A method and apparatus for solving catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain.

Abstract

This disclosure provides a method and a computing device for operating a computing device that applies the metaplasticity rules of the biological brain to prevent catastrophic forgetting in artificial neural network learning. [Solution] The computing device method involves randomly assigning different variability values ​​to multiple synapses of an artificial neural network, and sequentially learning for multiple tasks using the artificial neural network while storing information for each task via the synapses. Specifically, while learning for each task progresses, the strength of each synapse is adjusted according to the variability value assigned to each synapse, and task information based on the strength is stored via the synapses.

Description

[Technical Field]

[0001] This disclosure relates to a method and apparatus for resolving catastrophic forgetting in artificial neural network learning by applying metaplasticity rules of the biological brain. [Background technology]

[0002] In various fields of artificial intelligence that utilize deep neural networks, such as AI assistants and large-scale language models, it is crucial to store previously learned information while simultaneously learning and applying newly given information. However, when neural networks learn new information, they can sometimes suffer from catastrophic forgetting, a problem where they forget previously learned information.

[0003] Several methods have been proposed to solve this problem. The first method attempted to solve catastrophic forgetting by storing information about the learning task in synapses and selectively saving the connection strength. However, this required a large amount of computation because the connection strength of past synapses had to be readjusted every time new information was learned. The second method proposed widening the neural network or adding external memory when the amount of new input information increased, but this was difficult to apply to the realistic situation where the capacity of the neural network cannot be arbitrarily changed during the learning process. The third method introduced additional processes such as determining whether the information to be learned is necessary before storing it as long-term memory, but this was difficult to apply to actual work in terms of learning efficiency and speed. The fourth method presented the principle that it is possible to fluidly store information in a neural network with a mixture of various types of synapses. However, this was a model based on biological spike neural networks and could not be applied to convolutional neural networks used for practical tasks such as recognition / identification.

[0004] Most conventional techniques have focused solely on the goal of memorizing learned information with high accuracy. However, the actual brain exhibits a fluid information memory ability, sacrificing accuracy to secure memory capacity depending on the situation, rather than memorizing all learned information with perfect accuracy. Various biological information memory characteristics observed in the brain, such as the ability to fluidly adjust accuracy and memory capacity, the ability to strengthen memories to avoid the effects of contaminated information and noise through repeated learning, or the ability to prioritize the memory of frequently repeated information over less frequently repeated information, are extremely effective in various tasks in which deep neural networks are actually used, but have not yet been fully utilized in general artificial neural network learning. [Overview of the project] [Problems that the invention aims to solve]

[0005] This disclosure provides a method and apparatus for resolving catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain. [Means for solving the problem]

[0006] According to this disclosure, a method of operating a computing device that applies the metaplasticity rules of the biological brain to resolve catastrophic forgetting in artificial neural network learning includes the steps of randomly assigning different variability values ​​to multiple synapses of an artificial neural network, storing information about at least one task via the synapses while the artificial neural network learns about at least one task, wherein the step of storing information about at least one task may include adjusting the strength of each synapse according to the variability values ​​assigned to each synapse while learning for each task, and storing task information via the synapses based on the strength.

[0007] According to this disclosure, a computing device that applies the metaplasticity rules of the biological brain to resolve catastrophic forgetting in artificial neural network learning includes memory and a processor configured to be linked to memory and to execute at least one instruction recorded in memory, to randomly assign different variability values ​​to multiple synapses of an artificial neural network, and to store information for at least one task via the synapses while the artificial neural network is learning for at least one task, the processor may be configured to adjust the strength of each synapse according to the variability value assigned to each synapse as learning for each task progresses, and to store task information based on the strength via the synapses.

[0008] A computer program recorded on a non - transient computer - readable recording medium for causing a computing device to execute a method for solving catastrophic forgetting in artificial neural network learning by applying the meta - plasticity rules of a biological brain. The method includes steps of randomly assigning different variability values to a plurality of synapses of an artificial neural network respectively, and while advancing learning for at least one task by the artificial neural network, storing information of at least one task through the synapses. The step of storing information of at least one task may include steps of adjusting the strength of each synapse according to the variability value respectively assigned to the synapse while advancing learning of each task, and storing information of the task based on the strength through the synapse.

Advantages of the Invention

[0009] According to the present disclosure, by advancing learning for a plurality of tasks using an artificial neural network to which meta - plasticity rules are applied, information at each stage can be stored with an accuracy above a certain level and stored up to the maximum storage capacity, and catastrophic forgetting in which information of past tasks is lost can be prevented. In the artificial neural network of the present disclosure, without adding any calculation process, the function of fluidly storing information can be automatically realized. The performance of the artificial neural network of the present disclosure is enhanced by iterative learning, and even if noisy or contaminated data is presented as a learning dataset later, the loss is negligible.

Brief Description of the Drawings

[0010] [Figure 1] It is a diagram schematically showing a computing device for solving catastrophic forgetting in artificial neural network learning by applying the meta - plasticity rules of a biological brain in the present disclosure. [Figure 2] It is a diagram for explaining the characteristics of the operation of the computing device in FIG. 1. [Figure 3]This is a diagram for explaining the performance of the computing device in FIG. 1. [Figure 4] This is a diagram for the first experiment comparing the artificial neural network of the present disclosure with a general artificial neural network and for explaining the results thereof. [Figure 5] This is a diagram for the second experiment comparing the artificial neural network of the present disclosure with a general artificial neural network and for explaining the results thereof. [Figure 6] This is a diagram for the third experiment comparing the artificial neural network of the present disclosure with a general artificial neural network and for explaining the results thereof. [Figure 7] This is a diagram for the fourth experiment comparing the artificial neural network of the present disclosure with a general artificial neural network and for explaining the results thereof. [Figure 8] This is a diagram for the fifth experiment comparing the artificial neural network of the present disclosure with a general artificial neural network and for explaining the results thereof. [Figure 9] This is a diagram schematically showing an operation method of a computing device that applies the meta - plasticity rule of a biological brain to solve catastrophic forgetting in artificial neural network learning in the present disclosure.

Embodiments for Carrying out the Invention

[0011] Hereinafter, the present disclosure provides a method and an apparatus for applying the meta - plasticity rule of a biological brain to solve catastrophic forgetting in artificial neural network learning.

[0012] This disclosure focuses on the characteristics of synaptic plasticity observed in the human brain and realizes a sequential learning method that can be universally applied to artificial neural networks without going through complex intermediate processes such as considering physical structural changes of neural networks or connectivity between information. Specifically, this disclosure shows that by applying synaptic metaplasticity rules to artificial neural networks, each synapse acquires different degrees of flexibility and stability, which minimizes catastrophic forgetting in sequential information learning and enables fluid information storage similar to human working memory. In short, by simply multiplying the general learning rules of an artificial neural network by a single simple metaplasticity function, it is possible to maintain the memory accuracy of each piece of information above a certain level, store information up to the maximum memory capacity, and suppress catastrophic forgetting in this process. An artificial neural network to which the technology of this disclosure is applied can automatically realize the fluid information storage function without any additional computational processes. The information stored in this way is strengthened by a small amount of iterative learning, and even if noisy or contaminated data is later presented as a training dataset, the loss is minimal.

[0013] This disclosure is applicable to various neural network models that mimic the brain, but it is also applicable to almost all types of neural network models currently in widespread use (such as AlexNet, ResNet, and large language models), making it general-purpose unlike conventional techniques. In other words, this disclosure proposes a general learning or information storage algorithm and is not limited to any particular type / structure of neural network or learning model. This is a model that is not only universally applicable to all kinds of neural network models, but can also be applied in the same way to hardware systems such as neuromorphic systems without any additional processes.

[0014] Various embodiments of this disclosure will be described below with reference to the accompanying drawings.

[0015] Figure 1 is a schematic diagram of the computing device 100 in this disclosure, which solves catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain. Figure 2 is a diagram illustrating the operational characteristics of the computing device 100 in Figure 1. Figure 3 is a diagram illustrating the performance of the computing device 100 in Figure 1.

[0016] Referring to Figure 1, the computing device 100 is for solving catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain, and may include at least one of a communication module 110, an input module 120, an output module 130, a memory 140, or a processor 150. In some embodiments, at least one of the components of the computing device 100 may be omitted, and at least one other component may be added. In some embodiments, at least two of the components of the computing device 100 may be implemented in a single integrated circuit.

[0017] The communication module 110 may perform communication with an external device on the computing device 100. The communication module 110 may establish a communication channel between the computing device 100 and the external device and perform communication with the external device via the communication channel. For example, the external device may include at least one of other computing devices, servers, base stations, or satellites. The communication module 110 may include at least one of a short-range communication module or a long-range communication module. The short-range communication module may communicate with the external device using a short-range communication method. For example, the short-range communication method may include at least one of Bluetooth, Wi-Fi Direct, or infrared data association (IrDA). The long-range communication module may communicate with the external device using a long-range communication method. Here, the long-range communication module may communicate with the external device via a network. For example, the network may include at least one of a cellular network, the Internet, or a computer network such as a LAN (local area network) or a WAN (wide area network).

[0018] The input module 120 may receive signals used for at least one component of the computing device 100. The input module 120 may be configured to detect signals directly input by a user or to detect changes in the surroundings and generate signals. For example, the input module 120 may include at least one of a mouse, a keypad, a microphone, or a sensing module having at least one sensor. In some embodiments, the input module 120 may include at least one of a touch circuitry configured to detect touches, or a sensor circuitry configured to measure the intensity of force generated by a touch.

[0019] The output module 130 may output information to the outside of the computing device 100. The output module 130 may include at least one of the following: a display module configured to output information visually, or an audio output module that outputs information as an audio signal. For example, the audio output module may include at least one of a speaker or a receiver.

[0020] Memory 140 may store various data used by at least one component of the computing device 100. For example, memory 140 may include at least one of volatile memory or non-volatile memory. The data may include at least one program and its associated input or output data. The program may be stored in memory 140 as software including at least one instruction, and may include at least one of an operating system, middleware, or application.

[0021] The processor 150 may execute a program in memory 140 and control at least one component of the computing device 100. This allows the processor 150 to perform data processing or calculations. At this time, the processor 150 may execute instructions stored in memory 140. In various embodiments, the processor 150 may apply the metaplasticity rules of the biological brain to resolve catastrophic forgetting in artificial neural network learning. The metaplasticity rules may be implemented in such a way that the strengths of multiple synapses in the artificial neural network are adjusted according to their respective variability.

[0022] The processor 150 may impart the adaptive sequential learning ability of humans by realizing a state in which the metaplasticity values, i.e., variability values ​​(flexibility), of multiple synapses in the artificial neural network are randomly mixed. Specifically, the processor 150 may assign different variability to each synapse of the artificial neural network. Here, the variability value may be within a range between a lower limit and an upper limit. For example, the lower limit may be 0 and the upper limit may be 1. The closer the variability value is to the lower limit, the more stable the corresponding synapse may be, such that its strength (w) does not change easily. The closer the variability value is to the upper limit, the more flexible the corresponding synapse may be, i.e., unstable, such that its strength (w) changes easily. Based on this, the processor 150 may sequentially advance learning for multiple tasks (e.g., Task 1, Task 2, ...) using the artificial neural network, and store information for each task via the synapses. Specifically, the processor 150 may, while learning for each task, adjust the synaptic strength (w) according to the variability value assigned to each synapse, and store task information based on the strength (w) via the synapses.

[0023] As shown in Figure 2, during learning for each task, the processor 150 may adjust the synaptic strength (w) in the following way: That is, the processor 150 adjusts the strength (w) of each synapse to a new strength (w) n+1 ) may be adjusted. For each synapse, the processor 150 adjusts the intensity range (△w) from the initial intensity (w0) before learning to a range according to the variability value. n+1 n=0, 1, 2, ... For example, △w1 may be determined for learning Task 1, and △w2 for learning Task 2. The closer the variability value is to the lower limit, the greater the intensity adjustment range (△w n+1 ) can be small, and the closer the variability value is to the upper limit, the wider the intensity adjustment range (△w n+1 ) can be large. Also, processor 150 has a variable value, previous intensity (w n ), and strength adjustment range (△w n+1 ) by combining new strengths (w n+1) may be determined. In some embodiments, the processor 150 detects a learning rate reduced from a previous learning rate based on a variability value and an intensity adjustment width (Δw n+1 ) and determines a new intensity (w n ) based on a combination of the previous intensity (w n+1 ) and the reduced learning rate. The closer the variability value is to the lower limit value, the greater the width by which the learning rate is reduced may be, and the closer the variability value (flexibility) is to the upper limit value, the smaller the width by which the learning rate is reduced may be. For example, the new intensity (w n+1 ) may be determined by the following equations (1) and (2).

[0024]

Equation

[0025] Here, w represents the synaptic strength 、 w n represents the previous strength 、 w n+1 represents the new strength, Δw n+1 represents the intensity adjustment width, n represents the previous learning iteration number, n + 1 represents the new learning iteration number, TIFF2026116654000003.tif54 represents the learning rate, J(·) represents the loss function, S(·) represents the learning rate reduction function, and may be determined as in the following equation (2).

[0026]

Equation

[0027] Here, flexibility represents the variability value, and α may be a hyperparameter that adjusts the width of S(·).

[0028] In various embodiments, by applying metaplasticity rules, artificial neural networks spontaneously exhibit learning phenomena that appear in the brain's working memory, such as the serial position effect and the Hebb repetition effect. That is, as shown in Figure 3, when learning multiple tasks, information for new tasks can be learned while storing information for past tasks. Also, as shown in Figure 3, the memory accuracy for each task improves with each repetition of learning for that task. In other words, metaplasticity rules enable artificial neural networks to adaptively utilize memory capacity in the same way as human fluid memory. Specifically, the processor 150 can use the artificial neural network to repeatedly advance learning according to the learning frequency set for each task, and store information via synapses according to the memory capacity of each task. In this case, if the same learning frequency is set for the tasks, the memory capacity is distributed equally among the tasks, but if different learning frequencies are set for the tasks, a portion of the memory capacity of the task with the lower learning frequency may be reallocated for the task with the higher learning frequency.

[0029] Figure 4 shows the first experiment to compare the artificial neural network of this disclosure with a general artificial neural network, and a diagram illustrating the results.

[0030] Referring to Figure 4, an experiment was conducted in which learning was sequentially performed on 10 tasks in order to compare the artificial neural network of this disclosure with a general artificial neural network. In this experiment, both the artificial neural network of this disclosure and a general artificial neural network, AlexNet, were used, and in the artificial neural network of this disclosure, a metaplasticity rule was applied to the fully connected layer. That is, a state was achieved in which the variability values ​​of the synapses in the fully connected layer were randomly mixed. Here, the variability values ​​may be within the range between 0 and 1. Furthermore, as shown in Figure 4, learning was sequentially performed on 10 tasks using the artificial neural network of this disclosure and a general artificial neural network. Here, a task sequence in which 10 tasks are defined sequentially was established, and the tasks were to classify different 2-digit images.

[0031] As a result, as shown in Figure 4, differences were observed between the artificial neural network of this disclosure and a general artificial neural network in terms of memory accuracy for each task and the number of tasks for which information was stored. Specifically, the general artificial neural network exhibited a catastrophic forgetting effect, strongly remembering information from recent tasks and forgetting information from past tasks to below the baseline. On the other hand, the artificial neural network of this disclosure exhibited a serial position effect, where the memory accuracy of past tasks was preserved, maintaining the memory accuracy of all 10 tasks above the baseline.

[0032] Figure 5 shows a second experiment comparing the artificial neural network of this disclosure with a general artificial neural network, and illustrates the results.

[0033] Referring to Figure 5, an experiment was conducted in which the number of tasks was increased while sequentially training on tasks was performed in order to compare the artificial neural network of this disclosure with a general artificial neural network. In this experiment, the artificial neural network of this disclosure and the general artificial neural network were configured in the same way as in the above experiment. Furthermore, using the artificial neural network of this disclosure and the general artificial neural network, sequential training on tasks was performed while increasing the number of tasks, as shown in Figure 5. Here, different task sequences were defined in which different numbers of tasks were defined sequentially, and the tasks were to classify different two-digit images.

[0034] As a result, in situations where the total number of tasks to be learned increases, as shown in Figure 5, differences were observed between the artificial neural network of this disclosure and a general artificial neural network in terms of the overall memory accuracy based on the number of tasks and the number of tasks for which information was stored. Specifically, in the artificial neural network of this disclosure, the overall storage accuracy based on the number of tasks remained at a constant level, while the number of tasks for which memory accuracy was stored above the baseline increased. In other words, in situations where the number of tasks to be learned is unknown in advance, this disclosure enables the artificial neural network to automatically redistribute memory capacity so that past tasks and new tasks are stored above the baseline.

[0035] Figure 6 shows a third experiment to compare the artificial neural network of this disclosure with a general artificial neural network, and illustrates the results.

[0036] Referring to Figure 6, in order to compare the artificial neural network of this disclosure with a general artificial neural network, we conducted an experiment in which we repeatedly trained on the same task sequence, in particular to improve the memory accuracy of tasks located in the middle of the task sequence in the artificial neural network of this disclosure. In this experiment, the artificial neural network of this disclosure and the general artificial neural network were configured in the same way as in the experiment described above. Furthermore, using the artificial neural network of this disclosure and the general artificial neural network, we repeated training on the same task sequence nine times, as shown in Figure 6. Here, a task sequence was defined in which 10 tasks were defined sequentially, and the tasks were to classify different two-digit images.

[0037] As a result, as shown in Figure 6, a difference was observed in the memory accuracy of tasks in a task sequence between the artificial neural network of this disclosure and a general artificial neural network. Specifically, the general artificial neural network did not show a significant enhancement effect of memory accuracy through repetition. On the other hand, the artificial neural network of this disclosure reproduced the Hebb repetition effect of task memory, and the memory accuracy of the tasks constituting the task sequence increased overall. In particular, the memory accuracy of the task located in the middle of the task sequence improved significantly. This means that by repeatedly learning the task sequence, the memory capacity of tasks that previously showed relatively low memory accuracy was reinforced, reducing the disparity in memory accuracy among the tasks constituting the task sequence and resulting in an even distribution of memory capacity among the tasks.

[0038] Figure 7 shows a fourth experiment comparing the artificial neural network of this disclosure with a general artificial neural network, and illustrates the results.

[0039] Referring to Figure 7, in order to compare the artificial neural network of this disclosure with a general artificial neural network, and in particular to demonstrate that the memory strengthened by iterative learning in the artificial neural network of this disclosure is not easily lost even when exposed to contaminated data, an experiment was conducted in which the same task sequence was trained nine times, and then training was performed on intentionally contaminated (incorrect) data (data poisoning attack). In this experiment, the artificial neural network of this disclosure and the general artificial neural network were configured in the same way as in the experiment described above. Furthermore, using the artificial neural network of this disclosure and the general artificial neural network, as shown in Figure 7, training on the same task sequence was repeated nine times, and then training was performed on contaminated data. Here, a task sequence in which 10 tasks are defined sequentially was defined, and the tasks were to classify different two-digit images, and contaminated data was generated based on this.

[0040] As a result, as shown in Figure 7, a difference was observed in the accuracy of task memory in a task sequence between the artificial neural network of this disclosure and a general artificial neural network. Specifically, in the general artificial neural network, after exposure to contaminated data, memory for all tasks was lost to the baseline level, and it was unable to remember any of the tasks constituting the task sequence. On the other hand, the artificial neural network of this disclosure remembered all tasks above the baseline level, even after exposure to contaminated data. This suggests that task information learned through iterative learning is not easily lost even when exposed to contaminated data.

[0041] Figure 8 shows the fifth experiment to compare the artificial neural network of this disclosure with a general artificial neural network, and illustrates the results.

[0042] Referring to Figure 8, an experiment was conducted to compare the artificial neural network of this disclosure with a general artificial neural network, in which different learning frequencies were set for each task and learning for the task was advanced according to the learning frequency. In this experiment, the artificial neural network of this disclosure and the general artificial neural network were configured in the same way as in the experiment described above. Furthermore, using the artificial neural network of this disclosure and the general artificial neural network, learning for tasks with different learning frequencies was advanced sequentially, as shown in Figure 8. Here, a task sequence was defined in which 10 tasks were defined sequentially, and the tasks were to classify images with different two-digit numbers.

[0043] As a result, as shown in Figure 8, the changes in memory accuracy differed between the artificial neural network of this disclosure and a general artificial neural network depending on the difference in the learning frequency of tasks in the task sequence. Specifically, in a general artificial neural network, the memory accuracy of each task showed a high correlation with its order within the task sequence, rather than its learning frequency. On the other hand, in the artificial neural network of this disclosure, the memory accuracy of tasks showed a high correlation with their learning frequency, rather than their order within the task sequence. In the artificial neural network of this disclosure, memory capacity is allocated according to the importance of the tasks, not the learning order of the tasks. That is, when there is a difference in learning frequency between tasks, the artificial neural network of this disclosure reallocates the memory capacity of tasks with low learning frequency to tasks with high learning frequency, selectively forgetting tasks with low learning frequency and strongly storing tasks with high learning frequency.

[0044] Figure 9 schematically illustrates how the computing device 100 described herein operates, which applies the metaplasticity rules of the biological brain to resolve catastrophic forgetting in artificial neural network learning.

[0045] Referring to Figure 9 along with Figure 1, in step 910, the computing device 100 may realize a state in which the metaplasticity values, i.e., variability values, of multiple synapses in the artificial neural network are randomly mixed. Specifically, the processor 150 may assign different variability values ​​to each synapse of the artificial neural network. Here, the variability value may be within a range between a lower limit and an upper limit. For example, the lower limit may be 0 and the upper limit may be 1. The closer the variability value is to the lower limit, the more stable the corresponding synapse may be, such that its strength (w) does not change easily. The closer the variability value is to the upper limit, the more flexible the corresponding synapse may be, i.e., unstable, such that its strength (w) changes easily.

[0046] Next, in step 920, the computing device 100 may sequentially learn about multiple tasks using an artificial neural network and store information about each of the tasks via synapses. Specifically, the processor 150 may adjust the strength (w) of each synapse according to the variability value assigned to each synapse while learning about each task, and store task information based on the strength (w) via the synapses.

[0047] As shown in Figure 2, during learning for each task, the processor 150 may adjust the synaptic strength (w) in the following way: That is, the processor 150 adjusts the strength (w) of each synapse to a new strength (w) n+1 ) may be adjusted. For each synapse, the processor 150 adjusts the intensity range (△w) from the initial intensity (w0) before learning to a range according to the variability value. n+1 n=0, 1, 2, ... For example, △w1 may be determined for learning Task 1, and △w2 for learning Task 2. The closer the variability value is to the lower limit, the greater the intensity adjustment range (△w n+1 ) can be small, and the closer the variability value is to the upper limit, the wider the intensity adjustment range (△w n+1 ) can be large. Also, processor 150 has a variable value, previous intensity (△w n ), and strength adjustment range (△w n+1) by combining new strengths (w n+1 ) may be determined. In some embodiments, the processor 150 determines the variability value and the intensity adjustment range (Δw n+1 Based on this, the reduced learning rate from the previous learning rate is detected, and the previous intensity (w n ) and a combination of reduced learning rates create a new strength (w n+1 ) can be determined. The closer the variability value is to the lower limit, the larger the reduction in the learning rate can be, and the closer the variability value is to the upper limit, the smaller the reduction in the learning rate can be. For example, a new intensity (w n+1 ) may be determined using the above formulas (1) and (2).

[0048] In various embodiments, by applying metaplasticity rules, artificial neural networks spontaneously exhibit learning phenomena that appear in brain task memory, such as the serial position effect and the repetition learning effect. That is, as shown in Figure 3, when learning multiple tasks, information for new tasks can be learned while storing information for past tasks. Also, as shown in Figure 3, the memory accuracy for each task improves the more that task is learned. In other words, metaplasticity rules enable artificial neural networks to adaptively utilize memory capacity in the same way as human fluid memory. Specifically, the processor 150 can store information via synapses according to the memory capacity of each task, while repeatedly advancing learning according to the learning frequency set for each task by the artificial neural network. In this case, if the same learning frequency is set for the tasks, the memory capacity is distributed equally among the tasks, and if different learning frequencies are set for the tasks, a portion of the memory capacity of the task with the lower learning frequency may be reallocated for the task with the higher learning frequency.

[0049] According to this disclosure, by using an artificial neural network to which metaplasticity rules are applied to learn multiple tasks, information at each stage can be stored with a certain level of memory accuracy, while simultaneously being able to store up to the maximum memory capacity, and the problem of catastrophic forgetting, in which information from past tasks is lost, can be prevented in this process. The artificial neural network in this disclosure can automatically realize a fluid memory function of information without any additional computational process. The performance of the artificial neural network in this disclosure is enhanced by iterative learning, and there is no loss even if noisy or contaminated data is later presented as the training dataset.

[0050] In short, this disclosure provides a computing device 100 and a method of operation thereof that solves catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain.

[0051] In this disclosure, the operation method of the computing device 100 may include the steps of assigning different variability values ​​to multiple synapses of an artificial neural network (step 910), and storing information about at least one task via the synapses while learning for at least one task using the artificial neural network (step 920).

[0052] In this disclosure, training on at least one task using an artificial neural network may mean training on a single task or training on multiple tasks sequentially.

[0053] In this disclosure, the step of storing information for at least one task (step 920) may include the step of adjusting the synaptic strength (w) according to the variability value assigned to each synapse while learning for each task, and the step of storing task information based on the strength (w) via the synapses.

[0054] In this disclosure, the step of adjusting the strength (w) of each synapse involves adjusting the strength range (△w) from the initial strength (w0) before learning to each synapse, according to the variability value. n+1 The steps include determining the variability value and the previous intensity (w n ), and strength adjustment range (△w n+1 ) by combining new strengths (w n+1 This may include a step to determine the following:

[0055] In this disclosure, the variability value is within the range between the lower limit and the upper limit, and the closer the variability value is to the lower limit, the greater the intensity adjustment range (△w n+1 The smaller the variable value, the closer the variability value is to the upper limit, and the greater the strength adjustment range (△w n+1 ) can be large.

[0056] In this disclosure, a new intensity (w n+1 The step of determining the variability value and the intensity adjustment range (△w) is n+1 A step to detect a reduced learning rate from the previous learning rate based on the previous intensity (w n ) and a combination of reduced learning rates create a new strength (w n+1 This may include a step to determine the following:

[0057] In this disclosure, the closer the variability value is to the lower limit, the greater the reduction in the learning rate; and the closer the variability value is to the upper limit, the smaller the reduction in the learning rate may be.

[0058] In this disclosure, the lower limit may be 0 and the upper limit may be 1.

[0059] In this disclosure, the new intensity may be determined using formulas (1) and (2).

[0060] In this disclosure, the step of storing information for at least one task may include the step of storing information via synapses according to the memory capacity of each task, while repeatedly learning by an artificial neural network according to the learning frequency set for each of the multiple tasks.

[0061] In this disclosure, if the same learning frequency is set for the tasks, the memory capacity is distributed equally among the tasks; if different learning frequencies are set for the tasks, a portion of the memory capacity of the task with the lower learning frequency may be reallocated for the task with the higher learning frequency.

[0062] In this disclosure, the computing device 100 may include a memory 140 and a processor 150 configured to be linked to the memory 140 and to execute at least one instruction stored in the memory 140, to assign different variability values ​​to multiple synapses of an artificial neural network, and to store information for at least one task via the synapses while the artificial neural network is learning for at least one task.

[0063] In this disclosure, training an artificial neural network on at least one task may mean training on a single task or training on multiple tasks sequentially.

[0064] In this disclosure, the processor 150 may be configured to adjust the synaptic strength (w) according to the variability value assigned to each synapse while learning for each task, and to store task information based on the strength (w) via the synapses.

[0065] In this disclosure, the processor 150 adjusts the intensity range (△w) from the initial intensity (w0) before learning to each synapse, according to its variability. n+1 ) is determined, and the variability value, previous intensity (w n ), and strength adjustment range (△w n+1) by combining new strengths (w n+1 ) may be configured to determine.

[0066] In this disclosure, the variability value is within the range between the lower limit and the upper limit, and the closer the variability value is to the lower limit, the greater the intensity adjustment range (△w n+1 The smaller the variable value, the closer the variability value is to the upper limit, and the greater the strength adjustment range (△w n+1 ) can be large.

[0067] In this disclosure, the processor 150 has a variability value and an intensity adjustment range (Δw n+1 Based on this, the reduced learning rate from the previous learning rate is detected, and the previous intensity (w n ) and a combination of reduced learning rates create a new strength (w n+1 ) may be configured to determine.

[0068] In this disclosure, the closer the variability value is to the lower limit, the greater the reduction in the learning rate; and the closer the variability value is to the upper limit, the smaller the reduction in the learning rate may be.

[0069] In this disclosure, the lower limit may be 0 and the upper limit may be 1.

[0070] In this disclosure, the new intensity may be determined using formulas (1) and (2).

[0071] In this disclosure, the processor 150 may be configured to store information via synapses according to the memory capacity of each task, while repeatedly performing learning according to a learning frequency set for each of the multiple tasks using an artificial neural network.

[0072] In this disclosure, if the same learning frequency is set for the tasks, the memory capacity is distributed equally among the tasks; if different learning frequencies are set for the tasks, a portion of the memory capacity of the task with the lower learning frequency may be reallocated for the task with the higher learning frequency.

[0073] The apparatus described above may be implemented by hardware components, software components, and / or combinations of hardware and software components. For example, the apparatus and components described in the embodiments may be implemented using one or more general-purpose or special-purpose computers, such as processors, controllers, ALUs (arithmetic logic units), digital signal processors, microcomputers, FPGAs (field programmable gate arrays), PLUs (programmable logic units), microprocessors, or various devices capable of executing and responding to instructions. The processing unit may execute an operating system (OS) and one or more software applications running on the OS. The processing unit may also respond to software execution, access data, record, manipulate, process, and generate data. For convenience of understanding, it may be described as if a single processing unit is used, but those skilled in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements. For example, the processing unit may include multiple processors or one processor and one controller. Other processing configurations, such as parallel processors, are also possible.

[0074] Software may include computer programs, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or which may instruct the processing unit independently or collectively. Software and / or data may be embodied in any kind of machine, component, physical device, computer recording medium, or device for interpretation based on the processing unit or for providing instructions or data to the processing unit. Software may be distributed across a networked computer system, and may be recorded or executed in a distributed manner. Software and data may be recorded on one or more computer-readable recording media.

[0075] The methods according to the embodiment may be implemented in the form of program instructions executable by various computer means and recorded on a computer-readable medium. In this case, the medium may continuously record computer-executable programs or may temporarily record them for execution or download. Furthermore, the medium may be various recording or storage means in the form of a combination of one or more hardware components, and may be a medium directly connected to a computer system or distributed on a network. Examples of media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and media configured to record program instructions such as ROM, RAM, and flash memory. Other examples of media include recording media and storage media managed by app stores that distribute applications, and sites and servers that supply and distribute various other software.

[0076] The various embodiments and the terminology used herein should not be understood as limiting the technology described herein to any particular embodiment, but rather as including various modifications, equivalents, and / or substitutes of the embodiments. In the description of the drawings, similar components are referred to by similar reference numerals. Singular expressions may include plural expressions unless the context clearly indicates otherwise. In this specification, expressions such as “A or B,” “A and / or B,” “A, B, or C,” or “A, B, and / or C” may include all possible combinations of items listed in parallel. Expressions such as “first,” “second,” “first,” or “second” list the components in question, regardless of order or importance, and are used only to distinguish one component from another, not to limit that component. When a component (e.g., the first) is described as being "connected (functionally or communicatively)" or "linked" to another component (e.g., the second), the first component may be directly connected to the other component or connected via another component (e.g., the third component).

[0077] According to various embodiments, each component of the components described (e.g., a module or a program) may include one or more individuals. According to various embodiments, one or more components or steps of the components described above may be omitted, or one or more other components or steps may be added. Alternatively or additionally, multiple components (e.g., a module or a program) may be integrated into a single component. In such a case, the integrated component may perform one or more functions of each of the multiple components in the same or similar manner as they were performed by the corresponding component of the multiple components before integration. According to various embodiments, steps performed by a module, program or other component may be performed sequentially, in parallel, iteratively, or heuristically, one or more of the steps may be performed in a different order, omitted, or one or more other steps may be added. [Explanation of Symbols]

[0078] 100: Computing device 110: Communication module 120: Input Module 130: Output Module 140: Memory 150: Processor

Claims

1. A method for operating a computing device that solves catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain, A step of assigning different variability values ​​to multiple synapses in an artificial neural network, and The step of storing information about at least one task via the synapses while the artificial neural network is learning about at least one task. Includes, The step of storing the information of at least one task is: The steps include: adjusting the strength of each synapse according to the variability value assigned to each synapse while progressing with learning for each task, and The step of storing information about the task based on the intensity via the synapse. including, How computing devices operate.

2. The step of adjusting the strength of each of the synapses is: For each of the aforementioned synapses, the steps include determining the range of intensity adjustment from the initial intensity before the learning process, according to the variability value, and The step of determining a new intensity by a combination of the aforementioned variability value, the previous intensity, and the intensity adjustment range. including, A method for operating the computing device according to claim 1.

3. The aforementioned variability value is within the range between the lower limit and the upper limit. The closer the variability value is to the lower limit, the smaller the intensity adjustment range. The closer the variability value is to the upper limit, the larger the intensity adjustment range. A method for operating the computing device according to claim 2.

4. The step of determining the new strength is, A step of detecting a learning rate reduced from the previous learning rate based on the variability value and the intensity adjustment range, and The step of determining the new intensity by a combination of the previous intensity and the reduced learning rate. including, A method for operating the computing device according to claim 3.

5. The closer the variability value is to the lower limit, the greater the reduction in the learning rate. The closer the variability value is to the upper limit, the smaller the range by which the learning rate is reduced. A method for operating the computing device according to claim 4.

6. The lower limit is 0, and the upper limit is 1. A method for operating the computing device according to claim 3.

7. The aforementioned new intensity is determined as shown in the following formula (1): Here, w represents the synaptic strength, △w represents the strength adjustment range, n represents the previous learning iteration, and n+1 represents the new learning iteration. represents the learning rate, J(•) represents the loss function, and S(•) represents the learning rate reduction function, which are determined as shown in equation (2) below. Here, flexibility represents the variability value, and α is a hyperparameter that adjusts the width of S(•). A method for operating the computing device according to claim 2.

8. The step of storing the information of at least one task is: The artificial neural network repeatedly performs learning according to the learning frequency set for each of the multiple tasks, and stores the information via the synapses according to the memory capacity of each of the tasks. including, A method for operating the computing device according to claim 1.

9. If the same learning frequency is set for the aforementioned tasks, the memory capacity is distributed equally among the aforementioned tasks. If different learning frequencies are set for the aforementioned tasks, a portion of the memory capacity of the task with the lower learning frequency is reallocated for the task with the higher learning frequency. A method for operating the computing device according to claim 8.

10. A computing device that solves catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain, memory, and A processor configured to be linked to the memory, execute at least one instruction stored in the memory, assign different variability values ​​to multiple synapses of the artificial neural network, and store information about the at least one task via the synapses while the artificial neural network learns about the at least one task. Includes, The aforementioned processor, As learning progresses for each task, the strength of each synapse is adjusted according to the variability value assigned to each synapse. The system is configured to store information about the task based on the intensity via the synapse, Computing device.

11. The aforementioned processor, For each of the aforementioned synapses, the range of intensity adjustment from the initial intensity before the learning process is determined according to the variability value. The system is configured to determine a new intensity based on a combination of the aforementioned variable value, the aforementioned previous intensity, and the aforementioned intensity adjustment range. The computing device according to claim 10.

12. The aforementioned variability value is within the range between the lower limit and the upper limit. The closer the variability value is to the lower limit, the smaller the intensity adjustment range. The closer the variability value is to the upper limit, the larger the intensity adjustment range. The computing device according to claim 11.

13. The aforementioned processor, Based on the aforementioned variability value and the aforementioned intensity adjustment range, a learning rate reduced from the previous learning rate is detected. The new intensity is determined by a combination of the previous intensity and the reduced learning rate. The computing device according to claim 12.

14. The closer the variability value is to the lower limit, the greater the reduction in the learning rate. The closer the variability value is to the upper limit, the smaller the range by which the learning rate is reduced. The computing device according to claim 13.

15. The lower limit is 0, and the upper limit is 1. The computing device according to claim 12.

16. The aforementioned new intensity is determined by the following formula (1): Here, w represents the synaptic strength, △w represents the strength adjustment range, n represents the previous learning iteration, and n+1 represents the new learning iteration. represents the learning rate, J(•) represents the loss function, and S(•) represents the learning rate reduction function, which are determined as shown in equation (2) below. Here, flexibility represents the variability value, and α is a hyperparameter that adjusts the width of S(•). Computing device according to claim 11

17. The aforementioned processor, The artificial neural network is configured to repeatedly advance learning according to the learning frequency set for each of the multiple tasks, and to store the information via the synapses according to the memory capacity of each of the tasks. The computing device according to claim 10.

18. If the same learning frequency is set for the aforementioned tasks, the memory capacity is distributed equally among the aforementioned tasks. If different learning frequencies are set for the aforementioned tasks, a portion of the memory capacity of the task with the lower learning frequency is reallocated for the task with the higher learning frequency. The computing device according to claim 17.

19. A computer program recorded on a non-temporary computer-readable recording medium to cause a computing device to execute a method for resolving catastrophic forgetting in artificial neural network learning by applying the metaplasticity rules of the biological brain, The aforementioned method, A step of assigning different variability values ​​to multiple synapses in an artificial neural network, and The step of storing information about at least one task via the synapses while the artificial neural network is learning about at least one task. Includes, The step of storing the information for each of the aforementioned tasks is: The steps include: adjusting the strength of each synapse according to the variability value assigned to each synapse while progressing with learning for each task, and The step of storing information about the task based on the intensity via the synapse. including, Computer program.

20. The step of adjusting the strength of each of the synapses is: For each of the aforementioned synapses, the steps include determining the range of intensity adjustment from the initial intensity before the learning process proceeds, according to the variability value, and The step of determining a new intensity by a combination of the aforementioned variability value, the previous intensity, and the intensity adjustment range. including, The computer program according to claim 19.

Images