Artificial neural network learning device and learning method

Abstract

In an artificial neural network comprising a plurality of neurons and a plurality of synapses that connect the neurons in a single direction, wherein the neurons fire when a potential satisfies a condition that is indicated by a parameter including a firing threshold value and a time, and the synapses have the property of changing information transmission efficiency in accordance with the firing timing of each neuron connected to both ends thereof, one aspect of the present invention: monitors generation of a neuron group composed of specific neurons that are from among the plurality of neurons, that have fired within a prescribed period from an arbitrary time, and that are connected by synapses at a preset number of stages or fewer; acquires, for the generated neuron group, time change information pertaining to the number of specific neurons constituting the neuron group and time change information pertaining to information transmission efficiency of synapses connecting the specific neurons; and calculates the completeness of a concept represented by the neuron group on the basis of the acquired time change information pertaining to the number of specific neurons and the time change information pertaining to the information transmission efficiency of the synapses.

Description

Artificial neural network learning device and learning method

[0001] One aspect of the present invention relates to a learning device and a learning method for an artificial neural network that realizes the function of the brain of a living creature such as a human being.

[0002] A neural network that constitutes artificial intelligence is composed of multiple nodes and edges connecting these nodes, as described in Patent Document 1, for example. Each node has an internal potential, a firing threshold, and a reset value. When the potential exceeds the firing threshold in response to input from other nodes inside or outside the network, the node fires, and the firing bit is transmitted to other nodes via the edges. At this time, because edges have plasticity, the information transmission efficiency of the edges is autonomously strengthened or weakened. A group of nodes that are tightly coupled due to changes in the information transmission efficiency of the edges is defined as a node assembly.

[0003] Specifically, the measure of the strength of a group of tightly coupled nodes is time. A specific threshold time (for example, 10 milliseconds) is set, and a node assembly is a set of element nodes included in a node set that fire within the threshold time from a certain arbitrary time, and that are connected to other element nodes by a single edge. A state in which one or more nodes in a node assembly are firing is expressed as the activation of the node assembly. When multiple inputs are repeatedly given to the network, a node assembly that is likely to be activated in response to a certain input is generated. The generated node assembly represents a concept related to the input.

[0004] Node assemblies not only receive input but also output to other nodes. For example, input A from outside the network and input from node assembly B within the network change the information transmission efficiency w of the edge, and a new node assembly C is generated. This node assembly C represents a concept related to the concept represented by input A from outside the network and the concept represented by another node assembly B within the network. The concept represented by node assembly C is a new concept, and this is how the network generates concepts on its own.

[0005] Furthermore, after a node assembly is generated, the information transmission efficiency w of the edges may change due to inputs from outside the network and from other node assemblies within the network, and the number of nodes included in the existing node assembly may change. This represents the expansion or contraction of a concept. The expansion or contraction of a concept is one way of changing a concept and helping to create a new concept.

[0006] On the other hand, when a network of nodes includes a recursive circuit, a state occurs in which nodes within the network continue to fire even without input from outside the network. A state in which nodes continue to fire without input is called a self-propelled network. There can be multiple paths for nodes to continue firing. When the network continues to run self-propelled, multiple node assemblies within the network, for example D and E, become active, and the output of these node assemblies D and E changes the information transmission efficiency w of the edge, resulting in the generation of a new node assembly F. The concept represented by this node assembly F is a new concept, and this is the mechanism by which new concepts are created without input.

[0007] Furthermore, as the network continues to operate autonomously, the information transmission efficiency w of the edges between nodes may change, and the boundaries of existing node assemblies may change. This represents the expansion or contraction of a concept. The expansion or contraction of a concept helps to change the concept and generate new concepts. This is how the network expands and contracts concepts without input, helping to generate new concepts.

[0008] The above mechanism enables the network to perform operations similar to those of the human brain, which can make new discoveries and inventions while asleep.

[0009] International Publication No. 2024 / 069720

[0010] However, the technology described in Patent Document 1 can provide information on whether a node assembly group representing a concept has been generated, but cannot provide information on the extent to which a node assembly group representing a concept has been generated.

[0011] The present invention has been made in light of the above circumstances, and aims to provide a technique that makes it possible to know the degree of completion of a concept being generated.

[0012] In order to solve the above problems, a first aspect of the learning device or learning method for an artificial neural network according to the present invention is directed to an artificial neural network comprising a plurality of neurons and a plurality of synapses unidirectionally connecting the neurons, the neurons firing when the potential satisfies a condition expressed by parameters including a firing threshold and time, and the synapses having the property that the efficiency of information transmission changes depending on the firing timing of each neuron connected to both ends.

[0013] Then, the generation of a neuron group consisting of specific neurons from the plurality of neurons that fire within a predetermined time from an arbitrary time and are connected by the synapses within a predetermined number of stages is determined, and for the generated neuron group, time change information on the number of the specific neurons and time change information on the information transmission efficiency of the synapses connecting the specific neurons are obtained, and the degree of completion of the concept represented by the neuron group is calculated based on the time change information on the number of the specific neurons and the time change information on the information transmission efficiency of the synapses.

[0014] According to the first aspect of the present invention, it is possible to know the degree of completion of a concept being generated from the time change in the number of specific neurons in a neuron group representing a concept and the time change in the information transmission efficiency of synapses.

[0015] In a second aspect of the present invention, when the artificial neural network includes a reward-generating neuron having the function of generating a reward and providing it to a synapse, the amount of reward provided by the reward-generating neuron to each synapse connecting a plurality of the neurons is controlled based on the calculated degree of completion of the concept so as to increase the degree of completion of the concept. For example, the reward is increased when the degree of completion of the concept is low, and decreased when the degree of completion of the concept is sufficiently high.

[0016] According to a second aspect of the present invention, by adjusting the amount of reward given to synapses based on the calculated concept completeness, it is possible to construct a higher performance artificial neural network. Here, high performance refers to a large number of concepts and a high degree of completeness for each concept. Furthermore, by using information representing the completeness of concepts, it is possible to quantify the performance of the artificial intelligence, in other words, its intelligence.

[0017] That is, according to this aspect of the present invention, it is possible to provide a technology that makes it possible to know the degree of completion of a concept being generated.

[0018] FIG. 1 is a block diagram showing an example of the functional configuration of an artificial neural network learning device according to an embodiment of the present invention. FIG. 2 is a block diagram showing the functional configuration of a control unit of the artificial neural network learning device shown in FIG. 1. FIG. 3 is a flowchart showing an example of the procedure and content of the learning process executed by the control unit shown in FIG. 2. FIG. 4 is a flowchart showing an example of the procedure and content of the neuron list management process included in the learning process shown in FIG. 3. FIG. 5 is a flowchart showing an example of the procedure and content of the neuron flag management process included in the learning process shown in FIG. 3. FIG. 6 is a flowchart showing an example of the procedure and content of the neuron group generation determination process included in the learning process shown in FIG. 3. FIG. 7 is a flowchart showing an example of the procedure and content of the dopamine reward control process included in the learning process shown in FIG. 3. FIG. 8 is a flowchart showing an example of the procedure and content of the dopamine reward level check process included in the dopamine reward control process shown in FIG. 7. FIG. 9 is a diagram showing a state in which no concepts have been generated in the network. FIG. 10 is a diagram showing a state in which one concept has been generated in the network. FIG. 11 is a diagram showing a state in which the dopamine reward for the concept shown in FIG. 11 has been reduced.

[0019] Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

[0020] In one embodiment of the present invention, to clarify the correspondence of artificial neural networks to biological brains, the nodes and edges used in typical neural networks are referred to as neurons and synapses, respectively, and conceptual node assemblies are referred to as neuron populations.

[0021] (Configuration Example) FIG. 1 is a block diagram showing an example of the functional configuration of an artificial neural network learning device according to an embodiment of the present invention.

[0022] An artificial neural network learning device according to one embodiment includes a neural network NW and a control unit CS that controls learning in the neural network NW.

[0023] The neural network NW comprises, as its basic components, a plurality of general neurons NU1, ..., NU6, at least one reward-generating neuron NU7, and a plurality of synapses SY12, SY23, ..., SY76 that unidirectionally connect these neurons NU1, ..., NU7.

[0024] The internal state v of a general neuron NU1, ..., NU6 is expressed as the sum of a predetermined function with the internal state v and time t as parameters and the sum of the energy of all synapses SY12, SY23, ..., SY76 input to the general neuron NU1, ..., NU6.

[0025] The external states of the general neurons NU1, ..., NU6 are controlled by an external state control function provided in the neural network NW. The external state control function controls the external states of the general neurons NU1, ..., NU6 so that they are either fired or not fired, depending on the result of comparing the internal state v of the general neurons NU1, ..., NU6 with a predetermined threshold.

[0026] The information transmission efficiency w of the synapses SY12, SY23, ..., SY76 is controlled by a function of the neural network NW that controls the information transmission efficiency w. That is, the function that controls the information transmission efficiency w changes the information transmission efficiency w of the synapses SY12, SY23, ..., SY76 according to the difference in firing time between the neurons connected to both ends of the synapse.

[0027] The control unit CS includes, for example, a hardware processor constituting a central processing unit (CPU), a program storage unit, and a data storage unit. The program storage unit stores middleware such as an operating system (OS), as well as application programs required for the processor to execute various processes related to learning. The data storage unit stores various data required for the processor to execute the learning process.

[0028] FIG. 2 is a block diagram showing an example of the configuration of the processing functions executed by the control unit CS.

[0029] The control unit CS has, as processing functions related to the learning process of one embodiment, a neuron list management unit 11, a neuron flag management unit 12, a neuron group generation judgment processing unit 13, a neuron number acquisition processing unit 14, an information transmission efficiency acquisition processing unit 15, a concept completion calculation processing unit 16, and a dopamine reward control processing unit 17. The control unit CS also has a neuron list storage unit 21 and a neuron flag storage unit 22 in the data storage unit.

[0030] The neuron list management unit 11 generates list information for each general neuron NU1, ..., NU6, showing a list of neurons connected to the general neuron NU1, ..., NU6 via synapses within a predetermined number of stages, and stores the generated list information in the neuron list storage unit 21. The neuron list management unit 11 also updates the list information only when the information transmission efficiency of the synapses connected to the neurons included in the list information changes.

[0031] The neuron flag management unit 12 generates flag information for each of the general neurons NU1, ..., NU6, including information indicating whether the neuron fired within the predetermined time and information indicating the firing time, and stores the generated flag information in the neuron flag storage unit 22. The neuron flag management unit 12 also updates the flag information every predetermined time.

[0032] The neuron group generation determination processor 13 first refers to the list information stored in the neuron list storage unit 21 to detect a set of neurons connected within a predetermined number of stages from among the general neurons NU1, ..., NU6. The neuron group generation determination processor 13 then refers to the flag information stored in the neuron flag storage unit 22 to detect a neuron group that has fired within a predetermined time from an arbitrary time from the detected set of neurons. If the neuron group is detected, the neuron group generation determination processor 13 determines that a neuron group representing a concept has been generated.

[0033] When a neuron group representing the above concept is generated, the neuron number acquisition processing unit 14 obtains the time change rate of the number of neurons constituting the neuron group over a predetermined period of time.

[0034] When a neuron group representing the above concept is generated, the information transmission efficiency acquisition processing unit 15 calculates the time change rate of the information transmission efficiency at the synapses connecting each neuron that makes up the neuron group over the above specified time.

[0035] The concept completion calculation processing unit 16 calculates the completion of the concept represented by the neuron group using the time change rate of the number of neurons and the time change rate of the information transmission efficiency. An example of the calculation operation of the concept completion will be described in the operation example.

[0036] Based on the calculated concept completion, the dopamine reward control processing unit 17 controls the amount of reward that the reward-generating neuron NU7 provides to the synapses SY12, SY23, ... connecting the general neurons NU1, ..., NU6. An example of the reward amount control process will also be described below with an operational example.

[0037] Each of the processing units 11 to 17 is realized by causing a hardware processor to execute an application program stored in a program storage unit. Some or all of the processing units 11 to 17 may be realized using hardware such as an LSI (Large Scale Integration) or an ASIC (Application Specific Integrated Circuit).

[0038] (Example of Operation) Next, an example of operation of the artificial neural network learning device configured as above will be described.

[0039] To calculate the number of concepts generated at any given time and the degree of completion of the generated concepts, calculations must be performed using the presence or absence of firing of all neurons and the information transmission efficiency w of all synapses at a specified time interval from a past time to the given time.

[0040] However, this process requires a massive amount of computation. For example, if there are 10,000 neurons, the number of synapses connecting these neurons is 10,000 x 10,000 = 100 million. If the calculation interval is set to, say, 1 millisecond, shorter than the 5 milliseconds between the firings of a single neuron in a biological brain, the total number of synapses per second would be 100 million x 1,000 = 100 billion. Furthermore, to build an advanced AI system closer to the human brain, 10,000 neurons would be insufficient; hundreds of millions or even trillions would be required. In this case, the number of synapses, which is the square of the number of neurons, would be astronomically large. As a result, determining whether a concept has been generated and calculating the completeness of the generated concept would require a massive amount of computation.

[0041] In addition, in order to utilize the concept completion rate in real time during the learning of the neural network NW, the calculation of the concept completion rate must be performed in real time, which makes the calculation extremely difficult.

[0042] Therefore, in one embodiment of the present invention, we propose an algorithm that can reduce the amount of calculation required to determine whether a neuron group representing a concept has been generated and to calculate the degree of completion of the generated concept.

[0043] FIG. 3 is a flowchart showing an example of the procedure and content of the learning process executed by the control unit CS of the artificial neural network learning device.

[0044] (1) Managing the Neuron List To determine whether a group of neurons representing a concept has been generated, it is necessary to search and collect neurons that have fired within a predetermined threshold time from an arbitrary time from all neurons NU1, ..., NU6 that make up the neural network NW, and check whether these neurons are connected by synapses within a predetermined number of stages.

[0045] In a typical neural network, information representing whether a neuron is firing or not, expressed as "0" or "1," is stored in a list the length of which is the number of neurons. Furthermore, the synaptic information transmission efficiency is stored as a real value in a matrix of the number of neurons x the number of neurons. However, simply searching through the above list and matrix to collect the necessary values ​​requires a large amount of processing time and effort.

[0046] Therefore, first, in step S10, the control unit CS, under the control of the neuron list management unit 11, manages a neuron list specialized for the process of determining whether or not a concept has been generated as follows.

[0047] That is, the neuron list management unit 11 first focuses on one neuron at a certain time, detects other neurons connected to this neuron within a predetermined number of stages, and stores them in the neuron list. For example, using the initial value of the information transmission efficiency w of all synapses of the neural network NW, if the information transmission efficiency w is 0, it is considered not connected, and if it is other than 0, it is considered connected. Then, the neuron that is determined to be connected is stored in the list.

[0048] The neuron list management unit 11 then updates the generated neuron list in accordance with the value of the synaptic information transmission efficiency w as follows:

[0049] FIG. 4 is a flowchart showing an example of the processing procedure and processing contents of the list update processing executed by the neuron list management unit 11.

[0050] That is, the neuron list management unit 11 first acquires the synaptic information transmission efficiency w in step S11. At this time, if the neuron list is to be updated, for example, every 1 millisecond during the operation of the neural network NW, it takes too much time to acquire the information transmission efficiency w by referring to the matrix of the information transmission efficiency w.

[0051] Therefore, one of the functions of the neural network NW is to use a function that changes the synaptic information transmission efficiency w by the difference in firing time between the neurons at both ends of the synapse. That is, in step S11, the neuron list management unit 11 acquires the information transmission efficiency w from the function that changes the synaptic information transmission efficiency w, and determines in step S12 whether the acquired information transmission efficiency w has changed. Then, in step S13, the neuron list management unit 11 updates the neuron list only when the information transmission efficiency w has changed based on the above determination result.

[0052] For example, the neuron list management unit 11 issues a signal when the information transmission efficiency w changes to 0 and when it changes from 0 to a value other than 0. Then, only when a signal is issued, the neuron list management unit 11 updates the list relating to the neurons at both ends of the target synapse.

[0053] By generating and managing the neuron list as described above, the number of times the neuron list is updated is significantly reduced, thereby making it possible to reduce the processing load and time required for managing the list.

[0054] (2) Neuron Flag Management To determine whether a neuron group has been generated, it is necessary to know whether the neuron has fired within the period from a predetermined threshold time before to the current time. However, searching the neuron list every millisecond, for example, would result in a very large processing load.

[0055] Therefore, the control unit CS prepares a flag indicating the firing state of each neuron, and manages information indicating whether the neuron has fired from a predetermined threshold time before to the present time using this neuron flag.

[0056] FIG. 5 is a flowchart showing an example of the processing procedure and processing contents of the neuron flag management processing executed by the neuron flag management unit 12.

[0057] That is, the neuron flag management unit 12 first acquires the firing information for each neuron in step S21. To acquire this firing information, for example, a function provided in the neural network NW is used, which determines whether the external state of the neuron is firing or not by comparing the internal state v with a predetermined threshold.

[0058] The neuron flag management unit 12 determines whether the neuron has fired or not fired for each neuron based on the acquired firing information in steps S22 and S25. If the neuron has fired as a result of this determination, the neuron flag management unit 12 sets the flag to "true" in step S23 and stores the firing time in association with the flag in step S24. On the other hand, if the neuron has not fired, the neuron flag management unit 12 sets the flag to "false" in step S26 and stores the time when the neuron changed to "not fired" in association with the flag in step S27.

[0059] In other words, the neuron flag management unit 12 updates the flag information only when a neuron fires or changes to non-firing, and does not update it when the neuron does not change, which makes it possible to keep the processing load required for managing the neuron flag low.

[0060] (3) Determination of whether a neuron group representing a concept has been generated In step S30, the control unit CS, under the control of the neuron group generation determination processing unit 13, determines whether a neuron group representing a concept has been generated as follows.

[0061] FIG. 6 is a flowchart showing an example of the processing procedure and processing contents of the determination processing executed by the neuron group generation determination processing unit 13.

[0062] First, in step S31, the neuron group generation determination processing unit 13 reads the neuron list from the neuron list storage unit 21. Then, in step S32, the neuron group generation determination processing unit 13 refers to the neuron list and searches, for each neuron, for all sets of firing neurons that are connected to the neuron via synapses within a predetermined number of stages.

[0063] Next, in step S33, the neuron group generation determination processing unit 13 reads the neuron flag from the neuron flag storage unit 22. Then, in step S34, the neuron group generation determination processing unit 13 refers to the read neuron flag and searches for neurons that fired during a period from a predetermined threshold time to the current time, that is, neuron groups that fired almost simultaneously, from the previously searched set of firing neurons. Then, in step S35, the searched neuron groups are determined to be neuron groups that represent a concept.

[0064] In other words, in one embodiment, by determining the generation of a neuron group representing a concept by referring to the neuron list and neuron flag, it is possible to significantly reduce the processing load required to determine the generation of a neuron group, even when taking into account the processing load required for managing the neuron list and neuron flag, compared to checking the firing state and number of stages relative to other neurons for all neurons at 1 millisecond intervals each time.

[0065] (4) Determining the completeness of the concept To calculate the completeness of the concept represented by the generated neuron group, it is necessary to determine the time change rate of the number of neurons constituting the neuron group and the time change rate of the information transmission efficiency of the synapses connecting each neuron constituting the neuron group.

[0066] Therefore, in step S40, the control unit CS first detects the number of neurons constituting the neuron group representing the above concept under the control of the neuron number acquisition processing unit 14. Then, it calculates the time change rate of the detected number of neurons for the period from a predetermined threshold time before to the current time.

[0067] In step S50, the control unit CS detects synapses connecting the neurons that make up the neuron group representing the above concept under the control of the information transmission efficiency acquisition processing unit 15. Then, it calculates the time change rate of the information transmission efficiency of each detected synapse for the period from the predetermined threshold time before the current time.

[0068] Next, in step S60, the control unit CS, under the control of the concept completion calculation processing unit 16, calculates the completion of the concept based on the time change rate of the number of neurons calculated by the neuron number acquisition processing unit 14 and the time change rate of the information transmission efficiency calculated by the information transmission efficiency acquisition processing unit 15.

[0069] For example, the concept completeness calculation processing unit 16 defines a function such that the closer the time rate of change of the number of neurons and the time rate of change of the information transmission efficiency are to 0, the higher the concept completeness. Now, if the number of neurons is n and the information transmission efficiency is w, the calculation formula for concept completeness using the above function is expressed as follows: Concept completeness = 1 / |w'| + 1 / |n'| Note that ' in the formula represents time differentiation. The range of this formula is from 0 to infinity, and the maximum value of the concept completeness is infinity.

[0070] (5) Control of Dopamine Reward The number of neuron groups representing concepts determined to have been generated by the neuron group generation determination processing unit 13 and the concept completeness of each neuron group calculated by the concept completeness calculation processing unit 16 represent the performance of the neural network NW as an artificial intelligence. That is, the more concepts there are in the neural network NW and the higher the completeness of each concept, the higher the performance of the artificial intelligence can be considered to be.

[0071] In order to further improve the performance of the neural network NW as artificial intelligence, in step S70, the control unit CS controls the dopamine reward given to the synapses of the neural network NW under the control of the dopamine reward control processing unit 17 as follows.

[0072] FIG. 7 is a flowchart showing an example of the processing procedure and processing content of the control processing executed by the dopamine reward control processing unit 17.

[0073] That is, the dopamine reward control processing unit 17 first obtains the number of neuron groups representing concepts determined to have been generated by the neuron group generation determination processing unit 13 and the concept completion rate of each neuron group calculated by the concept completion rate calculation processing unit 16. Then, in step S71, the dopamine reward control processing unit 17 determines whether a sufficient number of concepts have been generated as desired. If the result of this determination is that a sufficient number of concepts have not been generated as desired, in step S73, the dopamine reward control processing unit 17 increases the dopamine reward value to be provided to each synapse from the reward-generating neuron NU7.

[0074] Furthermore, the dopamine reward control processing unit 17 adjusts the increase in the dopamine reward value in step S74 as follows.

[0075] FIG. 8 is a flowchart showing an example of the processing procedure and processing content of the dopamine reward value check processing executed by the dopamine reward control processing unit 17.

[0076] That is, the dopamine reward control processing unit 17 determines in step S741 whether the dopamine reward value is greater than a preset maximum value, and if it is greater than the maximum value, sets the dopamine reward value to the maximum value in step S742. This controls the dopamine reward value so that it does not exceed the upper limit value.

[0077] 9 shows a state in which a group of neurons representing a concept has not yet been generated within the neural network NW. In this example, the dopamine reward control processor 17 increases the dopamine reward value given by the reward-generating neuron NU7 to each synapse within the neural network NW, with the maximum value as the upper limit.

[0078] In Fig. 9, the magnitude of the dopamine reward value is indicated by the thickness of the arrow. For simplicity of illustration, Fig. 9 shows dopamine reward as being given only to some synapses SY12 and SY52, but in reality, it is given to all synapses.

[0079] As a result of increasing the dopamine reward value, a neuron group representing a concept is generated within the neural network NW. In this case, the dopamine reward control processing unit 17 determines in step S71 whether the number of generated concepts has reached a desired sufficient number based on the neuron group generation determination result by the neuron group generation determination processing unit 13. At the same time, the dopamine reward control processing unit 17 determines whether the concept completion level represented by the generated neuron group is equal to or greater than a threshold value, for example, based on the concept completion level of the neuron group calculated by the concept completion level calculation processing unit 16.

[0080] If the result of the above judgment is that the number of generated concepts does not reach the desired sufficient number, or if it does reach the desired number but the concept completion level does not reach the threshold, then in step S73, the dopamine reward control processing unit 17 further increases the dopamine reward value provided to each synapse from the reward-generating neuron NU7 by a predetermined unit amount.

[0081] 10 shows the state in which a neuron group representing a concept, consisting of neurons NU1', NU2', and NU3', is generated by increasing the dopamine reward value. However, in this example, although the neuron group is generated, its concept completion does not reach the threshold, so the increased dopamine reward value is continuously given to the synapses.

[0082] However, if dopamine reward continues to increase, the number of neurons that make up a concept will not stabilize due to large changes in synaptic information transmission efficiency w, and therefore the level of concept completion will not increase any further.

[0083] Therefore, in step S71, when the dopamine reward control processing unit 17 determines that the number of neuron groups representing a concept has reached the desired sufficient number and that the degree of concept completion has reached a predetermined threshold, the dopamine reward control processing unit 17 reduces the dopamine reward value given to each synapse from the reward-generating neuron NU7 by a predetermined unit amount in step S72.

[0084] At this time, the dopamine reward control processing unit 17 determines in step S743 whether the dopamine reward value is smaller than a preset minimum value, and if it is smaller than the minimum value, sets the dopamine reward value to the minimum value in step S744. This controls the dopamine reward value so that it does not fall below the lower limit value.

[0085] 11 shows the state in which the dopamine reward value is decreased. In this example, the decreased dopamine reward value is indicated by a thin arrow.

[0086] As described above, when the dopamine reward value provided to each synapse from the reward-generating neuron NU7 decreases, the change in synaptic information transmission efficiency w stabilizes, and the number of neurons in the neuron group that constitutes the concept stabilizes. As a result, the performance of the artificial intelligence constituted by the neural network NW remains high.

[0087] As described above, in one embodiment, for each neuron, a list of neurons connected to the neuron via synapses within a predetermined number of stages and a flag indicating the firing / non-firing state of each neuron from a predetermined threshold time before to the present time are referenced to determine whether a neuron group representing a concept has been generated in the neural network NW. Then, the degree of completion of the concept represented by the neuron group is calculated based on the time rate of change in the number of neurons constituting the generated neuron group and the time rate of change in the information transmission efficiency of the synapses connecting the neurons constituting the neuron group.

[0088] Therefore, it is possible to know not only whether a neuron group representing a concept has been generated in the neural network NW, but also the degree of completion of the concept represented by the generated neuron group. In addition, the process of determining whether a neuron group representing the concept has been generated can be performed with a small processing load by referring to the neuron list and flag.

[0089] Furthermore, the neuron list is updated only when the synaptic information transmission efficiency changes, and the neuron flag is updated only when the neuron fires or goes into non-firing state within a predetermined period, which reduces the processing load for managing the neuron list and neuron flag.

[0090] Furthermore, based on the result of determining whether a neuron group representing a concept has been generated and the result of calculating the degree of completion of the generated concept, the dopamine reward value given from the reward-generating neuron NU7 to each synapse in the network NW is controlled so that it is increased if the degree of completion of the concept is low, and decreased if the degree of completion of the concept reaches a threshold. This makes it possible to efficiently generate a neuron group representing a concept in the neural network NW and to increase its degree of completion, thereby effectively improving the performance of the neural network NW as an artificial intelligence.

[0091] Other Embodiments (1) In one embodiment, the dopamine reward value given to a synapse is controlled based on the number of concepts generated and their degree of completion. However, the present invention is not limited to this, and learning strategies other than reward may be changed using the degree of completion of the concept. For example, when the degree of completion of a concept exceeds a threshold, the function that changes the synaptic information transmission efficiency w based on the difference in firing times of the neurons at both ends of the synapse may be stopped. In this way, the neural network NW can be prevented from performing new learning.

[0092] (2) When the size of the neural network NW and the neuron connections can be varied, the size of the neural network NW and the neuron connections can be adjusted based on the calculation results of the concept completeness. Concepts are more easily generated the larger the size of the neural network NW, and the stronger the neuron connections, i.e., the more neurons are connected to other neurons and the greater the synaptic information transmission efficiency w. Therefore, if the concept completeness is lower than desired, the size of the neural network NW can be increased or the neuron connections can be strengthened until the concept completeness reaches the desired level. In this way, it is possible to automatically determine the size of the network NW required depending on the intended use.

[0093] (3) In addition, various modifications can be made to the hardware configuration of the control unit CS, the processing procedures and processing contents of each process performed by each processing unit provided in the control unit CS, the number of neurons and synapses that make up the neural network NW and how they are connected, etc., without departing from the spirit of this invention.

[0094] Although the embodiments of the present invention have been described in detail above, the above description is merely an example of the present invention in every respect. It goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. In other words, when implementing the present invention, specific configurations according to the embodiments may be appropriately adopted.

[0095] In short, this invention is not limited to the above-described embodiments, and in the implementation stage, the components can be modified and embodied without departing from the spirit of the invention. Furthermore, various inventions can be formed by appropriately combining multiple components disclosed in the above-described embodiments. For example, some components may be omitted from all the components shown in the embodiments. Furthermore, components from different embodiments may be appropriately combined.

[0096] NW... neural network CS... control unit NU1 to NU6... general neurons NU1' to NU3'... neurons that constitute concepts NU7... reward-generating neuron 11... neuron list management unit 12... neuron flag management unit 13... neuron group generation determination processing unit 14... neuron number acquisition processing unit 15... information transmission efficiency acquisition processing unit 16... concept completion calculation processing unit 17... dopamine reward control processing unit 21... neuron list storage unit 22... neuron flag storage unit

Claims

1. A learning device for use in an artificial neural network, comprising a plurality of neurons and a plurality of synapses unidirectionally connecting the neurons, wherein the neurons fire when their potentials satisfy conditions expressed by parameters including a firing threshold and time, and the synapses have the property of changing their information transmission efficiency depending on the firing timing of the neurons connected to both ends, the device comprising: a first processing unit that determines whether to generate a neuron group composed of specific neurons from the plurality of neurons that fire within a predetermined time from an arbitrary time and are connected via the synapses within a predetermined number of stages; a second processing unit that, when the neuron group is generated, acquires time-varying information on the number of the specific neurons that compose the neuron group and time-varying information on the information transmission efficiency of the synapses connecting the specific neurons that compose the neuron group; and a third processing unit that calculates the degree of completion of the concept represented by the neuron group based on the time-varying information on the number of the specific neurons and the time-varying information on the information transmission efficiency of the synapses.

2. The artificial neural network learning device according to claim 1, wherein the first processing unit comprises: a processing unit that detects, from among the plurality of neurons, a set of neurons connected to the neuron within the predetermined number of stages, by referring to list information that holds, for each neuron, a list of neurons connected to the neuron within the predetermined number of stages and that is updated only when the information transmission efficiency of the synapse changes; and a processing unit that detects, from the set of neurons, a neuron that has fired within the predetermined time, by referring to flag information that holds, for each neuron, information indicating whether the neuron has fired within the predetermined time, and that is updated every predetermined time.

3. The artificial neural network learning device according to claim 1, further comprising a fourth processing unit that, when said artificial neural network comprises a reward-generating neuron having the function of generating a reward and providing it to said synapse, controls the amount of reward provided by said reward-generating neuron to said synapse connecting a plurality of said neurons based on the degree of completion of said concept calculated by said third processing unit, so as to increase the degree of completion of said concept.

4. A learning method used in an artificial neural network comprising a plurality of neurons and a plurality of synapses unidirectionally connecting the neurons, the neurons firing when their potentials satisfy conditions expressed by parameters including a firing threshold and time, and the synapses having the property of changing their information transmission efficiency depending on the firing timing of the neurons connected to both ends, the method comprising the steps of: determining the generation of a neuron group consisting of specific neurons from the plurality of neurons that fire within a predetermined time from an arbitrary time and are connected via the synapses within a predetermined number of stages; when the neuron group is generated, acquiring time-varying information on the number of the specific neurons constituting the neuron group and time-varying information on the information transmission efficiency of the synapses connecting the specific neurons constituting the neuron group; and calculating the degree of completion of the concept represented by the neuron group based on the time-varying information on the number of the specific neurons and the time-varying information on the information transmission efficiency of the synapses.

Images