goal through the interaction between the environment and the machine.

[0107] At this point, the machine can use the above capabilities to complete the understanding and response of input information. This process is a minimum cycle of interaction between the machine and the outside world. The machine repeats this process to achieve a greater goal, which is manifested as a continuous interaction process between the machine and the outside world, showing the intelligence of the machine.

[0108] 10. Machine experience summary

[0109] The experience of the machine is not only to form connections in the relationship network through memory and forgetting mechanisms, but also to actively strengthen such connections. This active strengthening of connections can be manifested in many forms: for example, learning the experience of others through language. The activation information flow formed by language, together with language, constitutes the learned experience of others. This experience is stored in the memory as new input information, and they are part of the memory. For another example, the machine actively turns these memories into long-term experience by repeatedly memorizing the information that is closely related to the relationship between "benefits" and "harms".

[0110] The machine can use a preset algorithm to store those memories that can appear repeatedly and have a greater impact on the relationship between "benefits" and "harms". By inputting the information of the corresponding events in sequence virtually, the machine can go through the virtual processing process again, strengthening the relevant connections in the relationship network, thereby enhancing the experience. In this enhancement process, the connection between the common parts of similar experiences may gradually strengthen, so the experience will become more and more concise and universal, and eventually form a certain rule for the machine to summarize itself. Therefore, the self-summary of the machine can be realized by using preset programs and adjustable parameters. These adjustable parameters can be adjusted according to the connection relationship between "benefits" and "harms" and specific things that the machine gradually summarizes during learning, or they can be adjusted according to the intensity and frequency of the machine being stimulated.

[0111] Figure 1 The S105 module is the communication connection module of the machine. The machine can exchange data with other machines (including computers) in the outside world according to the preset protocol through the communication module. Through data exchange, the machine can realize distributed computing. For example, the machine on site may process part of the data, and may also transmit part or all of the information to the central brain, and make decisions with the help of the central brain's powerful information storage and processing capabilities. Through data exchange, the machine can also share memory. For example, the experience gained by a machine can be passed on to other machines, or the memory of multiple machines can be integrated to form a more complete experience. Machines can also share computing power, thereby forming distributed thinking ability. Through the communication connection module of the machine, machines can realize cognitive sharing, decision sharing and behavior coordination capabilities, thereby realizing super artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0112] Figure 1 A schematic diagram of one possible component of a machine. DETAILED DESCRIPTION

[0113] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the text of this application mainly proposes the main steps for realizing general artificial intelligence. Each of these main steps can be implemented using currently known structures and technologies. Therefore, the focus of this application is on these steps and their composition, rather than being limited to the details of each step using known technologies. Therefore, these embodiment descriptions are only exemplary and are not intended to limit the scope of this application. In the following description, in order to avoid unnecessary confusion of the focus of this application, we have omitted the description of known structures and technologies. All other embodiments obtained by those skilled in the art without making creative work should fall within the scope of protection of this application.

[0114] Sensor system establishment: Figure 1 In the S101 module, the sensor module of the machine. In order to make the machine produce a similar cognitive method as humans, the S101 module needs to use one or more general sensors: visual sensors, auditory sensors, taste and olfactory sensors, tactile sensors, gravity direction sensors and posture information sensors, etc. Sensors for specific applications can also be added (for example, autonomous driving can add infrared sensors, lidar sensors, etc.). The machine also needs to use sensors to monitor its own status. These sensors are also part of the machine's perception of information. S101 is a module mainly composed of sensor hardware and sensor corresponding software. Its purpose is to perceive information outside the machine and the machine itself through sensors. The data generated by the machine's sensors is input into the machine's processing unit and is called input data.

[0115] The communication between the sensor data of the machine and the processing unit of the machine may adopt any existing communication system or a customized communication system. These specific communication forms do not affect the implementation of this patent application.

[0116] Establishment of preset motivation system: The machine uses symbols to represent a type of motivation. For example, language symbols, or some sensory information data are directly used, or a symbol is directly artificially defined to represent a type of motivation of the machine. For example, the "Sa" symbol is used to represent the safety needs of the machine, and a value from 0 to 100 is used to represent the state of the machine's safety needs. Among them, 0 represents a very insecure feeling, and 100 represents a completely secure feeling. For example, the "danger" symbol is used to represent danger, and H, HM, M, ML, and L are used to represent the degree of danger from high to low. For example, a "smiley face" is used to represent a pleasant emotion, and 0 to 10 are used to represent the degree of pleasure. In the present invention application, how many motivation types are established, what form of symbols are used to represent the motivation types, and what form is used to express the state of the symbols, does not affect the claims of the present invention application. Because in the present invention application, the processing methods of all symbols and corresponding states are completely similar. For example, we can use different symbols to establish different emotions such as excitement, anger, sadness, tension, anxiety, embarrassment, boredom, calmness, confusion, disgust, pain, jealousy, fear, happiness, romance, sadness, sympathy and satisfaction for the machine. Each emotion has its own quantitative space. The motivation of the machine to seek benefits and avoid harm is specifically to weigh the pros and cons in the space established by various motivational needs and find an acceptable range of space. Then, based on experience, it pushes events to develop in a space that it can accept. This push strategy may be to directly promote the direct goal, or it may be a prerequisite for promoting the realization of the direct goal, thereby increasing the probability of the direct goal occurring.

[0117] Multi-resolution information feature extraction system: Multi-resolution information feature extraction can use the current artificial intelligence perception processing capabilities to extract the features of input information at different resolutions. The difference from the current artificial intelligence perception processing method is that the current algorithm realizes the mapping from data space to label space. In the present invention application, the goal of S102 module is to extract local common features in the input data.

[0118] The existing artificial intelligence perception processing method is to optimize the mapping network by mapping a large amount of data to labels and optimizing the error function. This mapping is from data to specific labels, and the optimized mapping relationship is closely related to the specific labels, so this algorithm is not universal. In the present invention application, all input data are directly mapped to local common features. And the local common features are basic feature libraries established through pre-training. The basic assumption for establishing these basic feature libraries is that the evolution of human beings is developing in the direction of efficient use of computing power. Because only in this way can we maximize energy savings while increasing the complexity of the algorithm to deal with complex environments, thereby increasing the chance of survival. The innate extraction algorithm for widely existing local features is a concrete manifestation of this evolutionary direction. Because in this way, these algorithms can be reused as much as possible to improve the energy efficiency of computing.

[0119] The reason for using multiple resolutions is that "similarity" is a vague concept. Only when a specific resolution is established can the degree of "similarity" between two things be explained. In the evolution of human beings, humans have concluded that there may be other similarities between some similar things on our planet. So we need to infer other possible similarities from known similarities. For example, inferring the possible behavior of a dog from its appearance. So multi-resolution is to establish the similarity between things at different resolutions. For example, at a coarse resolution, all dogs are similar. But at a further resolution, we will think that there are differences between different dogs. So through coarse resolution, we can generalize some common behaviors of dogs. For example, when we see a strange dog, we can activate related memories through the information features of the input dog. First, the information about the dog at a coarse resolution in all memories will be activated, and this coarse information will further activate other features of the dog. Then the more detailed features of the input dog information will activate the memories related to these detailed features. The final activation accumulation state is that the public information about the dog and the information about the detailed features of the dog are activated, thereby obtaining a preliminary information evaluation of the dog. Other detailed features about dogs that do not match the detailed features of this dog will not be activated, so this information will not interfere with cognition. Similarly, when we see an animal that is similar to a dog, from a coarse resolution, we will activate information about life, information about animals, common information about dogs, and information related to the characteristics exhibited by this particular animal. For example, if its claws are very similar to cat claws, then we may activate the prediction that it may scratch us. Therefore, the extraction of multi-resolution features, especially the extraction of multi-resolution action features, is the key to knowledge generalization. Because multi-resolution actions are widely present in different things. Therefore, multi-resolution actions are a widely existing bridge for knowledge generalization.

[0120] The specific method of the multi-resolution extraction method is to compress the data at different resolutions, and then use data extraction windows of different sizes to repeatedly find local similarities. In the patent invention application with application number 202010400739.7, entitled "A method for imitating human memory to achieve general machine intelligence", there is a specific method on how to implement the multi-resolution extraction method, which will not be repeated here.

[0121] So although we can use any specific feature extraction algorithm currently available, such as data coordinate basis transformation and filtering processing, such as convolutional neural networks, or various forms of time-delay neural networks and other specific algorithms to achieve multi-resolution feature extraction. However, in the present invention application, these algorithms are one of the steps to achieve multi-resolution feature extraction. Multi-resolution feature extraction and existing artificial intelligence are different in algorithmic goals, outputs, and the amount of data required is also different. Multi-resolution feature extraction does not require a large number of samples because it is a mapping from features to concepts, and usually a concept contains a limited number of features.

[0122] Associative chain activation system: Associative activation refers to the chain activation process under the principles of "proximity activation", "similar activation" and "strong memory activation". When information is input, the machine assigns its initial activation value according to the motivation, and performs chain activation through the principles of "proximity activation", "similar activation" and "strong memory activation". In the relational network, when a node (i) is assigned a certain activation value, if this value is greater than its preset activation threshold Va(i), then the node (i) will be activated. It will pass the activation value to other feature graph nodes that are connected to it. The transfer coefficient is determined according to the principles of "proximity activation", "similar activation" and "strong memory activation". If a node receives the transmitted activation value and accumulates its own initial activation value, and the total activation value is greater than the preset activation threshold of its own node, then it will also be activated and will also pass the activation value to other feature graphs that are connected to it. This activation process is chained until no new activation occurs and the entire activation value transfer process stops. This process is called the associative activation process.

[0123] The associative activation process is a search method for related memories, so it can be replaced by other search or lookup methods that can achieve similar functions. In the patent invention application with application number 202010400739.7, entitled "A method for imitating human memory to achieve general machine intelligence", a specific method on how to implement the associative chain activation system is proposed, which will not be repeated here.

[0124] Establishment of preset basic response system: The preset basic response system is a machine response system realized by preset programs. These systems are the instinctive reactions of the machine to respond. These instinctive reactions can be gradually optimized in acquired learning.

[0125] 1. Basic action system. Basic actions include giving machines programs that can imitate basic human actions, including language pronunciation, facial expressions, and body movements.

[0126] 2. Preset instinctive response. Instinctive response refers to the realization of the output response of the machine under specific input through preset programs. For example, the machine's avoidance response to high temperature, instinctive avoidance response to falling, avoidance response to sudden impact, etc. can be realized through preset programs. These responses are adjusted through experience in subsequent learning. For example, through a posteriori results, it is learned that under certain specific information input conditions, instinctive responses may cause serious losses. Therefore, when the input conditions that stimulate instinctive responses are accompanied by specific information that may cause serious losses due to instinctive responses, the machine can suppress instinctive responses under the drive of seeking benefits and avoiding harm, and seek a more optimized balance of gains and losses.

[0127] There are not many emergency situations that require instinctive responses, so they can be implemented using preset programs. However, the environment that triggers instinctive responses may be complex, so machines need to learn and build relationship networks through a posteriori experience, and respond differently to different situations under the control of prediction, planning and execution systems.

[0128] 3. Initial activation value assignment system. Figure 1 In step S102, after the machine extracts multi-resolution features, it uses the initial activation value assignment program to assign initial activation values ​​to the input information. The initial activation value assignment program is a preset program, and its input parameters include input information, the motivation and motivational state of the machine at this time, such as needs and need states, emotions and emotional states. Its output includes the initial activation value assigned to the input information. A simple method is to directly assign equal activation values ​​to all input information based on the machine's expected information, needs, emotions and other motivational states at this time. Another method is to make a simple classification based on the machine's expected information, and use different initial activation values ​​for different expected information. It is also possible to use a method of assigning different initial activation values ​​to information of different resolutions. For example, a larger activation value is assigned to low-resolution information, and a lower activation value is assigned to high-resolution information. It can also be assigned in reverse, depending on the machine's expectations of the input information. The machine's expectations of the input information come from the previous information processing results.

[0129] The machine can also adjust the initial activation value assigned to the input information according to the frequency of the input information. For example, for those frequent stimuli, the initial activation value assigned to them gradually decreases. For these information with low initial activation values, the activation value obtained by other information during the chain activation process is also low. When a new memory is stored, the initial memory value obtained is positively correlated with its activation value when the memory occurs. Therefore, those commonplace things in daily life, because they have low initial activation values, the activation values ​​given to other information by the chain activation they initiate are also low, and the memory value obtained when they are stored is also low. Memory storage is first placed in a temporary memory bank. Only information that has obtained enough memory value in the temporary memory bank will be moved to other memory banks. Therefore, those commonplace things in daily life are difficult to survive in the temporary memory bank and form long-term memory. This is also a result of the evolution of intelligence, because intelligent bodies need those general rules and special exceptions to cope with the external environment. The general rules are summarized through repetition. Once the summary is completed, there is no need to remember similar situations over and over again. However, those exceptional or occasional strong stimuli may bring about a strong chain activation process due to the high initial activation value, so that the single memory value reaches the long-term memory threshold and becomes the long-term memory of the machine.

[0130] The initial activation value of the machine can also be assigned multiple times. For example, the aforementioned method is used to simply assign an initial activation value to the input information. Then, based on the preliminary processing results of the input information, the degree of its connection with "benefits" and "harms" is preliminarily determined, and then according to the degree of its connection with "benefits" and "harms", the initial activation value is assigned again according to the preset procedure. For example, for information that may bring great "benefits" and "harms", the machine's strategy may be to further analyze the information. Further analysis means extracting the input information with more resolutions again, and assigning a higher initial activation value to the multi-resolution information features again, so that the entire activation range is larger, thereby searching for more related memories. The above process can also be iterated, resulting in multiple initial activation value assignments.

[0131] The above basic action system, preset instinctive response and initial activation value assignment system can be completed using currently mature computer programs. These specific implementation methods can be realized through the common knowledge of the industry. The patent of this invention mainly discloses the methods and ways to realize general machine intelligence using existing technologies, and the specific implementation details of specific steps that can be realized by technicians in the industry based on the common knowledge in the existing industry do not need to be further explained in the application of this invention.

[0132] Environmental reconstruction: Environmental information is part of the machine's input information. This part of information will activate similar environmental information in the machine's memory. The environmental information in this memory may be memories of other parts or other angles of the same environment, or memories of similar environments. The activated information may include visual information, auditory information, tactile information, motivation, and motivation state information.

[0133] The machine needs to adopt (1) overlapping similar parts of the environment to construct a 3D environment. (2) Using the common structural features in similar environments as a model to predict missing information. Through steps (1) and (2), the environment assembled by the machine is the current environment constructed by integrating the currently input environmental information and memory-related environmental information. This is the machine's reconstruction of the environment. The machine's reconstruction of the environment is essentially the machine's prediction of the environment. After the environmental space is established, we can understand the parts of the real space that are currently invisible based on other parts of the borrowed memory space. For example, when we look at a familiar cabinet, it seems that we can see the image inside the cabinet. This is actually because we have superimposed the memory image inside the cabinet. At the same time, in the process of reconstructing these environments, the machine's corresponding auditory memories, tactile memories, motivations, and motivational states (such as memories of various emotions and emotional states) may be triggered through associative activation, so the environment reconstructed by the machine is an environment with subjective emotions.

[0134] The size of the machine's prediction range of the environment is related to the machine's expected goal. Generally, when the expected goal is large, the prediction range of the environment is also large. When the resolution of the expected goal is high, the prediction resolution of the environment is also high. The machine's expected goal is the goal generated by the machine in the previous prediction, decision-making and execution process.

[0135] There are many algorithms that reconstruct the whole from the local, such as GAN neural network, and 3D reconstruction technology commonly used in games. The machine can use these existing technologies to combine the input multi-resolution environmental information and the activated multi-resolution feature information in the memory for 3D reconstruction. Technical personnel in the industry can achieve these goals based on the existing industry common knowledge, and these technologies are not within the scope of the claims of the present invention application, so no further explanation is needed here.

[0136] Self-reconstruction: Similar to the method of environmental reconstruction, the machine also uses the same method to reconstruct its self-image. In reconstructing its self-image, in addition to using the current input information, the machine also needs to integrate the relevant information in the memory activated by the current input information: such as vision and hearing, touch, smell, taste and feeling, gravity feeling, limb state feeling, motivation and motivation state (such as emotions and emotional states, needs and need states, etc.). The machine needs to use (1) overlapping similar parts of this information to construct a 3D image. (2) Using the common structural features in the memory of itself, such as a human model or a typical image of oneself, as a model to organize this information and supplement the missing information. Through steps (1) and (2), the whole information combined by the machine is the image and feeling of the machine that it has established about itself. For example, when we put our hands behind our backs and make movements, we seem to be able to see these movements. This is because after we issue neural commands and obtain tactile perception, we activate the vision connected to similar neural commands in memory, the vision and touch connected to similar body posture perception information, and the vision connected to similar tactile perception information. These information are integrated into our overall image by overlapping similar parts. We create a mirror image of ourselves in our minds, and we seem to be able to see the movements of our mirror image.

[0137] The machine first has a specific concept of "self" through the preset software and hardware system: limbs or vocal organs that can be driven under its own command, sensory organs that can send various information to itself, and visual and auditory information that is always together with its various senses. Since these information always appear at the same time, the relationship between them is very close in the relationship network, thus forming the concept of "self". Therefore, the concept of "self" is formed in the same way as other concepts, and it is not mysterious. For example, if a person can feel pain when a table is hit, then he will definitely regard this table as part of his "self".

[0138] With the concept of "self" in a narrow sense, the machine gradually obtains the relationship between various "benefits" and "harms" and the "self" through learning. These "benefits" and "harms" are closely connected with the underlying motivations of the machine (such as needs and emotions). Therefore, driven by the motivation of "seeking benefits and avoiding harm", the behavior pattern of the machine may be to "possess" those things that bring "benefits" to itself, and "avoid" those things that bring "harm" to itself. Therefore, only after the concept of "self" can there be concepts such as "possession" and "avoidance". Because the concepts of "possession" and "avoidance" are extended under the principle of maximizing benefits and minimizing losses. With the concepts of "possession" and "avoidance", the machine can understand the organization, laws, behaviors and morals of our society. Because the core content of the organizational form of our society is the various expressions of "possession" and "avoidance".

[0139] Therefore, the self-awareness of a machine is essentially a behavior of the machine. The machine does not need to add mysterious self-awareness to itself. It is a behavior of the machine that determines its interaction with the outside world in a way of "seeking benefits and avoiding harm" after learning various information and its own "benefit" and "harm" relationships through relationship networks and associative activation.

[0140] The focus of the present invention patent application is to innovatively reveal the way to achieve strong artificial intelligence, and the focus is on the proposed implementation path. The specific software and algorithms required for the self-reconstruction of the machine can be achieved by technical personnel in the industry based on the existing industry common knowledge to achieve the sub-goal of achieving self-awareness proposed in the present invention patent application, so no further explanation is needed here.

[0141] Language reconstruction: Language plays an important role in machine intelligence. Language is a set of symbols created by humans to better communicate experience. Each symbol represents some specific things, processes and scenarios. When the machine receives language information from the outside world, the S102 module performs multi-resolution feature extraction on the input language (such as voice or text). For example, the overall characteristics of the voice are obtained at low resolution, such as rising and falling tones, stress, tone, speed, voice size and changes, voice intonation frequency and changes, etc.; specific vocabulary is obtained at medium resolution; specific syllable pronunciation is obtained at high resolution.

[0142] Through the S103 module, these features will activate related memories in memory through associative activation. Because language is used very frequently, it is usually closely associated with typical images, dynamic images, feelings, sounds and other sensors of the things, processes and scenes they represent. These closely connected local networks are part of the relationship network, which is the concept.

[0143] When language is input, the related memories represented by the language are activated. These memories may contain information about the language itself, as well as memories about the use of the language (such as voice emphasis or text emphasis to emphasize key points, such as a tone of distrust or a mocking tone, etc.). These activated information constitute an activated information flow. In order to balance the contextual association of language and current semantic recognition, the activation value of the activated information will decay over time. The decay parameters are related to the motivation and state of the machine (such as needs and need states, emotions and emotional states). Chain activation realizes the contextual association recognition of all input information. The input information here includes both environmental information and activated memory information. The mutual activation assignment of these information reflects the contextual association. This association is more extensive than the content of the semantic library generated by statistics. It not only involves language, but also all sensory inputs and related memories. Therefore, through the multi-resolution information formed by S102 and the common sense network formed by S103, the machine can realize the connection between language and static and dynamic images, feelings, needs and emotions, and also realize the connection between language and related language and memory. When this connection is incorporated into the machine's understanding of language input and it responds based on its understanding of the language and relevant experience, it means that the machine truly understands the true meaning of the input language.

[0144] Language input constitutes an input information flow, and the corresponding activated memory also constitutes an activated information flow. When the machine understands language, it needs to reconstruct this activated information flow. Reconstruction is to reconstruct the environment information and reconstruct the self-involved information. The machine views the reconstructed activated information flow from a third-party perspective. The reconstructed information flow is a new information input, and the machine can store this new information flow back into the memory as part of the memory. Therefore, what the machine learns through language is a virtual experience. It is a virtual scene viewed from a third-party perspective. The reconstructed self-image is a representation of self-consciousness, and the reconstructed environment is virtual environmental information. The machine can learn cognition from these virtual experiences, and learn the connection between various information and "benefits" and "harms". This virtual experience will also activate relevant information in the memory, and they will also exist in the memory. Together with other memories, as part of the relationship network, they are also optimized by the memory and forgetting mechanisms.

[0145] Language reconstruction is the basis for machine learning and language understanding, and it is also the basis for machine learning to learn all the experience accumulated in human history. With the ability to learn language, machines can learn all human knowledge, which is the basic ability to lead to super intelligence.

[0146] Language is a way to convey other people's experience. Machines can not only convey experience through language, but also directly share cognition through relationship networks, so the learning process of machines can be very rapid. Humans must use preset rules to avoid the unfavorable human behavior caused by various cognitive combinations of machines. Since the machine intelligence technology involved in the present invention application is visible to humans, the machine's activation process, the machine's prediction results, the machine's decision-making algorithm and decision results are all directly readable and understandable to humans. This is different from the current deep learning. In current deep learning, the decision-making process is difficult to explain because the number of parameters involved in decision-making is too large and the meaning is not clear. In the present invention application, the machine is established from perception to cognition through shared local features, memory and forgetting mechanisms. The number of parameters involved in decision-making is much smaller than the current multi-layer neural network, and each step of the decision-making is understandable, so the artificial intelligence proposed in the present invention application is an explainable and controllable artificial intelligence.

[0147] Machine prediction, decision-making and response process: The machine prediction, decision-making and response process is based on the relationship network, limiting the search scope through associative activation, and using past experience (including the causal probability between events, including the "benefits" and "harms" brought by events) to use statistical methods to calculate the size and probability of "benefits" and "harms". Then, under the predetermined algorithm, the principle of seeking benefits and avoiding harm is used to increase or reduce the probability of a certain event. This is the entire process of machine prediction, decision-making and response.

[0148] When information is input, the machine does not need to exhaustively predict all possible outcomes, which is an impossible task. The machine only needs to evaluate the possible outcomes that are activated and related to the input information. This is equivalent to using common sense to limit the search scope.

[0149] In past experience, each possible outcome is connected to motivational states such as demand states and emotional states through a relationship network, which represent the "benefits" and "harms" that the outcome brings to the machine. Therefore, by combining the probabilities of possible occurrence with the relevant needs and emotional states, the machine can infer the size, type, and probability of the "benefits" and "harms" that each outcome of the development of things may bring to itself. Therefore, when the machine predicts the "benefits" and "harms" that the possible outcomes will bring to itself, it needs to take the probabilities of possible occurrence into account at the same time. Therefore, the machine must evaluate both the "benefits" and "harms" and their corresponding probabilities, and make decisions based on both.

[0150] For example, within a limited scope, the machine can use any current artificial intelligence prediction method, such as Monte Carlo search, decision tree, Bayesian estimation, rule-based machine reasoning methods, to infer the probability of the current event developing into the most likely outcome of the next step, as well as the "benefits" and "harms" this outcome will bring to itself.

[0151] One possible evaluation method is to use various types of needs and emotional states to establish a multidimensional space. In this space, the machine is preset with various preset spaces such as the best pursuit space, acceptable area, and unacceptable area. This is equivalent to establishing various calculation rules for the machine to seek benefits and avoid harm. However, due to the variety of needs and emotional types, such rules are difficult to express clearly, so the method of calculating spatial distance can be used. The strategy adopted by the machine can be to get as close to the best pursuit space as possible and stay away from the unacceptable space in terms of spatial distance. These rules are then quantified by calculating spatial distance.

[0152] Another simpler method is to calculate Y=∑ f i (x1,x2,…)* p1 i (x1,x2,…)-∑ G j (y1,y2,…) *p2 j (y1,y2,…). Where Y is the last evaluated value, f i (x1, x2, …) are various benefit values ​​(determined by demand, emotion type and state according to preset rules), p1 i (x1,x2,…) is the probability of occurrence of the corresponding return value. j (y1,y2,…) are various loss values, p2 j (y1, y2, ...) is the probability of the corresponding loss value. ∑ is the summation symbol, summing i and j respectively. The purpose of the machine is to maximize Y. This method is to first quantify the needs, emotional types and states into gain values ​​and loss values, and then simply calculate the probability sum.

[0153] The above evaluation of the possibility of a single step is a technical evaluation, which can be achieved by using existing machine intelligence methods and existing statistical algorithms, such as calculating the probability statistical average or the minimum spatial distance.

[0154] After evaluating the possibility of a single step, the machine also needs to predict and plan the "benefits" and "harms" of its subsequent responses and the responses of the environment. This is a multi-step "benefits" and "harms" prediction and evaluation process. This is a strategic planning, equivalent to finding the optimal path. The principle of selecting a path is to seek benefits and avoid harm. The selection methods can include existing decision-making algorithms such as decision trees, rule-based expert systems, Bayesian networks, and evolutionary algorithms. The basis of these algorithms is built on a relationship network with causal probabilities and "benefits" and "harms" relationships.

[0155] When the machine evaluates the "benefits" and "harms" of multiple steps, a simple evaluation method can be: take the response of each planned step as a virtual process and use it as input information again, go through all the information extraction, association activation, environment, self and language reconstruction processes again, and then use the same single-step evaluation method again to evaluate the possible "benefits" and "harms". This is a way to use one's own experience to evaluate the possible impact of one's own response. These impacts include the feedback from the outside world on one's own response (including the feedback from others in the environment), and the impact that these feedbacks may have on one's own "benefits" and "harms".

[0156] With this approach, the machine can use an iterative method to fully evaluate the input information and plan its response according to the principle of seeking benefits and avoiding harm. This planning may be repeated multiple times, and multiple iterations are essentially a process of expanding the search range and finding the optimal path. This is the essence of machine planning.

[0157] The optimal path search range of the machine during planning is determined by the possible "benefits" and "harms". Usually, if the possible "benefits" and "harms" are very high, the machine will expand the search range under the preset algorithm principle of seeking benefits and avoiding harms, such as lowering the activation threshold or increasing the number of evaluation iterations, so that a wider range of information is activated, and a wider range of experience is incorporated into the path prediction process to select the optimal path.

[0158] Therefore, the decision-making process of the machine is not mysterious. With the help of common sense, the decision-making process of the above machine is relatively simple to implement. Pre-set programs and existing machine reasoning methods can be used to calculate the probabilities of the main possible results, as well as the probabilities of various gains and losses. The methods used by machines to evaluate "benefits" and "harms" only involve probability and statistical methods, which are mature public knowledge. However, the methods used by machines to evaluate "benefits" and "harms" are not limited to the above two methods. The machine can use any existing statistical decision-making algorithm.

[0159] It should be pointed out that the decision-making process of the machine is based on the relationship network. The relationship network is closely related to the education the machine has received. Even the same learning materials and different learning orders will produce different relationship networks, thus forming different cognitive and decision-making processes of the machine. The decision-making ability of the machine is also affected by the machine's associative activation parameters, statistical algorithm parameters, underlying motivations and motivational states (including needs and need states, emotions and emotional states, etc.). Therefore, machines with different experiences and different settings may make different decisions.

[0160] The machine can also adjust the algorithm parameters of its decision-making based on the feedback of its own experience. For example, after the machine executes a series of responses according to its own decision, if the actual results are quite different from the expected results, then these memories will also enter the memory and become part of the relationship network. The next time the machine makes a decision, these memories will become new experiences, and these experiences will affect the machine's next decision-making behavior in a similar state.

[0161] The machine can also learn and summarize experience under the preset algorithm. The specific implementation method is: the preset algorithm can select the processes in which the machine obtains greater "benefits" and "harms". Let the memory of these processes be virtually input multiple times in the machine's processor, so that the relevant memories will be activated each time. The common features that exist in such processes will appear again and again, and activate the relevant memories again and again, enhancing their memory value. Therefore, these enhanced memories are easier to be activated by association and easier to be used as reference. However, those related memories that only exist in a single specific process will gradually be forgotten due to long-term difficulty in activation. This is the process of the machine actively summarizing experience. Of course, without such a preset program, the machine itself will also summarize these experiences in the same way, but the samples and time required will become longer.

[0162] Because the machine's decision-making system searches for the optimal path over multiple steps, the machine can make decisions that accept small losses in the hope of obtaining large gains later.

[0163] For example, the trainer warns the machine that hitting people will be "severely punished." The machine further activates relevant information by explaining the word "severely punished," and can understand that hitting people will bring huge losses. The cognition that "hitting people" may bring losses can also be the direct experience gained by the machine through direct hitting and feedback, or it can be the relevant experience preset for the machine, or it can be a mixed experience of multiple methods. However, when the machine faces the situation that "its owner is being attacked and is very dangerous", then the machine will have to weigh the "loss value" of hitting people and the huge "benefit value" of saving the owner's life. At this time, the machine needs to conduct multiple steps of search and expand the search range to determine the probability of bringing huge "benefit value" and whether it will bring other losses. Then make a comprehensive decision based on the probability of gain and loss.

[0164] For example, according to experience (or preset experience), if the owner's life is threatened, then the "benefit value" of saving the owner will be very high. But if the attacker is a "policeman", then you cannot intervene, otherwise you will suffer huge "losses" with a 100% probability. This knowledge can be acquired or directly preset experience.

[0165] The machine execution system is similar to the machine decision system, but the difference is that the machine execution system concretizes the goals generated by the machine decision system; the concretization process is also a kind of decision of the machine, which is to subdivide the specific goals layer by layer to the bottom-level driving command level of the machine until the machine can directly execute the instructions. When the machine is subdivided layer by layer, it also uses the machine's decision system.

[0166] After having the above basic capabilities, the machine can execute specific responses according to its own decisions, such as language output, action output (including expression and body language output) or other forms of output (such as output data stream, image, etc.). The execution response step is a process of translating the plan into actual output.

[0167] If the machine chooses voice output in the selection of various possible response steps, it is relatively simple. It only needs to convert the image feature map to be output into voice, and then use the relational network and memory to combine the dynamic feature map (including the concept of representing the relationship) and the static concept (grammatical knowledge existing in the relational network) by concept replacement, organize it into a language output sequence, and call the pronunciation experience to implement it. It should be pointed out that the machine may choose some dynamic features to express the whole sentence based on experience (its own or others' experience) (such as using different movement patterns of tone, audio pitch or stress changes to express doubts, ridicule, distrust, emphasis and other common human ways. These methods are usually low-resolution features of a sentence or a whole speech). Because the machine learns these expressions from human life, any human expression can theoretically be learned by the machine.

[0168] If the machine chooses to output actions, or a combination of voice and actions, the problem becomes much more complicated. This is equivalent to organizing an event. The machine's response plan may only contain the main steps and the final goal, and the rest needs to be improvised in practice.

[0169] The machine needs to divide the goals of the sequence (including intermediate goals and final goals) according to the different time and space involved in these goals, so as to coordinate its execution efficiency. This also comes from experience, and it can also be executed using a preset algorithm. The method adopted is to select goals that are closely related in time and goals that are closely related in space as groups. Because the information combination formed by the combination of dynamic feature maps and static feature maps has a related memory environment space with time and space information, this step can be classified. This step is equivalent to rewriting from the general script to the sub-script. It should also be pointed out that this general script may also be just one of the stage goals of completing machine planning, such as increasing the probability of a condition that may bring good profit value.

[0170] The machine needs to combine the intermediate goals in each link with the real environment again, and use the segmented imitation method to expand them layer by layer. The response plan proposed by the machine at the top level is usually composed of highly generalized process features and highly generalized static concepts (because these highly generalized processes can find multiple similar memories, so the responses established by borrowing from them are also highly generalized). For example, under the total output response of "business trip", "going to the airport" is an intermediate link goal. But this goal is still very abstract, and the machine cannot perform imitation.

[0171] Therefore, the machine needs to divide according to time and space, and take the links that need to be executed in the current time and space as the current goal. And take the goals of other time and space as the inherited goals and put them aside temporarily. After the machine takes the intermediate link as the goal, the machine still needs to further subdivide time and space (write the hierarchical script again). This is a process of increasing time and space resolution. The process of the machine converting a goal into multiple intermediate link goals is still to use decision-making ability, analyze various possible results and possible probabilities, and choose its own response according to the principle of "seeking benefits and avoiding harm". The above process is constantly iterating, and the process of dividing each goal into multiple intermediate goals is a completely similar processing flow. It must be decomposed until the bottom experience of the machine. The bottom experience for language is to mobilize muscles to produce syllables. For actions, it is to decompose to issue driving commands to related "muscles". This is a tower-shaped decomposition structure. The machine starts from the top-level goal and decomposes a goal into multiple intermediate link goals. This process is to create virtual intermediate process goals. If these intermediate process goals "meet the requirements", they will be retained. If they "do not meet the requirements", they will be recreated. Whether it "meets the requirements" means that the machine analyzes the gains and losses and confirms whether the strategy has reached the preset acceptable standards.

[0172] The above process is iterated layer by layer, and eventually a rich and colorful response of the machine can be established.

[0173] In this process, the machine may encounter new information at any time, causing the machine to process various information, and these original goals become inherited motivations. This is equivalent to constantly encountering new situations in the process of organizing activities, which need to be solved immediately, otherwise the activity cannot be organized. So the director stops other activities and solves the problem at hand first. After solving it, the activity continues. Another situation is that in this process, the director suddenly receives a new task, so after weighing the pros and cons, the director decides to suspend the activity first and give priority to the new task.

[0174] The machine is performing the imitation task that can be performed while breaking down other goals into more detailed goals. So the machine is thinking while doing. This is because the real situation varies greatly, and the machine cannot know the external situation in advance and make plans. So this is a process of achieving a goal through the interaction between the environment and the machine.

[0175] At this point, the machine can use the above capabilities to complete the understanding and response of input information. This process is a minimum cycle of interaction between the machine and the outside world. The machine repeats this process to achieve a greater goal, which is manifested as a continuous interaction process between the machine and the outside world, showing the intelligence of the machine.

[0176] The prediction, decision-making and response process of the above machines do not require new algorithms. They are existing algorithms. Based on the relationship network, they can realize the prediction, decision-making and response process of the machine by organizing them reasonably. The focus of this invention is to reveal how to establish methods, processes and steps for realizing general machine intelligence (strong artificial intelligence) by organizing existing algorithms and data. The existing statistical algorithms and artificial intelligence algorithms are not within the scope of the application of this invention and will not be repeated here.

Claims

1. A method for establishing strong artificial intelligence, characterized in that: include: S101, establishing a sensor module of the machine to sense information outside the machine and the machine itself; S102, establishing a machine-to-sensor input information simplification module for simplifying the input information, mainly including extracting underlying features of the input information by the machine; S103, establishing an associative activation module, for activating other information by using the extracted underlying features in accordance with the "proximity activation" principle, the "similar activation" principle and the "strong memory activation" principle, so as to establish or / and optimize a relationship network; "proximity activation" means that after a specific information in the memory is activated, the information adjacent to the specific information will be activated; "similar activation" means that when a specific feature in the memory receives an activation signal from other features, the receiving ability is positively correlated with the similarity between them; "strong memory activation" means that the higher the memory value, the stronger the ability to receive activation signals from other features, so that the deeply remembered information is more easily activated; S104, establishing an application layer of a machine for applying the relationship network; S105, establishing a communication module of the machine, so that the machine can exchange data with other machines in the outside world through the communication module according to a preset protocol; The machine considers that the input information that is adjacent in input time is connected to each other; The connection relationship is optimized through a memory and forgetting mechanism; The application layer includes: an initial activation value assignment system, an associative activation search device, an associative activation reconstruction device, a machine's underlying motivation device, a language information reconstruction device, a machine's predictive capability generation device, a machine's decision-making and response system, an imitation capability device, an execution device, and a machine's experience summary device. The initial activation value assignment system is used to assign an initial activation value to the input information. The associative activation search device is used to find experience related to the input information, motivations and motivation states that are connected to related experience information, and other types of information related to the input information in memory. The associative activation reconstruction device is used to obtain experience information related to the input information for reference by the machine response. The associative activation reconstruction device is also used for environmental reconstruction and self-image reconstruction. The underlying motivation device is used to give the underlying motivation to the driving machine power source, the language information reconstruction device realizes the context association recognition of the input information in a chain activation manner of language, the machine prediction ability generation device infers the various possibilities of the development of things and the corresponding probabilities or the various possibilities and corresponding probabilities of other people's behaviors based on past experience or similar experience, the decision and response system is used to evaluate and respond to the input information, the imitation ability device is used to imitate the behavior of other people or other machines, and continuously improve according to its own evaluation or external feedback information, the execution device specifically executes the response according to its own decision, and the experience summary device is used to enable the machine's experience to form connections in the relationship network through memory and forgetting mechanisms, and actively strengthen the connections.

2. The method according to claim 1, characterized in that: The input information is represented by multi-resolution features, and the method for implementing the representation includes representing the combination of the multi-resolution features; the connection relationship is established on the basis of the multi-resolution features of the information, and the connection relationship between the input information is different at different resolutions.

3. The method according to claim 1, characterized in that: The input information includes external input information, internal monitoring information, machine's engine and engine activation status information.

4. The method according to claim 1, characterized in that: The information storage method adopted by the machine enables the machine to express the connection relationship between adjacent information in input time through information storage.

5. A method for using strong artificial intelligence, characterized in that: include: The strong artificial intelligence established by the method for establishing strong artificial intelligence according to any one of claims 1 to 4 performs machine prediction, decision-making and execution; wherein, After receiving external or internal information, the machine uses the associative activation method to activate related information in the relationship network; the machine predicts the impact that similar events may have on its own motivation by counting the probability of each activated motivation and the corresponding memory value; the machine selects and promotes the path of event development in a way that seeks benefits and avoids harm based on the impact that the statistically analyzed events may have on its own motivation. The sequence goals on these paths are the behavioral decisions of the machine.

6. The method according to claim 5, characterized in that: The machine takes the planned output response as its own virtual information input, and continues to use associative activation, information reconstruction and prediction, decision-making and execution to evaluate the possible feedback on its response.

7. The method according to claim 5, characterized in that: The machine takes the expected external feedback as its own virtual information input, and continues to use associative activation, information reconstruction and prediction, decision-making and execution methods to evaluate the "benefits" and "harms" that the external feedback on its responses may bring to itself.

8. The method according to claim 5, characterized in that: The machine expands the search range by lowering the activation threshold or increasing the number of prediction and evaluation iterations, thereby finding the optimal response path from a larger search range.

9. The method according to claim 5, characterized in that: The way machines drive the path of events is a dynamic process. During the execution process, the machine continuously follows the same prediction method under new information, incorporates new information into the prediction, decision-making and execution process, and updates behavioral decisions.

10. The method according to claim 5, characterized in that: In the process of executing the sequence goals, the machine decomposes the single goal in the sequence goals into more low-level goals according to the same machine prediction, decision-making and execution algorithm, and decomposes them layer by layer during execution until the low-level driving commands that the machine can directly execute are decomposed; The machine's behavior under a series of underlying driving commands constitutes the machine's response to input information.

11. The method according to claim 5, characterized in that: When making predictions, the machine not only predicts the impact that its own behavior may have on its motivational state, but also predicts possible behaviors of others and the impact that these behaviors may have on its motivational state. When the machine predicts possible behaviors of others, it bases its predictions on assumptions including past memories of others' behavior in similar states and the assumption that others also make decisions in a way that seeks benefits and avoids harm.

12. A strong artificial intelligence device, characterized in that: The method for establishing strong artificial intelligence according to any one of claims 1 to 4 is capable of autonomous learning, decision-making and execution, and the strong artificial intelligence device comprises: Sensor systems, including general sensor groups that simulate human senses, sensor groups for machine-specific purposes, and sensor groups for machines to monitor their own operating status; The pre-set motivation system includes using specific symbols to represent a type of motivation and using a specific method to represent the state of the motivation. The starting point of the machine when making decisions is to drive a specific motivation to a specific state. A relational network system, including multi-resolution simplification of input information, assuming that there is a relationship between adjacent input information, and expressing such relationship in a specific way in memory; the relationship between multi-resolution information is optimized by memory and forgetting mechanisms; the relational network includes the information itself, and also includes the motivation and motivational state stimulated by the input information; Associative activation systems include systems that assign initial values ​​to the input information once or multiple times, these systems are implemented in the form of preset programs, and according to the learning results, the parameters in these preset programs are adjusted in a way of seeking benefits and avoiding harm; they also include associative activation methods based on the "proximity activation" principle, the "similar activation" principle and the "strong memory activation" principle; when the associative activation system stores the input information, the memory value assigned to the input information is both a positively correlated statistical function of the initial activation value assigned once or multiple times, and a positively correlated function of the activation value of the motivation state activated by the input information in the memory; A machine decision system is used to use the input information and activate related memories through an associative activation method; the machine limits its own search scope for memory and evaluation scope for its own decision through an associative activation method, and counts the size and probability of "benefits" and "harms" that may be brought to itself based on the probability of these motivations and motivational states; including distinguishing between self and non-self, and predicting the possible responses of self and non-self based on its own experience and the principle of seeking benefits and avoiding harm, and further iteratively evaluating the size and probability of "benefits" and "harms" that these responses may bring to itself; the machine determines the possible development direction of the event and the impact of the event development direction on itself through such iterative evaluation; the machine responds in a way of seeking benefits and avoiding harm to increase the probability of the event development direction tending to a direction that is beneficial to itself, and to reduce the probability of the event development direction tending to a direction that is harmful to itself; when executing the response, the machine needs to update the internal and external information of the machine at any time, and adopt the same decision-making method to correct its own response, and the goal is still to increase the probability of the event development direction tending to a direction that is beneficial to itself, and reduce the probability of the event development direction tending to a direction that is harmful to itself; The machine execution system is similar to the machine decision system, but the difference is that the machine execution system concretizes the goals generated by the machine decision system; the concretization process is also a kind of decision of the machine, and the decision is to subdivide the specific goals layer by layer to the bottom-level driving command level of the machine until the machine can directly execute the instructions; when the machine is subdivided layer by layer, it also adopts the machine decision system; The preset response system of the machine includes part or all of the following functional systems: the association system between the machine's emotion and motivation state, the association system between emotion and emotion expression, the machine's preset pronunciation and action system, and the preset conditioned reflex system; the preset response system of the machine is composed of parameters adjusted by preset programs and experience gained through learning.

13. A general artificial intelligence device, characterized in that: According to the method for establishing strong artificial intelligence according to any one of claims 1 to 4, the general artificial intelligence device is configured to include the following functions: The machine maps input data to local common features; The machine maps local common features to concepts; the machine establishes the relationship between the machine's motivational needs and external information through a relationship network, and uses the relationship to make predictions, decisions and responses.

Images