JP7226497B2 - Information processing device, method for causing information processing device to simulate intellectual behavior, computer program, and storage medium storing it - Google Patents

Description

The present invention relates to an information processing device, a method for causing the information processing device to simulate intellectual behavior, a computer program, and a storage medium storing it.

For example, devices and systems that realize advanced information processing like the human brain are known (eg, Patent Documents 1 to 3, Non-Patent Documents 1 and 2). The system described in Patent Document 1 is a computer system that simulates intelligent information processing of the human brain. The system described in Patent Document 1 includes six functional circuits that perform long-term perceptual memory, short-term perceptual memory, selection of perceptual expressions and information processing of working memory, action selection, action output, and cognitive control. It has control logic (feedback control, feedforward control, etc.) to describe interactions between functionally relevant brain anatomical regions.

The information processing device described in Patent Document 2 uses physical quantities (video, audio, etc.) obtained by a sensor and chemical quantities (components and compositions, etc.) as input information, and an input information holding unit and an internal representation code holding unit. An analysis unit that performs bi-directional conversion processing (analysis/inverse mapping) between them, and an internal information code storage unit and a high-level integrated code storage unit that are placed between one or more internal representation codes and associate them with each other. an integrated storage unit for storing relationships, an input information storage unit, an internal representation code storage unit, and an analysis unit that monitors the input information held in the high-level integrated code storage unit, the internal representation code, and the high-level integrated code It is a brain-type computer system having a supervisory control unit that controls the integrated storage unit.

The information processing device described in Patent Document 3 has a mechanism for holding one state variable in one axon and one state variable in one synapse, and changes the state variable of the axon based on the occurrence of the first event. , and update the synaptic state variables based on the updated axonal state variables and the occurrence of a second event, which can include multiple axons and synapses. In the information processing device described in Patent Document 3, in order to express neural plasticity that is important for learning and memory, changes in added weight values with respect to the occurrence of events and adjustments by neurotransmitters such as dopamine are performed. Update the synaptic weight value based on This can provide a way to modulate neural plasticity.

The information processing apparatus described in Non-Patent Document 1 converts amounts and actions indicating sensory input to changes in the environment and the resulting states in the brain (that is, states of perception/inference, learning/memory, attention/awakening) into free energy. We have a model that determines to minimize , and seeks actions that minimize the free energy. Minimizing the free energy means, in other words, reducing the difference (prediction error) between the amount indicating the state of sensory perception with respect to environmental changes and the amount indicating the state of that cognition. In the information processing device described in Non-Patent Document 1, the amount indicating the state of cognition is the neural activity corresponding to the state of the brain under the principle of free energy, memory, attention, and synaptic efficiency (synaptic connectivity ), determined by optimizing synaptic gains, which represent properties of brain evoked activity, plasticity, and neuromodulation, respectively.

The information processing device described in Non-Patent Document 2 specifies information about one experience in the form of a conceptual structure related to it (integrated information quality), and quantifies it with Φmax (integrated information amount), and the information Integrated Information Theory evaluates the conscious state of the brain as integrated information based on how mechanisms and current states affect the probability of past and future states (cause-effect power). Theory). A single neuron in the brain has an internal state of being active or inactive. This internal state is caused by input information, and its output affects other neurons. It is possible to reproduce the activity of cranial nerves spatially, temporally, and according to the type of activity (Low, High, Burst).

U.S. Pat. No. 9,208,430 JP-A-2000-259597 U.S. Patent Application Publication No. 2015/0286925

Karl Friston, The free-energy principle: a rough guide to the brain?, Trends in Cognitive Science, 2009, Vol.13, No.7, p.293-301 G. Tononi, Integrated information theory: from consciousness to its physical substrates, Nature Review Neuroscience, 2016, Vol.17, No.7, p.450-461 Masami Kamiyama, Neuroinformatics and Retina Research, Biophysics, 2011, 51(3), p.112-117 Toru HAMASAKI, Takahiro YAMAGUCHI, Masami IWAMOTO, Estimating the influence of age-related changes in skin stiffness on tactile perception for static stimulations, Journal of Biomechanical Science and Engineering, 2018, 13(1), p.17-00575

However, in the technique described in Patent Document 1, the integration of information from the environment and various parts of the body is performed only by the connection of neural circuits and their control logic, and the brain as a whole integrates and processes multiple pieces of information. method is not considered. In addition, the human brain cannot handle mechanical information, including the structure of the brain (arrangement of nerves and blood vessels, etc.) and physical properties (mechanics of materials, fluid dynamics, electromagnetism, etc.), which vary depending on the person and environment. and individual differences in experience and learning cannot be taken into account. In addition, it does not include the relationship between brain structure and function, so it cannot be used to predict functional decline or functional improvement due to brain injury.

Further, in the technique described in Patent Document 2, although there is a description of integrating and storing information step by step, such as conversion from input information to internal representation code and high-level integration code, the brain as a whole stores a plurality of information. No consideration is given to how to integrate and process In addition, although it is possible to select the learning mode, it does not consider a specific method for rewriting weight values through learning, which is necessary for expressing neural plasticity. Inability to handle mechanical information including structure (arrangement of nerves and blood vessels, etc.) and physical properties (material mechanics, fluid dynamics, electromagnetism, etc.), so consider individual differences in the human brain and experience/learning It is not possible. In addition, it does not include the relationship between brain structure and function, so it cannot be used to predict functional decline or functional improvement due to brain injury.

Further, in the technique described in Patent Document 3, no consideration is given to a method of integrating and processing a plurality of pieces of information in the brain as a whole with respect to a plurality of sensory information inputs. In addition, the impact of neural circuits specific to each brain function on information processing in the brain has not been sufficiently considered. Therefore, it is difficult to apply it for efficient information processing considering the structure and function of the human brain. Furthermore, we provide a method to rewrite the weight values of neural networks to adjust plasticity, which is important in learning. Since it cannot handle dynamic information including dynamics, fluid dynamics, electromagnetism, etc.), it cannot consider individual differences in the human brain and experience/learning. In addition, it does not include the relationship between brain structure and function, so it cannot be used to predict functional decline or functional improvement due to brain injury.

In addition, in the technique described in Non-Patent Document 1, brain activity measurement data using electroencephalograph (EEG), magnetoencephalograph (MEG), and functional MRI (Magnetic Resonance Imaging) when sensory stimulation is applied is applied to this information processing device, it is possible to investigate the mechanism of cognitive activity. Therefore, it is possible to reflect the structure of an individual's brain to some extent, and it is possible to perform information processing in consideration of the structure and function of the human brain. However, the technology described in Non-Patent Document 1 can only consider the mechanism of cognitive activity with respect to multiple stimuli, and does not include a method of intentionally selecting actions from among multiple tasks. not In addition, the free energy is calculated based on a statistical model of experimental data representing brain activity, and does not include mechanical information derived from the structure of the brain (arrangement of nerves and blood vessels, etc.). It cannot be used to predict decline or functional improvement.

In addition, although the technique described in Non-Patent Document 2 is effective for examining how the level of consciousness changes due to structural and functional damage to the brain, neural circuits specific to each brain function are The degree of influence on information processing is not sufficiently considered, and a method of intentionally selecting an action from among multiple tasks is not included. In addition, although it is possible to know changes in the active parts of the brain and the state of consciousness, it cannot be used to predict the possibility of brain function improvement through rehabilitation, etc., because it does not retain the learning function. In addition, it cannot handle mechanical information including the structure of the brain (arrangement of nerves and blood vessels, etc.) and physical properties (mechanics of materials, fluid mechanics, electromagnetism, etc.) that differ depending on the person and environment. Individual differences in experience and learning cannot be considered.

The present invention has been made to solve at least part of the above-described problems, and aims to provide a technique capable of predicting functional changes in the brain associated with changes in brain mechanics.

The present invention has been made to solve at least part of the above problems, and can be implemented as the following modes. An information processing device capable of substituting intellectual behavior, comprising a plurality of functional units that have learned behavior patterns related to each of a plurality of different functions of the brain, and integrating the plurality of functional units. When electrical information is input to the whole brain, mechanical information is obtained as a mechanical quantity obtained by performing an electromagnetic analysis of the whole brain, and the mechanical information is integrated into the plurality of and an information integration unit that associates the weight values of the functional units with the information processing apparatus. In addition, the present invention can also be implemented as the following modes.

(1) According to one aspect of the present invention, an information processing apparatus capable of substituting intellectual behavior is provided. This information processing device integrates a plurality of functional units corresponding to each function, which have learned behavioral patterns related to each of a plurality of different functions of the brain, and the plurality of functional units for input electrical information. and an information integration unit that obtains dynamic information of the entire brain and associates the dynamic information with the integrated weight values of the plurality of functional units.

There are individual differences in the structure of the human brain (arrangement of nerves and blood vessels, etc.), and these directly affect the functions and functions of the brain. In addition, there are individual differences in experiences and learning content from birth to adulthood and old age, which influence differences in personality, decision-making, emotional behavior, motor skills, etc. . According to this configuration, the information integration unit associates the mechanical information of the entire brain with the integrated weight values of the plurality of functional units. The mechanical information of an individual's brain changes as the information from the environment and the body changes. By associating such mechanical information with the weight values of multiple functional parts of the brain, it is possible to re-learn the multiple functional parts of the brain based on the mechanical information, thereby improving the structure and function of the brain associated with plasticity. change can be expressed. This makes it possible to express personal experiences and learning content. In addition, the mechanical information is obtained from the MR
It can be obtained from a medical tomography apparatus such as I (Magnetic Resonance Imaging) and CT (Computed Tomography). Since the information integration unit associates such dynamic information with the weight values of multiple functional units of the brain, it is possible to perform learning that reflects the individual's motor function, emotional function, sensory function, and the like. In addition, changes in mechanical information may also accompany damage to the cranial nervous system. When the damaged part of the cranial nervous system is specified, the information integration unit 20 can reflect the change in the mechanical information of the damaged part in the weight values of the plurality of functional parts of the brain. It is possible to predict the content of functional deterioration related to the body and mind due to the appearance of changes. As a result, it is possible to provide an information processing apparatus capable of predicting functional changes in the brain associated with changes in brain mechanics.

(2) In the information processing apparatus of the above aspect, the information integration unit configures an integrated network that integrates the plurality of function units by superimposing elements simulating neural networks included in the plurality of function units. applying a mechanical model obtained from the structure and physical characteristics of the brain to the integrated network, and performing electromagnetic analysis using the input electrical information on the integrated network to which the mechanical model is applied. Thus, the dynamic information may be determined and the determined dynamic information may be associated with weight values of the integrated network. According to this configuration, the information integration unit applies a dynamic model obtained from the structure and physical characteristics of the brain to the integration network, and performs electromagnetic analysis to obtain the input electrical information from the entire brain. It is possible to obtain the dynamic information as In addition, the information integration unit can re-learn by associating the obtained dynamic information with the weight value of the integration network.

(3) The information processing apparatus of the above aspect may further include a plasticity section that updates the weight values of the plurality of functional sections when the dynamic information exceeds a preset threshold value. According to this configuration, the plastic part updates the weight values of the plurality of functional parts when the dynamic information exceeds a preset threshold value, so that functional changes in the cranial nervous system can be simulated. Also, the weight values of a plurality of functional parts are associated with the mechanical information of the brain as a whole. Therefore, the update of the weight value influences the update of the mechanical information of the whole brain, so that structural changes of the cranial nervous system can be simulated in addition to the functional changes of the cranial nervous system. As a result, it becomes possible to make adjustments that take into consideration the interaction of each functional network, just like in the real brain. Further, by updating the weight values, it is possible to update the weight values in the learning model of a plurality of functional parts, and by updating the mechanical information, the mechanical model obtained from the structure and physical characteristics of the brain is updated. be able to. Therefore, processing using the updated learning model and dynamic model enables learning that reflects changes in the structure and function of cranial nerves.

(4) In the information processing device of the above aspect, the threshold may be determined based on a yield point in material mechanics. According to this configuration, since the threshold value for updating the weight value is determined based on the yield point in material mechanics, the weight value and mechanical information of the neural network, in other words, informatics and physics are combined. By matching, it is possible to simultaneously express functional changes and structural changes in the brain nervous system in neuroplasticity.

(5) In the information processing device of the above aspect, the plastic portion may further display the dynamic information. According to this configuration, since the plastic portion displays the dynamic information, convenience can be improved.

(6) The information processing apparatus of the above aspect may further include an action selection unit that selects an action to be preferentially executed from a plurality of actions executable by the plurality of function units. According to this configuration, the action selection unit selects an action to be preferentially executed from a plurality of actions that can be executed by a plurality of function units. For this reason, for example, a digital human model that simulates a person (a model that can output internal body information including blood pressure and muscle strength from each model data such as skeleton, bones, muscles, blood vessels, heart, and organs), and virtual space It is possible to simulate the physical actions of people and characters such as "walking, running, running away, fighting" and multiple behaviors related to emotional behavior, considering the priority behavior based on intentions. can express individual differences in

(7) In the information processing device of the above aspect, the action selection unit selects each of the function units under a multitasking condition in which the free energy based on the dynamic information is constant and a plurality of the actions can be executed. The action to be preferentially executed may be determined by solving a combinatorial optimization problem of minimizing the prediction error with respect to the target value of and maximizing the energy required for the plurality of actions. According to this configuration, the action selection unit minimizes the prediction error with respect to the target value of each function unit under multitasking conditions in which the free energy based on the dynamic information is constant and a plurality of actions can be executed. , and maximizing the energy required for a plurality of actions. Therefore, the action selection unit can handle a large amount of information and data in a format that can be analyzed by a mathematical model called dynamic information. In addition, when the action selection unit selects an action, it is possible to use a quantum computer that is good at solving combinatorial optimization problems. And efficient information processing on action selection becomes possible.

It should be noted that the present invention can be implemented in various aspects, for example, an information processing device capable of substituting intellectual behavior, a method/computer program for allowing an information processing device to simulate intellectual behavior, and this computer program. Realized in the form of a server device for distribution, a non-temporary storage medium storing this computer program, a digital human model equipped with this computer program, a virtual space system, a neurorehabilitation support system, a simulation system, a brain function evaluation system, etc. can do.

1 is an explanatory diagram illustrating the configuration of an information processing apparatus as one embodiment of the present invention; FIG. It is a figure explaining a function part and an information integration part. It is a figure explaining learning of a function part. It is a figure which shows an example of a recursive neural network. It is a figure which shows an example of the function map for every time interval, and a weight value. FIG. 4 is a diagram for explaining integration of a functional part and a dynamic model; FIG. 4 is a diagram for explaining the association between the hardness distribution of the whole brain and the weight values of the neural network; FIG. 4 is a diagram showing weight values and elastic moduli in brain functional regions where neural networks of functional regions intersect; It is a figure explaining a plastic part. It is a figure explaining an action selection part. It is a figure which shows an example when mechanical damage generate|occur|produces in a brain.

<Embodiment>
FIG. 1 is an explanatory diagram illustrating the configuration of an information processing device 1 as one embodiment of the present invention. The information processing apparatus 1 is a computer capable of performing human intellectual actions on behalf of a human, and is also called AI (Artificial Intelligence). The information processing apparatus 1 of the present embodiment can predict functional changes in the brain that accompany mechanical changes in the brain by having the following configuration. In the present embodiment, a person is taken as an example of a creature, but the information processing apparatus 1 may be made to realize the intellectual behavior of other creatures, not limited to humans.

The information processing device 1 includes a function unit 10, an information integration unit 20, a plasticity unit 30, and an action selection unit 4.
0. An input 100 to the information processing apparatus 1 is electrical information representing stimuli and signals from the environment and the body. The input 100 includes, for example, information on the external environment (light, sound, heat, vibration, acceleration, etc.), information on the five senses (visual, auditory, tactile, gustatory, olfactory), information on the inside of the body (blood pressure, heart rate, carbon dioxide (carbon concentration, etc.) can be exemplified. The output 200 from the information processing device 1 is electrical information representing reactions and signals to the environment and the body. As the output 200, for example, the muscle activation level and patterns of movements and facial expressions based thereon can be exemplified.

The functional unit 10 is composed of a plurality of functional units N1, N2, . It is Each of the functional units N1 to NN is implemented by a configuration capable of simulating a neural network of the brain, eg, a neural network, and stored in a storage unit (not shown). Each of the functional units N1 to NN is prepared by learning in advance for a large number of inputs 100 given in advance using a learning method suitable for each function of the brain. Also, each of the functional units N1 to NN can perform learning through a series of processes from receiving a new input 100 to outputting an output 200. FIG. An example of reinforcement learning as an example of learning will be described later in detail.

The information integration unit 20 integrates the functional units N1 to NN of the functional unit 10 to obtain mechanical information for the input 100 (electrical information) for the brain as a whole. The information integration unit 20 further associates the obtained dynamic information with the weight values of the integrated neural network of the entire brain. Specifically, the information integration unit 20 of the present embodiment integrates each of the functional units N1 to NN by superimposing elements (nodes, links) that simulate neural networks included in the functional units N1 to NN. An integrated network is constructed, and a dynamic model 21 obtained from the structure and physical properties of the brain is applied to this integrated network. The information integration unit 20 obtains dynamic information by performing electromagnetic analysis using the input 100 on the integrated network to which the dynamic model 21 is applied, and associates the obtained dynamic information with the weight value of the integrated network. Details will be described later.

When the mechanical information obtained by the information integration unit 20 exceeds a preset threshold value, the plasticity unit 30 re-learns information such as the weight value of the neural network of the whole brain corresponding to the mechanical information. By updating by, the structure and function of the brain realized by the information processing apparatus 1 are changed. Details will be described later.

The action selection unit 40 selects an action to be preferentially executed from among a plurality of actions that can be executed by each of the function units N1 to NN of the function unit 10. FIG. Specifically, the action selection unit 40 of the present embodiment has a constant free energy based on dynamic information, and under multitasking conditions in which a plurality of actions can be executed, the goals of each function unit N1 to NN Actions to be preferentially executed are determined by solving a combinatorial optimization problem that minimizes prediction errors for values and maximizes the energy required for multiple actions. Actions that can be selected by the action selection unit 40 are stored in advance in the action selection list 41 . For example, actions N1_A1 to A4 are actions related to the functional part N1, and actions N2_A1 to A4 are actions related to the functional part N2. The action selection list 41 is updated by adding new action candidates during the process. Details will be described later.

The information processing apparatus 1 selects the next motion for the given input 100 by the function unit 10, the information integration unit 20, the plasticity unit 30, and the action selection unit 40, and repeats the loop of outputting the output 200. By doing so, AI that learns and grows can be realized. According to this information processing device 1, for example, it is possible to predict functional deterioration of the brain associated with mechanical damage in a specific part of the brain, repeat action selection necessary for functional improvement, and give mechanical changes to the brain. It is also possible to predict functional improvement from the results of re-learning and re-prediction by.

FIG. 2 is a diagram explaining the function unit 10 and the information integration unit 20. As shown in FIG. 2A to 2D are diagrams showing examples of the functional units N1 to NN included in the functional unit 10. FIG. In FIG. 2A, a functional part N1 related to vision as a sensory function is represented in the form of a neural network. Similarly, FIG. 2(B) represents the functional part N2 related to hearing as a sensory function, FIG. 2(C) represents the functional part N3 related to emotional function, and FIG. A functional part N4 related to the motor function is represented. The black circles in the figure represent sets of neurons representing functional parts of the brain such as the thalamus, somatosensory cortex, basal ganglia, and amygdala (including connections between neurons in each functional part). call. A link LK connecting each node N represents a set of nerve fibers connecting sets of nerve cells. The arrangement of neural networks in the functional units N1 to NN may reflect, for example, the measurable shape of an individual, the arrangement of anatomical parts, the arrangement of blood vessels and cerebrospinal fluid, and the like. Each of the functional units N1 to NN receives as input 100 electrical information such as light, sound, smell, heat, vibration, acceleration, blood pressure, heart rate, and carbon dioxide concentration. An input 100 is transmitted through a node N and a link LK in each of the functional units N1 to NN, processed, and output as electrical information (output 200) such as reactions to the environment and body.

There are about 100 billion nerve cells in the brain, but in the information processing device 1 of the present embodiment, these nerve cells are divided into functional units N1 to NN (FIG. 1) and treated as aggregates of nerve cells. . Since the functional unit 10 is composed of a neural network that simulates the neural network of the brain as shown in FIGS. Instead, it is possible to consider from the input to the nervous system to the output, and the brain stem can also be included in the neural circuit. The motor function N4 shown in FIG. 2(D) is a neural network for performing a desired motion while controlling posture, so it is possible to realize a function such as picking up a cup and moving it to a specific place. , can correspond to what is traditionally modeled as “experience”.

When creating each of the functional units N1 to NN (FIG. 1) of the functional unit 10, learning is performed using a neural network representing each function. At this time, an appropriate learning method (unsupervised learning, supervised learning, reinforcement learning, etc.) is selected according to each function. For example, reinforcement learning was selected to represent the basal ganglia function, which is important in motor function, supervised learning was selected to represent the cerebellar function, and supervised learning was selected to represent the function of the motor and somatosensory areas. is selected for unsupervised learning. On the other hand, in order to represent the functions of the amygdala and nucleus accumbens, which are related to emotional functions, a learning method related to reinforcement learning may be selected according to each function. 2(A) and 2(B), a functional part N1 related to vision as a sensory function and a functional part N2 related to hearing as a sensory function are distinguished. The correspondence between (the sensory functions in the above example) and the functional units N1 to NN in the functional unit 10 may be one-to-many. Further, by grouping a plurality of functions of the brain (e.g., emotional function and sensory function) into one functional unit NN within the functional unit 10, the correspondence may be many-to-one.

FIG. 3 is a diagram for explaining learning of the function unit 10. As shown in FIG. FIG. 3A is a diagram showing an example of a neural network having units arranged in two layers. FIG. 3B is a diagram showing an example of a neural network having one layer of units. FIG. 3(C) is an explanatory diagram showing an example of the functional part N4 related to motor function. FIG. 3D is an explanatory diagram showing how reinforcement learning is performed by the functional unit N4. FIG. 3(E) is a diagram showing an example of a function map for muscle control of the function unit N4.

As described above, each functional unit of the functional unit 10 performs learning through processing of the new input 100 . Reinforcement learning of the functional part N4 related to motor function shown in FIGS. As shown in FIG. 3(D), a neural network including the basal ganglia is simulated in the functional part N4 related to the motor function. When performing reinforcement learning for the purpose of maintaining posture, which is one of the functions of this basal ganglia, in using the reinforcement learning algorithm, Fig. 3
By using a plurality of neural networks each having one-layer units shown in (B), the functional unit N4-1 shown in FIG. 3(D) is made into a learning model. For example, the input 100 from the body is the joint angle of a single joint and the joint angular velocity, and the output 200 is to obtain the muscle activity of each muscle for maintaining posture under gravity. Do reinforcement learning. Then, as shown in FIG. 3(E), a muscle control function, that is, a function map, is obtained that obtains the optimum activation level for maintaining posture for each muscle with respect to the inputs of the joint angles and joint angular velocities. be able to. Also, the function map when the optimum value is obtained and the weight value W _ij are stored in the storage unit for each function unit.

FIG. 4 is a diagram showing an example of a recurrent neural network. In the learning modeling of the functional unit N4-1 shown in FIG. Using a neural network having units arranged in two layers or a multi-layered neural network of three or more layers, a learning model can be created according to the connectivity of the neural network. At this time, if the recurrent neural network RNN (Recurrent Neural Network) shown in FIG. 4 is used, not only spatial learning but also time-series learning can be performed. Recurrent neural networks are used in the field of natural language processing for processing continuous information such as sentences. Furthermore, in addition to the recursive neural network RNN, if a long short-term memory (LSTM) method is used in combination, long-term memory can be realized, which is preferable.

FIG. 5 is a diagram showing an example of function maps and weight values W _ij for each time interval. As described above, if the recursive neural network RNN _and the long/short-term memory LSTM are used, as shown in FIG. Information learned over time can be stored as memory using functional maps. Information about the temporal change of the function map and the temporal change of the weight value W _ij when the optimum value is obtained is stored in the storage unit for each functional unit. Human beings are thought to accumulate experiences from birth as memories for each function. However, not all memories can be retrieved instantly, and it is thought that there is a limit to the memories that can be retrieved immediately. In the information processing apparatus 1 of the present embodiment, stored information that can be retrieved immediately can be obtained from the function map by using it as a lookup table, and normal information can be obtained from calculation using the weight value W _ij .

The example of FIG. 3E shows an example in which joint angles and joint angular velocities are input. However, information about these joint angles and joint angular velocities is detected by joint capsule receptors, which are musculoskeletal mechanoreceptors in joints, and transmitted to the brain as electrical information. Thus, in order to handle stimuli and signals from the environment and the body as electrical information, a method of converting mechanical stimuli into electrical information, such as mechanoreceptors, is required. In this regard, it is possible to use a mathematical model that converts environmental stimuli into electrical information using nonlinear differential equations proposed by Hodgkin-Huxley. Several mathematical models have been proposed for the senses of sight, hearing, and touch. In addition, if the technique described in Non-Patent Document 4 is used, the impulse response of Merkel cells when pressure or two-point stimulation is applied to the finger skin by a skin mechanics model using the finite element method is correlated with it. It can be output as Mises stress. By using such a mathematical model of each sensory organ, it is possible to obtain electrical information on stimulation.

FIG. 2E is a diagram showing an example of an integrated network Nj that integrates neural networks of functional units N1 to NN. FIG. 2(F) is an explanatory diagram of the structure and physical properties of the brain. FIG. 2F conceptually represents a computational model 21 that expresses the actual brain structure and physical properties based on material mechanics and fluid mechanics. The calculation model 21 is a basic model for performing physical calculations of material dynamics and fluid dynamics, and is hereinafter also referred to as a "dynamic model 21" (FIG. 1: information integration unit 20, dynamic model 21). The dynamic model 21 is calculated in advance and stored in a storage unit (not shown) of the information processing device 1 . The dynamic model 21 reflects, for example, the structure of the brain and the characteristics of blood vessel arrangement. The brain characteristics reflected in the dynamic model 21 may be those of a certain individual or generalized ones. FIG. 2G is a diagram explaining the integration of the integrated network Nj and the computational model 21. As shown in FIG. FIG. 2F is a diagram showing an example of stress distribution with respect to the input 100. FIG.

The information integration unit 20 integrates the electrical information received as the input 100 by each of the functional units N1 to NN of the functional unit 10 using the mechanical model 21, and converts it into mechanical information (for example, stress distribution) of the entire brain. . Specifically, the information integration unit 20 arranges the integrated network Nj (integrated neural networks of the functional units N1 to NN) shown in FIG. (2 (G)). Thereby, the information integrating section 20 can obtain the stress distribution Ns(t ¹ ) at a certain time t ¹ when the input 100 is given to the integrated network Nj.

FIG. 6 is a diagram illustrating the integration of the functional part N1 and the dynamic model 21. As shown in FIG. The information integration unit 20 applies the dynamic model 21 to each of the functional units N1 to NN shown in FIGS. 2A to 2D separately from the integrated network Nj to obtain the respective stress distribution. may In the example of FIG. 6, the information integration unit 20 performs an electromagnetic analysis on the dynamic model 21 using the arrangement of the function unit N1. A stress distribution Ns(t ¹ ) at t ¹ is obtained. In the case of FIG. 6, the stress distribution of the brain when vision as a sensory function receives stimulation from the outside is obtained.

The information integration unit 20 may further associate the integrated dynamic information (the stress distribution in the above example) with the neural network weight value W _ij or the like. In the past, there was no method for directly connecting physics, which deals with the theoretical calculation of dynamical information, and informatics, which deals with estimating the relationships between multiple pieces of information such as neural networks. . However, since the density distribution of nerve fibers in the brain is strongly related to the hardness distribution of the brain, it can be said that a hard region in the hardness distribution of the brain has a high density of nerve fibers in the distribution of the nerve fibers in the brain. It can be seen that a high nerve fiber density facilitates nerve transmission, that is, the weight value of the neural network increases. Hereinafter, a method of associating the stiffness distribution of the whole brain as dynamic information obtained based on this idea with the weight values of the neural network will be described.

FIG. 7 is a diagram for explaining the association between the hardness distribution of the whole brain and the weight values of the neural network. FIG. 7A shows a mechanical model 21 representing the hardness distribution of the entire brain. FIG. 7(B) represents the learning model 22 of the whole brain represented as a neural network. Here, in the mechanical model 21 (FIG. 1: information integration unit 20, mechanical model 21, FIG. 7A), the brain exhibits nonlinear mechanical characteristics, but only linear mechanical characteristics in small deformation are considered. , the dynamics model 21 is expressed as in equation (1). In equation (1), σ _i is the equivalent stress, E _ij
is the elastic modulus, ε _i is the equivalent strain, and σ _j0 is the initial equivalent stress. Note that i,
j=1 to n, where n represents the total number of brain regions.

On the other hand, the learning model 22 (FIG. 1: information integration unit 20, learning model 22, FIG. 7(B)) represented by a neural network is expressed as shown in Equation (2). In equation (2), y _i represents the output value, W _ij the weight value, x _i the input value, and b _j the bias, respectively. It should be noted that i
, j=1 to n, where n represents the total number of brain regions. The summation convention was used for the right term of equation (2).

Considering the above-described correspondence relationship between the hardness distribution of the whole brain and the weight values of the neural network, the elastic modulus E _ij of the dynamic model 21 shown in FIG. can be considered to be equivalent to the weight value W _ij of . Here, the elastic modulus E _ij indicates the elastic modulus in the direction from a part i to a part j, and the weight value W _ij indicates a weight value when a signal is transmitted from a part i to a part j. Thus, if the elastic modulus E _ij and the weight value W _ij are equivalent, the equivalent stress value σ _i of the dynamic model 21 shown in FIG. 7(A) and the learning model 2 shown in FIG. 7(B)
2 output values y _i can be considered to be equivalent.

However, E _ij is usually expressed in a Cartesian coordinate system, for example in a finite element model. However, the weight value W _ij in the learning model 22 in FIG. 7B and the elastic modulus E _ij in the dynamic model 21 in FIG. is represented by Therefore, in order to associate with a finite element model or the like, coordinate transformation is required for basis vectors τ' i and _{τ i of} new coordinates x' _i and old coordinates x _i , respectively, as shown in equation (3). By performing the coordinate transformation of Equation (3), it becomes possible to associate the elastic modulus E _ij and the equivalent stress σ _i obtained by the dynamic model 21 with the weight value W _ij and the output value y _i of the learning model 22. Become.

FIG. 8 is a diagram showing weight values and elastic moduli in brain functional regions where the neural networks of the functional regions N1 to N4 intersect. Among the nodes N constituting each of the functional units N1 to N4 shown in FIGS. 2A to 2D, there are also nodes N that are duplicated among the plurality of functional units N1 to NN. The thalamus and insular cortex, which are hubs of brain functions, correspond to this. In such a case, as indicated by thick-framed circles in FIG. 8, a plurality of overlapping nodes N exist in one functional brain region Np. Therefore, the correspondence relationship between the hardness distribution of the whole brain and the weight values of the neural network (correspondence relationship between the dynamic model 21 and the learning model 22) described with reference to FIG. 7 can be applied as it is.

Using MRE (Magnetic Resonance Elastography), it is possible to determine the distribution of elastic properties of brain tissue. MRE is an MRI (Magnetic Resonance Imaging) apparatus with a high magnetic field. By applying an oscillating gradient magnetic field that is synchronized with the vibration by an external vibrator that can be driven in a high magnetic field MRI (Magnetic Resonance Imaging) apparatus, the viscoelasticity of the tissue is obtained from the wave image of the target tissue. This is a method for obtaining characteristics. Although there is a resolution problem, the use of MRE makes it possible to obtain the elastic modulus E _ij of individual brain tissue. On the other hand, the data obtained by DWI (Diffusion Weighted Imaging), a diffusion MRI-enhanced image using a high-field MRI apparatus, and DSI (Diffusion Spectrum Imaging), are used in DSI Studio.
By analyzing using software such as, it is possible to construct a structural network of an individual's brain, and obtain information on the direction of line segments connecting the above-mentioned connecting sites. This makes it possible to obtain the elastic modulus E _ij of the individual brain tissue corresponding to the direction of the nerve. The weight value W _ij can also be obtained by using the correspondence relationship between the hardness distribution of the whole brain and the weight value of the neural network (correspondence relationship between the dynamic model 21 and the learning model 22) described in FIG. Therefore, these can be given as initial values when reflecting the structure of an individual's brain.

The plastic part 30 displays the mechanical information converted by the information integrating part 20 as a mechanical quantity σ _p (e.g., equivalent stress value) of the related parts between the functional parts N1 to NN.
In addition, when the mechanical quantity σ _p of the related part becomes larger than a preset threshold value σ _y (Fig. 1: σ = σ _p > σ _y ), the plastic part 30 sets the weight value W _ij of the neural network to Update. Here, the threshold σ _y is represented by a mechanical quantity (e.g. equivalent stress value). The threshold σ _y is a value that reproduces neural plasticity in neurophysiology and can be adjusted to reproduce experimental data.

FIG. 9 is a diagram illustrating the plastic portion 30. FIG. FIG. 9(A) shows a schematic diagram of plasticity in material mechanics. FIG. 9(B) shows a schematic representation of neural plasticity. As shown in FIG. 9, in the information processing apparatus 1 of the present embodiment, the plasticity of nerves (that is, the setting of the threshold σ _y ) adopts a concept similar to the state of reaching the plasticity region in material mechanics, that is, the yield point. there is In material mechanics, in the elastic region, even if it is repeatedly deformed, it returns to its original state, but once it enters the plastic region, it does not return to its original state and changes its shape. Similarly, normal weak signal transmission in nerves does not affect memory, but when there is signal transmission that causes emotional experience or strong sensory reception, memory and action selection are significantly affected. may have an impact. In such cases, it is thought that structural and functional changes occur in the cranial nervous system. In this way, correlating mechanical information and neural network weight values, in other words, physics and informatics, can simultaneously express structural and functional changes in the brain and nervous system in terms of neural plasticity. There are advantages to be had.

In this way, the plastic portion 30 is configured such that when the mechanical quantity σ _p (for example, mechanical information such as the equivalent stress value) of the related portion between the functional portions N1 to NN becomes larger than the threshold value σ _y , as shown in FIG. , the equivalent stress σ _p at that time is associated as the output value y _p (Fig. 1: plastic portion 30, σ _p →y _p ). As a result, as indicated by the dashed frame in FIG _. 1, each of the functional units N1 to NN performs re-learning calculations to derive the weight value W _ij that satisfies the output value y _p . are updated, and the updated weight values W _ij are stored in storage units (not shown) of the functional units N1 to NN. This indicates that functional changes in the cranial nervous system have occurred. On the other hand, when the weight value W _ij is updated, the elastic modulus E _ij is also updated based on the correspondence described in FIG. 7 (FIG. 1: plastic portion 30, W→E). As a result, the hardness distribution of the entire brain changes, indicating that structural changes have occurred in the cranial nervous system.

The updated elastic modulus E _ij also updates the elastic modulus E _ij of the dynamic model 21 of the information integration part 20 , as indicated by the dashed arrow extending from the plastic part 30 to the information integration part 20 in FIG. 1 . Similarly, the updated weight values W _ij update the weight values W _ij of the learning models 22 of the functional units N1 to NN. Calculations after the update are performed using the post-update elastic modulus E _ij and weight value W _ij , enabling learning that reflects changes in the structure and function of cranial nerves. Normally, each functional part N1 to NN learns individually so as to realize each function, and the weight value W _ij etc. are determined, but in the actual brain, each functional network in the brain influences each other. is thought to affect In this regard, in the information processing apparatus 1 of the present embodiment, the weight value W _ij is updated by re-learning in the plastic part 30 . When updating, the weight values W _ij of the individual networks are adjusted so as to correspond to the stress distribution. Therefore, it is possible to make adjustments considering the interaction of each functional network as in the actual brain.

In daily actions, the human brain receives stimuli/signals from the environment and the body, such as the input 100 to the information processing apparatus 1, and takes actions in response to them. At that time, if you are in a normal state of consciousness, you always perceive the state (current value) such as current sensations, emotions, thoughts, and movements (equivalent to each functional part N1 to NN), and an internal model based on past experiences ( The action is selected based on the perceived error between the current value and the predicted value compared with the predicted value by the learning model 22). The information processing apparatus 1 of this embodiment also employs a method of selecting an action based on this principle.

With regard to neural activity in the brain, recent studies using electroencephalographs, functional MRI, etc. have revealed that brain activity is observed during sleep as well as during wakefulness. Also, although humans can perform multiple tasks at the same time, there is a limit to how much they can handle, and it is difficult to perform multiple tasks at the same time with the same quality. From this point of view and the point of view of homeostasis of the living body, the energy required for brain activity is considered to be constant. On the other hand, from the physical point of view of the brain, the electrical signals of nerves from the body, blood flow, spinal fluid, etc. enter the brain, but the free energy due to these actions is constant as long as there is no loss to the outside. It is conceivable that From the above point of view, hereinafter, we assume that the free energy based on the integrated mechanical information of the whole brain is constant.

The action selection unit 40 provides means for selecting the optimum action based on the intention of determining priority actions. In the following, it is assumed that a plurality of tasks are performed simultaneously, tasks related to functional networks N1, N2, and N3 are T1, T2, and T3, respectively, and three tasks are performed simultaneously. Among them, it is assumed that the task T2 is preferentially performed. In this case, in response to an input 100 such as a stimulus or signal from the environment or the body, in each of the functional units N1 to NN, for each of the tasks T1, T2, and T3, the target value of each task (assumed image of the task) and the internal model can be obtained from the following equation. Here, the internal model means a neural network model based on the weight values Wij stored in the storage _unit , or the function maps shown in FIGS. 3(E) and 5. FIG. In equations (4), (5), and (6), e ^N1 , e ^N2 , and e ^N3 are prediction errors, and y ^N1 , y ^N2 , and y ^N3 are prediction values by the internal model (output values of the learning model 22 y _i ), and y ^N10 , y ^N20 , and y ^N30 respectively represent the target values of each task.

Considering the correspondence relationship described in FIG. 7, the predicted values y ^N1 , y ^N2 and y ^N3 by the internal model are associated with the corresponding stresses σ ^N1 , σ ^N2 and σ ^N3 respectively. Therefore, the brain free energies (Helmholtz free energies) associated with tasks T1, T2, and T3 can be expressed as E(σ ^N1 ), E(σ ^N2 ), and E(σ ^N3 ), respectively, The above-mentioned condition that the free energy is constant can be expressed as in the following equation (7). In equation (7), the free energy E ^Total of the entire brain is the sum of the energy for all the functional parts N1 to NN to act and the energy Eh for maintaining the rest of the brain in a steady state. .

FIG. 10 is a diagram explaining the action selection unit 40. As shown in FIG. For example, assume that task T1 and task T3 are acting as intended, ie, in a state where e ^N1 =0 and e ^N3 =0, T2, which is a priority task, is added. At this time, as shown in FIG. 10, the action selection unit 40 selects from the plurality of possible actions A11 to A34 such that the prediction error e ^N2 of the task T2 is minimized, that is, the function unit N2 Select the action of task T2 that maximizes the activity of the action portion. At this time, if the free energy E(σ ^N2 ) is increased in order to increase the activity of the functional part N2 under the constant free energy condition of Equation (7), the energy related to the activities of the other tasks T1 and T3 will be Therefore, it is necessary to optimize the initial prediction errors e ^N1 and e ^N3 so that they do not become large. This problem is a combinatorial optimization problem, and various solutions have been proposed.

Recently, quantum computers using quantum annealing methods can solve combinatorial optimization problems quickly and efficiently. For this reason, for example, as shown in FIG. 10, after an input 100 is given, the above-described series of processes by the functional unit 10, the information integration unit 20, and the plasticity unit 30 are performed in the von Neumann computer. The processing of the combinatorial optimization problem by may be performed by a quantum computer. In the information processing device 1, when implementing a large number of functional units N1 to NN, in order to realize the optimal action selection based on the intention under the condition of expression (7), a combinatorial optimization problem is solved using a quantum computer. Efficient to solve.

When the prediction errors e ^N1 to ^NN are 0 or very small, the human acts unconsciously and reacts without perceiving it. When the prediction errors e ^N1 to ^NN are large, the human acts consciously. Therefore, the action selection unit 40 can determine whether the action is conscious or unconscious based on the magnitude of the prediction error. While it has been known that prediction errors account for a large part of human cognitive activities, it is also known that not only simple prediction errors but also the degree of integration of each function in the brain is important for consciousness. This point has not been clarified even in the current neuroscience field. In the information processing apparatus 1 of the present embodiment, the action selection unit 40 integrates the activities of the function units N1 to NN as dynamic information, and selects each function unit N1 so as to reduce the prediction errors e ^N1 to ^NN . Actions are selected by solving combinatorial optimization problems under the condition that the energy E ^Total of integrated brain activity is constant from among the multitasks that each NN is in charge of. For this reason, it contains a different form of the conventionally known concept of consciousness. When the prediction error is 0 or very small in all of the functional units N1 to NN, the action selection unit 40 assumes an unconscious action. do not implement

The actions A11 to A34 selected by the action selection unit 40 are expressed as an output 200, that is, reactions/signals to the environment/body. The output 200 is, for example, the sensory-motor activity represented by the functional part N4, and the muscle activity level and the body movement based on it, and the emotional behavior represented by the functional part N4, the muscle activity level of the facial muscles and the facial expression based on it. pattern (joy, anger, etc.).

In each functional unit N1 to NN, when considering a neural network related to thinking, a function map in which inputs 100 from various environments and outputs 200 associated with reactions and actions to each environment are associated is It is saved in a storage unit (not shown). Even in such a case, the action selection unit 40 uses the prediction values y ^N1 to ^NN of the action result by the internal model (equivalent to the learning model 22) reflecting the individual's way of thinking and thought, and the prediction error e ^N1 of the actual action result. ~ When ^NN is obtained, induce a conscious emotional change (joy, anger, etc.), or list action candidates (related thinking functional parts N1 to NN) for selecting the next action. Action N1#A1 in action selection list 41 shown in FIG.
NN#N4 and actions A11-A34 shown in FIG.
It is an action that can be cited as a candidate for each. In other words, when a large prediction error e N1 to NN occurs in the course of the function units ^N1 to ^NN continuing learning with respect to the new input 100, the action selection unit 40 interprets it as conscious brain activity. As a result, the action selection unit 40 adds them to the action selection list 41 as new action candidates for the function units N1 to NN. As a result, new action candidates can be selected when the action selection unit 40 solves the combinatorial optimization problem.

As described above, in the information processing apparatus 1, the next action is selected by the function unit 10, the information integration unit 20, the plasticity unit 30, and the action selection unit 40 using stimuli and signals from the environment and the body as the input 100, A loop for outputting a response or signal to the body as an output 200 is repeated. As a result, the information processing apparatus 1 can simulate a process in which a person learns and grows through experience. At this time, in each of the functional units N1 to NN, the learning result (weight value W _ij , function map when the optimum solution is obtained) regarding the neural network of each function is stored in the storage unit. In addition, the plastic part 30 changes the function of the brain (weight value W _ij ) when the mechanical quantity σ _p (mechanical information) becomes larger than a preset threshold value σ _y . The structure (elastic modulus E _ij ) is also changed (FIG. 1: flexible section 30, W→E, dashed arrow extending from plastic section 30 to information integrating section 20). The action selection unit 40 selects the next action using the changed, in other words, grown brain structure (dynamic model 21) and function (learning model 22). If an action causes prediction errors e ^N1 to ^NN , the action selection unit 40 adds the action to the action selection list 41 . These series of processes correspond to accumulating parts related to a person's personality, such as a person's experience and aggressiveness. Both the dynamic model 21 and the learning model 22 continue calculation as long as the input 100 is given. That is, the stress value is not reset to 0 in the dynamic model 21, and is updated even during repeated learning, and the stress value continues to be updated.

FIG. 11 is a diagram showing an example of mechanical damage to the brain. For example, when a nerve is mechanically damaged in a specific region of the brain surrounded by a dashed frame, the elastic modulus E _ij of that region becomes zero. Here, considering the correspondence relationship described with reference to FIG. 7, the weight value W _ij of the neural network at the site where the mechanical damage has occurred is also 0. This results in the inability to perform actions related to the impaired function. From this, in the information processing device 1, it becomes possible to express brain function deterioration accompanying the mechanical damage of the brain.

A similar consideration can also be given to the case where the elastic modulus E _ij is changed by giving some contribution from the outside to give a mechanical change in a specific part of the brain. In this case also, the weight value W _ij of the neural network is changed in the portion where the elastic modulus E _ij is changed, taking into consideration the correspondence relationship described in FIG. Re-learning is performed using the changed weight value W _ij , and from the obtained re-prediction result, it is possible to predict whether the contribution to the brain will lead to functional improvement. Here, as a contribution from the outside, action selection for function improvement, that is ^, as explained in ^FIG . By repeating action selection that increases the brain activity of , a mechanical change in the brain, that is, a change in the elastic modulus E _ij may occur. For this reason, the action selection described in FIG. 10 is important in improving functional decline associated with brain damage and mechanical changes.

1 and 10, the functional units N1 to NN (neural networks) set in advance have been described as targets. However, when the brain is actually damaged, re-learning may be performed by constructing a new neural network in a part unrelated to the original neural network. In order to realize this, it is necessary to rearrange the nodes N constituting the functional units N1 to NN shown in FIG. 2 and connect them by new links LK. The theory of self-organization is effective for rearrangement of nodes N and connection by new links LK.

A digital human model is effective for evaluating human comfort, workability, and mental state on a computer. A digital human model is a model that can output internal body information including blood pressure and muscle strength from each model data such as skeleton, bone, muscle, blood vessel, heart, and organ. The information processing apparatus 1 of the present embodiment can control input/output such as sensorimotor and emotional behavior of such a digital human model. By using a combination of the digital human model and the information processing apparatus 1 of the present embodiment, it is possible to evaluate human comfort and workability on a computer.

For example, suppose a sensory motion like a golf swing. Proprioceptive information from a plurality of joints/muscles of the body, tactile information from a hand, and the like are transmitted to the somatosensory area using the functional unit N4 (neural network) shown in FIG. 2(D). The motor cortex also provides the body with information about the muscle activity of each muscle that allows the golf swing to be achieved. At this time, posture is controlled in cooperation with the basal ganglia and the cerebellum, and muscle activity is adjusted so as to follow the movement trajectory of the target image. In this way, the learning for the functional unit N4 is performed, and the learning is held in the storage unit of the functional unit 10 as a function map. When the environment is windy or the footing is bad, combination optimization is performed with golf swing as the priority action from among the plurality of action selections shown in FIG. 10 . At this time, for example, when sadness is felt as an emotion, part of the activity energy is devoted to the related functional parts N1 to NN, and sufficient activity energy cannot be devoted to muscle control in the golf swing. It is possible to simulate a situation where a good shot cannot be hit.

In general, the nucleus accumbens of the brain stores the conditions for obtaining reward, and the amygdala stores the conditions for obtaining fear/anxiety. It is believed that a neural network including these two functional brain regions is involved in human evaluation of comfort. In the information processing apparatus 1 of the present embodiment, the method shown in FIG. Learning can be performed using the functional part of the reward system including the nucleus accumbens, and learning can be performed using reinforcement learning in the functional part including the amygdala for displeasure and fear.

In addition, a function map or the like that associates the input 100 and the output 200 when pleasure or discomfort is obtained is obtained, and prediction errors e ^N1 to ^NN are obtained from the current situation and the function map. Since the action selection unit 40 can obtain the pleasant/unpleasant ratio from the obtained prediction errors e ^N1 to ^NN , it is possible to evaluate the comfort/unpleasant in the current environment. Furthermore, behaviors for pleasure and displeasure are selected based on intention from among multiple behavior selections, for example, behaviors that provide more reward for pleasure and pleasure are selected based on displeasure and fear. In response, an action such as running away may be selected. According to the information processing apparatus 1 of the present embodiment, it is possible to evaluate pleasure and displeasure by a new method that has never existed in the past, and to select an action based on the intention.

In the information processing apparatus 1 of this embodiment, the brain structure of an individual may be given as an initial value for the layout of neural networks in the functional units N1 to NN. By doing so, by repeating learning with a plurality of stimuli/signals (input 100) from the environment and the body, it is possible to learn and grow differently for each individual. For example, in the case of the golf swing combined with a digital human model, it is possible to perform learning calculations that incorporate the structure of an individual's cranial nerves and the structural characteristics of an individual's body. can correspond to In addition, in the evaluation of comfort/discomfort, the structure of individual cranial nerves may be considered. By increasing the action selection list 41 accompanying learning as described in FIG. 10, it is also possible to evaluate actions according to individual characteristics. A function that measures an individual's physiological quantity (blood pressure, heart rate, respiratory rate, exercise, etc.) and outputs 200 the individual's reaction/response (sensory evaluation results, etc.) when that information is used as input 100. If we can create a map and associate it with learning calculations, the accuracy of individualized learning and growth models will be further improved.

In addition, when the head is damaged in an accident, etc., or when there is an organic change in the neural structure of the brain due to mental illness, etc., as explained in FIG. , the weight value W _ij changes, and each function of the brain (function parts N1 to NN) is also affected. For this reason, the information processing apparatus 1 of the present embodiment provides useful knowledge for diagnosis in the medical field and for formulating treatment plans. In addition, by repeatedly changing the stimulus/signal (input 100) received from the body, such as rehabilitation and improvement of diet, the person acts consciously. By repeatedly selecting an activity to improve, if a large change can be given to the brain activity of the related neural network (for example, functional part N1), the neural structure of the related brain (for example, functional part N4) can also be changed. It may also be possible to reconfigure neural networks that change and function better. Thus, the information processing apparatus 1 of the present embodiment can also be used as a neurorehabilitation support tool for assisting rehabilitation of physical disabilities and mental illnesses.

<Modification of this embodiment>
The present invention is not limited to the above-described embodiments, and can be implemented in various aspects without departing from the scope of the invention. For example, the following modifications are possible.

[Modification 1]
In the above embodiment, an example of the configuration of the information processing device 1 is shown. However, the configuration of the information processing apparatus 1 can be modified in various ways. For example, the information processing device 1 may be configured by cooperation of a plurality of information processing devices arranged on a network. In this case, for example, at least part of the functional unit 10, the information integration unit 20, the plasticity unit 30, and the action selection unit 40 may be implemented by different information processing devices. For example, the information processing device 1 may include a plurality of sets of functional units 10, information integration units 20, plasticity units 30, and action selection units 40, each corresponding to a different person. For example, the information processing device 1 does not have to include at least one of the plasticity section 30 and the action selection section 40 .

[Modification 2]
In the above embodiment, each functional unit N1 to NN of the functional unit 10 is configured by a neural network. However, each of the functional units N1 to NN may be configured by means other than the neural network (for example, the functional map shown in FIG. 3(E)). For example, the functional unit 10 may store an integrated network Nj obtained by previously integrating neural networks of the functional units N1 to NN separately from the functional units N1 to NN.

[Modification 3]
In the above-described embodiment, the dynamics model 21 of the information integration unit 20 is a model for performing physics calculations of material dynamics and fluid dynamics. However, the mechanical model 21 may be configured as a model that expresses at least one physical property such as material dynamics, fluid dynamics, and electromagnetism.

[Modification 4]
In the above embodiment, the plastic part 30 updates the weight value when the dynamic information becomes larger than the preset threshold value. However, the plastic part 30 may update the weight value each time the process is performed without using the threshold value. For example, the threshold for updating the weight values may be determined independently of the yield point in material mechanics, and may be user modifiable. For example, the plastic part 30 may omit the display of the dynamic quantity (dynamic information).

[Modification 5]
In the above embodiment, the action selection unit 40 selects a plurality of actions to be executed and an action to be preferentially executed under a multitasking condition in which a plurality of actions can be executed. However, the action selection unit 40 may select one action to be executed under the condition that only a single action is executed. For example, the action selection unit 40 selects an action by any method other than solving a combinatorial optimization problem that minimizes the prediction error with respect to the target value of the function unit and maximizes the energy required for the action. good too.

The present aspect has been described above based on the embodiments and modifications, but the above-described embodiments are intended to facilitate understanding of the present aspect, and do not limit the present aspect. This aspect may be modified and modified without departing from the spirit and scope of the claims, and this aspect includes equivalents thereof. Also, if the technical features are not described as essential in this specification, they can be deleted as appropriate. Some or all of the functions and processes implemented by software in the above embodiments may be implemented by hardware. Also, part or all of the functions and processes implemented by hardware may be implemented by software. As hardware, for example, various circuits such as integrated circuits, discrete circuits, or circuit modules combining these circuits can be used.

DESCRIPTION OF SYMBOLS 1... Information processing apparatus 10... Function part 20... Information integration part 21... Dynamic model 22... Learning model 30... Plasticity part 40... Action selection part 41... Action selection list

Claims (10)

An information processing device capable of substituting intellectual behavior,
a plurality of functional units corresponding to each function, each having learned a behavioral pattern related to each function for a plurality of different functions of the brain;
obtaining mechanical information, which is a mechanical quantity obtained by performing an electromagnetic analysis of the entire brain when electrical information is input to the entire brain in which the plurality of functional units are integrated; with the integrated weight values of the plurality of functional units;
An information processing device. The information processing device according to claim 1,
The information integration unit
configuring an integrated network that integrates the plurality of functional units by superimposing elements that simulate neural networks included in the plurality of functional units;
Applying a mechanical model obtained from the structure and physical properties of the brain to the integrated network,
obtaining the mechanical quantity as the mechanical information by performing electromagnetic analysis using the input electrical information on the integrated network to which the dynamic model is applied;
An information processing device that associates the obtained dynamic information with a weight value of the integrated network. The information processing apparatus according to claim 1 or claim 2, further comprising:
An information processing apparatus, comprising: a plasticity unit that updates weight values of the plurality of functional units when the dynamic quantity as the dynamic information exceeds a preset threshold value. The information processing device according to claim 3,
The information processing device, wherein the threshold is determined based on a yield point in material mechanics. The information processing device according to claim 3 or claim 4,
The information processing device, wherein the plastic part further displays the dynamic information. The information processing apparatus according to any one of claims 1 to 5, further comprising:
An information processing apparatus comprising an action selection unit that selects an action to be preferentially executed from among a plurality of actions executable by the plurality of function units. The information processing device according to claim 6,
The action selection unit
Under multitasking conditions where the free energy based on the mechanical information is constant and a plurality of the actions can be executed,
By solving a combinatorial optimization problem that minimizes the prediction error with respect to the target value of each of the functional units and maximizes the energy required for the plurality of actions, the action to be executed with priority is determined. Information processing equipment. A method for causing an information processing device to simulate intellectual behavior, comprising:
a step of learning behavioral patterns related to each of a plurality of different functions of the brain;
obtaining electrical information;
a step of obtaining mechanical information, which is a mechanical quantity obtained by performing electromagnetic analysis of the entire brain when the acquired electrical information is input to the entire brain integrated with the plurality of functions; ,
associating the mechanical information with weighted values of the plurality of features combined;
A method. A computer program,
a step of learning behavioral patterns relating to each of a plurality of different functions of the brain;
obtaining electrical information;
a step of determining mechanical information, which is a mechanical quantity obtained by performing electromagnetic analysis of the entire brain when the acquired electrical information is input to the entire brain integrated with the plurality of functions; ,
associating the mechanical information with weight values for the plurality of combined functions;
A computer program that causes a computer to execute a storage medium,
A storage medium storing the computer program according to claim 9 .

Images