A Continuous Learning Method, Device, Medium and Equipment Based on Memory Consolidation

Abstract

This specification discloses a continuous learning method, device, medium, and equipment based on memory consolidation. In this method, the sensory cortex module can sequentially learn the training sample sets of different tasks in each round of training, and transfer the parameters of the learned sensory cortex module to the hippocampal module. Subsequently, by loading the weights of the hippocampal module into the prefrontal cortex module, the prefrontal cortex module superimposes the weights of the hippocampal module, thereby realizing the memory replay of the hippocampus. Finally, by inputting slow-wave rhythm signals into the prefrontal cortex module, it can be adjusted in combination with the weights of the hippocampal module, thereby effectively alleviating the "catastrophic forgetting" problem of the neural network model in continuous learning and significantly improving the accuracy.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence technology, and in particular, to a continuous learning method, device, medium, and equipment based on memory consolidation. Background Art

[0002] With the development of machine learning technology, neural network models have achieved remarkable achievements in dealing with complex tasks. Among them, in order to improve the processing ability of neural network models for complex tasks, it is usually possible to make the neural network model learn new tasks or new data in a changing environment, so that the neural network model can continuously improve its own performance in a dynamic and unpredictable environment.

[0003] However, when continuously learning multiple tasks through a neural network model, there is usually a problem of "catastrophic forgetting", that is, when the neural network model learns a new task each time, its processing performance for the old task will drop sharply. This problem of "catastrophic forgetting" will result in poor accuracy of the neural network model in performing old task processing, thereby restricting the further development of artificial intelligence technology. Summary of the Invention

[0004] This specification provides a continuous learning method, device, medium, and equipment based on memory consolidation to partially solve the above problems existing in the prior art.

[0005] This specification adopts the following technical solutions:

[0006] This specification provides a continuous learning method based on memory consolidation. The method is applied to a multi-task continuous learning system, and the multi-task continuous learning system includes: a prefrontal cortex module, a hippocampal module, and a sensory cortex module. Each of the prefrontal cortex module, the hippocampal module, and the sensory cortex module contains various network layers. For each network layer, each neuron node is included in the network layer. The method includes:

[0007] Obtain each training sample set for different tasks, and each training sample set includes at least one of image data, text data, and audio data;

[0008] According to each training sample set, perform multiple rounds of parameter adjustment on the module parameters of the prefrontal cortex module to obtain a target prefrontal cortex module, and perform task execution through the target prefrontal cortex module; wherein,

[0009] For each round of parameter adjustment, select the target training sample set used in this round of parameter adjustment from each training sample set, and use the prefrontal cortex module after the previous round of parameter adjustment as the prefrontal cortex module to be adjusted used in this round of parameter adjustment;

[0010] Train the somatosensory cortex module according to the target training sample set to obtain a trained somatosensory cortex module. Transfer the module parameters of the trained somatosensory cortex module to the hippocampal module, and extract the connection weights between the neuron nodes included in the hippocampal module as the target weights. Load the target weights into the prefrontal cortex module to be adjusted, and input a preset slow-wave rhythm signal into the prefrontal cortex module to be adjusted. According to the spike firing sequence between the neuron nodes when the prefrontal cortex module to be adjusted receives the slow-wave rhythm signal, adjust the connection weights between the neuron nodes in the prefrontal cortex module to be adjusted to obtain the prefrontal cortex module after this round of adjustment. When it is determined that a preset termination condition is met, use the prefrontal cortex module after this round of adjustment as the target prefrontal cortex module.

[0011] Optionally, each network layer included in the hippocampal module corresponds one-to-one with each network layer included in the somatosensory cortex module. For each network layer included in the hippocampal module, each neuron node included in this network layer corresponds one-to-one with each neuron node included in the corresponding network layer of the somatosensory cortex module.

[0012] Optionally, for each neuron node included in the prefrontal cortex module to be adjusted, the neuron node updates the membrane potential corresponding to the neuron node according to the received electrical signal, the connection weight between the neuron node and the neuron node in the previous network layer, and the slow-wave phase of the slow-wave cycle in which the neuron node is currently located. And if the membrane potential corresponding to the neuron node exceeds a preset spike threshold, a spike is sent to the neuron node in the next network layer that has a connection relationship with the neuron node. The slow-wave phase of the slow-wave cycle in which the neuron node is currently located includes: the rising phase and the falling phase. Among them, if the slow-wave phase of the slow-wave cycle in which the neuron node is currently located is the rising phase, the update amplitude of the membrane potential corresponding to the neuron node each time is greater than the update amplitude of the membrane potential corresponding to the neuron node when it is currently in the falling phase. The electrical signal includes at least one of a slow-wave rhythm signal and a spike output by a neuron node in the previous network layer that has a connection relationship with the neuron node.

[0013] Optionally, for each neuron node included in the prefrontal cortex module to be adjusted, the neuron node updates the membrane potential corresponding to the neuron node according to the received electrical signal, the connection weight between the neuron node and the neuron node in the previous network layer, and the slow-wave phase of the slow-wave cycle in which the neuron node is currently located, specifically including:

[0014] For each neuron node included in the prefrontal cortex module to be adjusted, the neuron node updates the membrane potential corresponding to the neuron node according to the received electrical signal, the connection weight between the neuron node and the neuron node in the previous network layer, and the scaling factor corresponding to the slow wave phase of the slow wave cycle in which the neuron node is currently located.

[0015] Optionally, the prefrontal cortex module and the hippocampal module are composed of a spiking neural network model, and the sensory cortex module is composed of a spiking neural network model or an artificial neural network model.

[0016] This specification provides a task execution method, including:

[0017] Obtain task data of the task to be executed, where the task data is at least one of image data, text data, and audio data;

[0018] Input the task data into a pre-trained target model to process the task data through the target model to execute the task to be executed and obtain an execution result, where the target model is obtained based on the target prefrontal cortex module trained by the above continuous learning method based on memory consolidation.

[0019] This specification provides a continuous learning device based on memory consolidation, including:

[0020] An acquisition module, configured to acquire each training sample set for different tasks, where each training sample set includes at least one of image data, text data, and audio data;

[0021] A training module, configured to perform multiple rounds of parameter adjustments on the module parameters of the prefrontal cortex module according to the training sample sets, so as to obtain a target prefrontal cortex module, and perform task execution through the target prefrontal cortex module; wherein, for each round of parameter adjustment, a target training sample set used in this round of parameter adjustment is selected from the training sample sets, and the prefrontal cortex module after the previous round of parameter adjustment is used as the prefrontal cortex module to be adjusted in this round of parameter adjustment; the sensory cortex module is trained according to the target training sample set to obtain a trained sensory cortex module, the module parameters of the trained sensory cortex module are transferred to the hippocampal module, and the connection weights between the neuron nodes included in the hippocampal module are extracted as target weights, the target weights are loaded into the prefrontal cortex module to be adjusted, and a preset slow-wave rhythm signal is input into the prefrontal cortex module to be adjusted, so as to adjust the connection weights between the neuron nodes in the prefrontal cortex module to be adjusted according to the spike firing order between the neuron nodes when the prefrontal cortex module to be adjusted receives the slow-wave rhythm signal, and obtain the prefrontal cortex module after this round of adjustment; when it is determined that a preset termination condition is met, the prefrontal cortex module after this round of adjustment is used as the target prefrontal cortex module.

[0022] This specification provides a task execution device, including:

[0023] A task data acquisition module, configured to acquire task data of a task to be executed, where the task data is at least one of image data, text data, and audio data;

[0024] A task execution module, configured to input the task data into a pre-trained target model, so as to process the task data through the target model to execute the task to be executed and obtain an execution result, where the target model is obtained based on the target prefrontal cortex module trained by the above-mentioned continuous learning method based on memory consolidation.

[0025] This specification provides a computer-readable storage medium, where the storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned continuous learning method based on memory consolidation and task execution method are implemented.

[0026] This specification provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the program, the above-mentioned continuous learning method based on memory consolidation and task execution method are implemented.

[0027] At least one of the above technical solutions adopted in this specification can achieve the following beneficial effects:

[0028] In the continuous learning method based on memory consolidation provided in this specification, each training sample set for different tasks is obtained. Each of the training sample sets here includes at least one of image data, text data, and audio data. According to each training sample set, the module parameters of the prefrontal cortex module are adjusted in multiple rounds to obtain a target prefrontal cortex module, and task execution is performed through the target prefrontal cortex module. Among them, for each round of parameter adjustment, a target training sample set used in this round of parameter adjustment is selected from each training sample set, and the prefrontal cortex module after the previous round of parameter adjustment is used as the prefrontal cortex module to be adjusted in this round of parameter adjustment. The sensory cortex module is trained according to the target training sample set to obtain a trained sensory cortex module. The module parameters of the trained sensory cortex module are transferred to the hippocampal module, and the connection weights between each neuron node included in the hippocampal module are extracted as target weights. The target weights are loaded into the prefrontal cortex module to be adjusted, and a preset slow-wave rhythm signal is input into the prefrontal cortex module to be adjusted. According to the pulse firing sequence between each neuron node when the prefrontal cortex module to be adjusted receives the slow-wave rhythm signal, the connection weights between each neuron node in the prefrontal cortex module to be adjusted are adjusted to obtain the prefrontal cortex module after this round of adjustment. When it is determined that a preset termination condition is met, the prefrontal cortex module after this round of adjustment is used as the target prefrontal cortex module.

[0029] As can be seen from the above method, when the server needs to use the prefrontal cortex module to learn new task content each time, the sensory cortex module can be trained through the training sample set corresponding to the new task, and the module parameters after the sensory cortex module learns are transferred to the hippocampal module. Furthermore, the weights of the hippocampal module can be extracted and loaded into the prefrontal cortex module, so that the prefrontal cortex module superimposes the weights of the hippocampal module. Furthermore, by inputting a slow-wave rhythm signal into the prefrontal cortex module, each neuron node included in the prefrontal cortex module can emit pulses under the action of the slow-wave rhythm signal. Furthermore, according to the sequence of pulses emitted by each neuron node included in the prefrontal cortex module under the action of the slow-wave rhythm signal, the connection patterns between each neuron node and between each neuron node used to represent new task content and old task content are adjusted, so that the prefrontal cortex module can retain the content of the newly learned task and the content of the old task, so as to alleviate the "catastrophic forgetting" problem in the continuous learning process of the neural network model, and further improve the accuracy of the neural network model in performing tasks. Description of the Drawings

[0030] The drawings described herein are used to provide a further understanding of this specification and form a part of this specification. The schematic embodiments of this specification and their descriptions are used to explain this specification and do not constitute an improper limitation of this specification. In the drawings:

[0031] Figure 1 It is a schematic flowchart of a continuous learning method based on memory consolidation provided in this specification;

[0032] Figure 2 It is a schematic diagram of the first-round parameter adjustment process provided in this specification;

[0033] Figure 3 It is a schematic diagram of the second-round parameter adjustment process provided in this specification;

[0034] Figure 4 It is a schematic flowchart of a task execution method provided in this specification;

[0035] Figure 5 It is a schematic diagram of a continuous learning device based on memory consolidation provided in this specification;

[0036] Figure 6 It is a schematic diagram of a task execution device provided in this specification;

[0037] Figure 7 It is provided in this specification corresponding to Figure 1 schematic diagram of an electronic device. Specific embodiments

[0038] To make the purpose, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope protected by this specification.

[0039] The following will detail the technical solutions provided in each embodiment of this specification in conjunction with the drawings.

[0040] Figure 1 It is a schematic flowchart of a continuous learning method based on memory consolidation provided in this specification, including the following steps:

[0041] S101: Obtain each training sample set for different tasks, and each of the training sample sets includes at least one of the following data: image data, text data, and audio data.

[0042] In a biological brain, memory is one of the bases of intelligence, and the hippocampal region in the biological brain is a key area for the brain to form long-term memories. During sleep, the biological brain converts the content learned in the short term into long-term memories through a memory consolidation mechanism. During slow-wave sleep, the coordinated activities between the cerebral cortex and the hippocampus support this process.

[0043] Specifically, when a biological entity is in a learning state, the hippocampus in its biological brain first converts the received information into short-term memories. These short-term memories are initially processed and integrated in the hippocampus to gradually stabilize them, thus preventing them from quickly fading due to interference from other information. Further, when the biological entity is in a sleeping state, the hippocampus in its biological brain can transmit the stored short-term memories to the cerebral cortex in the form of memory replay, so that the cerebral cortex can, based on the memories replayed by the hippocampus, generate slow waves to regulate the whole, enabling the cerebral cortex to adjust the connection strength and structure between neural synapses in two forms, namely long-term potentiation (LTP) and long-term depression (LTD), to achieve memory consolidation.

[0044] Based on this, this specification provides a continuous learning method based on memory consolidation. By transferring the module parameters of the perceptual cortex modules trained with different training sample sets to the hippocampal module, and by simulating the activities of the hippocampus and the cerebral cortex of the brain during biological sleep through the hippocampal module and the prefrontal cortex module, the connection pattern between each neuron node in the prefrontal cortex module can be adjusted. Thus, during the slow-wave sleep simulated by the hippocampal module and the prefrontal cortex module, by combining the collaborative features of slow waves and their influence on network synaptic plasticity, a memory consolidation function with a bionic effect can be achieved, and further, the "catastrophic forgetting" problem during the continuous learning process of the neural network model can be alleviated to improve the accuracy of task execution by the neural network model.

[0045] In this specification, the execution entity for implementing the continuous learning method based on memory consolidation can be a specified device such as a server set up on a service platform. Of course, it can also be a terminal device such as a laptop or a desktop computer. For the sake of convenience in description, this specification only takes the server as an example of the execution entity to illustrate a continuous learning method based on memory consolidation provided by this specification.

[0046] Among them, the above-mentioned continuous learning method based on consolidation can be applied to a multi-task continuous learning system, and here the multi-task continuous learning system includes: a prefrontal cortex module, a hippocampal module, and a sensory cortex module.

[0047] The above-mentioned prefrontal cortex module and hippocampal module contain various network layers. For each network layer, each network layer contains various neuron nodes. Here, the prefrontal cortex module and the hippocampal module can be composed of a spiking neural network model (SNN).

[0048] The above-mentioned sensory cortex module contains various network layers. For each network layer, each network layer contains various neuron nodes. Here, the sensory cortex module can be composed of an artificial neural network model (ANNs) or a spiking neural network model.

[0049] Furthermore, the server can obtain each training sample set required during the process of continuous learning for the prefrontal cortex module.

[0050] In the above content, each training sample set corresponds to a different task, and the data contained in each training sample set can be at least one of: image data, text data, and audio data. The above tasks can be determined according to the actual situation.

[0051] For example: The above different tasks can be recognition tasks for different image data, such as: recognition tasks for image data containing the number "1" and image data containing the number "2", recognition tasks for image data containing the number "7" and image data containing the number "8", etc.

[0052] Of course, the above different tasks can also refer to image classification tasks for datasets such as CIFAR and ImageNet.

[0053] S102: Based on the respective training sample sets, perform multiple rounds of parameter adjustment on the module parameters of the prefrontal cortex module to obtain a target prefrontal cortex module, and perform task execution through the target prefrontal cortex module. Among them, for each round of parameter adjustment, select the target training sample set used in this round of parameter adjustment from the respective training sample sets, and use the prefrontal cortex module after the previous round of parameter adjustment as the prefrontal cortex module to be adjusted used in this round of parameter adjustment; train the sensory cortex module according to the target training sample set to obtain a trained sensory cortex module, transfer the module parameters of the trained sensory cortex module to the hippocampal module, and extract the connection weights between the respective neuron nodes included in the hippocampal module as the target weights, load the target weights into the prefrontal cortex module to be adjusted, and input a preset slow-wave rhythm signal into the prefrontal cortex module to be adjusted, so as to adjust the connection weights between the respective neuron nodes in the prefrontal cortex module to be adjusted according to the impulse firing sequence between the respective neuron nodes when the prefrontal cortex module to be adjusted receives the slow-wave rhythm signal, and obtain the prefrontal cortex module after this round of adjustment; when it is determined that a preset termination condition is met, then use the prefrontal cortex module after this round of adjustment as the target prefrontal cortex module.

[0054] In this specification, the server can perform multiple rounds of parameter adjustment on the prefrontal cortex module. The training sample sets used in each round of parameter adjustment are different. In this way, the server can obtain a prefrontal cortex module after continuous learning as the target prefrontal cortex module. Through the target prefrontal cortex module, not only can it process new tasks corresponding to new training sample sets in multiple rounds of parameter adjustment, but also can simultaneously process old tasks corresponding to old training sample sets.

[0055] Specifically, for each round of parameter adjustment in the above-mentioned multiple rounds of parameter adjustment, the server can determine the training sample set used in this round of parameter adjustment from the respective training sample sets of different tasks as the target training sample set, and use the prefrontal cortex module after the previous round of parameter adjustment as the prefrontal cortex module to be adjusted used in this round of parameter adjustment.

[0056] Furthermore, the server can input the target training sample set into the sensory cortex module used in this round of parameter adjustment to train the sensory cortex module used in this round of parameter adjustment with the target training sample set to obtain a trained sensory cortex module. Furthermore, the module parameters of the trained sensory cortex module can be transferred to the hippocampal module used in this round of parameter adjustment.

[0057] Among them, the target training sample sets used for each round of parameter adjustment are different, and the somatosensory cortex modules and hippocampal modules used in each round of parameter adjustment are also different. It can be understood that in each round, a new training sample set is used to train a new somatosensory cortex module, and the module parameters of the trained somatosensory cortex module are transferred to a new hippocampal module.

[0058] In practical application scenarios, the above-mentioned somatosensory cortex modules can be of two types, and the following will elaborate on these two types of somatosensory cortex modules in detail.

[0059] The first type of somatosensory cortex module can be composed of an artificial neural network model. At this time, the method for the server to transfer the module parameters of the trained somatosensory cortex module to the hippocampal module can be: according to the module parameters of the trained somatosensory cortex module, determine the number of each network layer included in the somatosensory cortex module, the number of each neuron node included in each network layer of the somatosensory cortex module, the connection relationship between each neuron node in the somatosensory cortex module, and the connection weights between each neuron node in the somatosensory cortex module.

[0060] Furthermore, the server can adjust the module parameters of the hippocampal module according to the number of each network layer included in the somatosensory cortex module, the number of each neuron node included in each network layer of the somatosensory cortex module, the connection relationship between each neuron node in the somatosensory cortex module, and the connection weights between each neuron node in the somatosensory cortex module, so as to transfer the module parameters of the somatosensory cortex module trained in this round to the hippocampal module.

[0061] Specifically, the server can use neuron nodes of the Leaky Integrate and Fire (LIF) type to form each neuron node in the spiking neural network model that constitutes the hippocampal module. Each neuron node in the hippocampal module corresponds one-to-one with each computing node in the somatosensory cortex module. The relationship between the membrane potential of each neuron node and time can be expressed as:

[0062]

[0063] Among them, is the initial membrane potential, is the membrane potential corresponding to time corresponding, is the current input corresponding to time at time, is the preset time parameter, is the preset resistance parameter. In this specification, , 。

[0064] It can be understood that each network layer included in the hippocampal module in the above content corresponds one-to-one with each network layer included in the sensory cortex module. For each network layer included in the hippocampal module, each neuron node included in this network layer corresponds one-to-one with each neuron node in the corresponding network layer in the sensory cortex module. For any two neuron nodes included in the hippocampal module, the connection relationship between these two neuron nodes is the same as the connection relationship between the two corresponding neuron nodes in the sensory cortex module.

[0065] In other words, for each connection existing in the artificial neural network model constituting the sensory cortex module, there is also a corresponding connection in the spiking neural network model constituting the hippocampal module. The server can assign initial membrane potential values and spike thresholds to all neuron nodes in the spiking neural network model constituting the hippocampal module. Preferably, the server can set the initial membrane potential and the spike threshold to = -75 mV, = -55 mV.

[0066] The second type of sensory cortex module can be composed of a spiking neural network model. At this time, the server can directly use the spiking neural network model constituting the trained sensory cortex module as the spiking neural network model constituting the hippocampal module, so as to transfer the module parameters of the sensory cortex module after this round of training to the hippocampal module.

[0067] It should be noted that the server can determine the spike time series of each neuron node in the input layer included in the spiking neural network model constituting the hippocampal module according to each training sample data in the target training sample set used in this round of training through the Poisson distribution.

[0068] The spike value corresponding to each time unit in the above spike time series can be represented as 1 or 0, that is, 1 represents emitting a spike or 0 represents not emitting a spike. Thus, the above spike time series can be input into the input layer of the hippocampal module, so that the input layer of the hippocampal module releases spike signals to the neuron nodes connected to the input layer included in other network layers of the hippocampal module, for initializing the hippocampal module.

[0069] Furthermore, the server can extract the connection weights between each neuron node included in the spiking neural network model constituting the hippocampal module as the target weights. Then, the target weights can be loaded into the spiking neural network model constituting the prefrontal cortex module to load the connection weights of the hippocampal module for the prefrontal cortex module, and the prefrontal cortex module can be initialized using the above spike sequence.

[0070] It should be noted that in the first-round parameter adjustment, the prefrontal cortex module to be adjusted is also obtained by transferring the module parameters of the sensory cortex module. And in the subsequent rounds of parameter adjustment, the prefrontal cortex module to be adjusted used is the prefrontal cortex module after the previous round of parameter adjustment.

[0071] Furthermore, the server can input a preset slow-wave rhythm signal into the prefrontal cortex module to be adjusted, so that each neuron node in the prefrontal cortex module to be adjusted emits pulses under the action of the slow-wave rhythm signal.

[0072] Among them, for each neuron node included in the prefrontal cortex module to be adjusted, the neuron node updates the membrane potential corresponding to the neuron node according to the received electrical signal, the connection weight between the neuron node and the neuron node in the previous network layer, and the slow-wave phase of the slow-wave cycle in which the neuron node is currently located. And after updating the membrane potential, if the membrane potential corresponding to the neuron node exceeds the preset pulse threshold, a pulse is emitted to the neuron node in the next network layer that has a connection relationship with the neuron node.

[0073] Among them, the above-mentioned electrical signal includes at least one of a slow-wave rhythm signal and a pulse output by a neuron node in the previous network layer that has a connection relationship with the neuron node.

[0074] In the above content, the slow-wave phases of the slow-wave cycle in which the prefrontal cortex module is currently located include: the rising phase and the falling phase. Among them, the rising phase and the falling phase alternate. For example, the prefrontal cortex module to be adjusted can be in the rising phase in the first unit time, and can be in the falling phase in the second unit time, and so on.

[0075] If the slow-wave phase of the slow-wave cycle in which the neuron node is currently located is the rising phase, the update amplitude of the membrane potential corresponding to the neuron node each time is greater than the update amplitude of the membrane potential corresponding to the neuron node when it is currently in the falling phase.

[0076] It should be noted that for each neuron node included in the prefrontal cortex module to be adjusted, the neuron node can update the membrane potential corresponding to the neuron node according to the received electrical signal, the connection weight between the neuron node and the neuron node in the previous network layer, and the scaling factor corresponding to the slow-wave phase of the slow-wave cycle in which the neuron node is currently located. Specifically, it can refer to the following formula:

[0077]

[0078] Among them, is the current input of the current received by the neuron node integrating the slow-wave rhythm signal at time t, is the pulse time series (0 / 1) corresponding to the slow-wave rhythm signal, is the connection weight between neuron nodes at time t, is a preset scaling factor used to scale it so that the neuron node can receive all the information contained in the slow-wave rhythm signal.

[0079] Specifically, for each neuron node in the prefrontal cortex module to be adjusted, the neuron node can update the membrane potential corresponding to the neuron node according to the integrated current input of the received slow-wave rhythm signal. If the membrane potential corresponding to the neuron node exceeds the preset pulse threshold (-55 mV), a pulse is sent to the next-layer neuron nodes that have a connection relationship with the neuron node, and it is reset to the initial membrane potential (-75 mV). Otherwise, no pulse is sent, and it waits to receive the next input of the slow-wave rhythm signal, and the membrane potential corresponding to the neuron node is updated again according to the next received slow-wave rhythm signal. Specifically, it can refer to the following formula:

[0080]

[0081] where, is the integrated current input received by each neuron node from the slow-wave rhythm signal at time t, is the pulse time series corresponding to the slow-wave rhythm signal, is the connection weight between neuron nodes at time t, is the scaling factor, [sign(z)] + is the rectification function, when is greater than 0, the return value is 1, and in other cases the return value is 0. Thus, the slow-wave phase of the slow-wave cycle in which the prefrontal cortex module to be adjusted is located at different unit times t is determined. 1 / T represents the slow-wave frequency corresponding to slow-wave sleep, and this slow-wave frequency determines the time length of a single slow-wave cycle, which can be set to 1 Hz in this specification. q represents the noise term, which follows a uniform distribution from -b to b, and b is the noise parameter.

[0082] As can be seen from the above, during the rising phase, the server can gradually increase the membrane potential of each neuron node in the prefrontal cortex module to be adjusted until it reaches a threshold, at which point the neuron node starts to fire impulses. During this process, the excitability of the neuron node increases, enabling more neuron nodes to be activated, facilitating the integration of information by the prefrontal cortex module to be adjusted, and helping to strengthen the connections between the neuron nodes related to the current task in the pulse neural network. During the falling phase, the membrane potential of each neuron node decreases, entering a relatively inhibitory state. During this process, the excitability of the neuron node decreases, and the impulse firing decreases or stops, so as to weaken the connection weights corresponding to the connections between the neuron nodes that are not frequently used in the pulse neural network, thereby reducing noise interference.

[0083] As can be seen from the above, in this specification, the server can temporarily store the module parameters learned by the sensory cortex module in each round of training in each hippocampal module, that is, for each round of training, the server can transfer the module parameters of the sensory cortex module after this round of training to the hippocampal module, so that in each round of parameter adjustment, the hippocampal module can store the information learned by the sensory cortex module in training in the short term. Furthermore, the connection weights of the hippocampal module can be loaded by the prefrontal cortex module to be adjusted and adjusted under the action of the slow-wave rhythm signal. Then, the prefrontal cortex module to be adjusted can consolidate according to the information stored in the hippocampal module to reduce the influence of the new information learned in the subsequent training process on the information stored in the hippocampal module learned by the adjusted prefrontal cortex module.

[0084] It should be noted that the method for the server to load the target connection weights of the hippocampal module into the prefrontal cortex module to be adjusted can be as follows: for each target connection weight of the hippocampal module, the corresponding connection weight is determined from the prefrontal cortex module to be adjusted. Then, the target connection weight and the corresponding connection weight can be fused to obtain the fused connection weight, and the fused connection weight is updated to the prefrontal cortex module to be adjusted, so as to obtain the prefrontal cortex module to be adjusted after loading the target connection weights of the hippocampal module.

[0085] Among them, the target connection weight and the corresponding connection weight are the connection weights between the same two neuron nodes. In other words, for the connection weights between each two neuron nodes included in the prefrontal cortex module, if it is determined that these two neuron nodes are the same as the two neuron nodes corresponding to the target connection weight, then it can be determined that the connection weight between these two neuron nodes is the corresponding connection weight to the target connection weight.

[0086] Further, the server can adjust the connection weights between the neuron nodes in the prefrontal cortex module to be adjusted according to the spike firing order between the connected neuron nodes in the prefrontal cortex module to be adjusted, so as to obtain the prefrontal cortex module after this round of adjustment. In the parameter adjustment of subsequent rounds, the prefrontal cortex module after this round of adjustment can be used as the prefrontal cortex module used in the next round of parameter adjustment.

[0087] In addition, the server can also convert the pulse neural network model that constitutes the prefrontal cortex module after this round of adjustment into an artificial neural network model to conduct tests during the period for the converted artificial neural network model.

[0088] Among them, the method of the above test can be that for each round of parameter adjustment, the server can also obtain the test set corresponding to the target training set used in this round of parameter adjustment, and test the artificial neural network model converted from the prefrontal cortex module after this round of adjustment through the test set to obtain the test result of the prefrontal cortex module after this round of adjustment.

[0089] If it is determined that the prefrontal cortex module after this round of adjustment does not meet the preset conditions according to the test result of the prefrontal cortex module after this round of adjustment, the parameter adjustment of the prefrontal cortex module can be re-performed through the above method. During this process, when it is determined that the preset termination condition is met, the prefrontal cortex module after this round of adjustment can be used as the target prefrontal cortex module; when it is determined that the preset termination condition is not met, the prefrontal cortex module after this round of adjustment can be used as the prefrontal cortex module used in the next round of parameter adjustment.

[0090] Among them, the above test result can be used to characterize whether the accuracy improvement degree of the task execution results output by the above artificial neural network model for each test sample included in the test set is higher than the preset improvement threshold compared with the accuracy of the task execution results output by the sensory cortex module in this round of parameter adjustment for each training sample included in the target sample set used in this round of parameter adjustment.

[0091] In the above content, the method for the server to adjust the connection weights between the neuron nodes in the prefrontal cortex module to be adjusted according to the spike firing order between the connected neuron nodes in the prefrontal cortex module to be adjusted can be that for each neuron node in the prefrontal cortex module to be adjusted, if the neuron node in the next network layer with a connection relationship with this neuron node fires a spike after this neuron node, the connection weight between this neuron node and the next neuron node is increased.

[0092] If a neuron node in the next network layer that has a connection relationship with this neuron node fires a pulse before this neuron node, then reduce the connection weight between this neuron node and the neuron node in the next network layer.

[0093] Among them, the weight adjustment value during the process of increasing or decreasing the connection weights between each neuron node can be determined according to the current connection weight, as well as the difference between the pulse firing time corresponding to this neuron node and the pulse firing time corresponding to the neuron node in the next network layer ( ), and specifically, it can refer to the following formula:

[0094]

[0095]

[0096]

[0097] In the above formula, is the weight adjustment value of the connection weight between the neuron node and the neuron node with a connection relationship. ε (w) represents the relationship that the degree of weight change also depends on the weight size itself. In this specification, = 0.0103, = 0.0051, = 14 ms, = 34 ms.

[0098] Furthermore, after the server monitors the pulse firing order corresponding to each neuron node every time, it can adjust the connection weights between the neuron nodes of the prefrontal cortex module to be adjusted based on the weight adjustment value determined by the above formula, obtain the adjusted connection weights, and determine the adjusted prefrontal cortex module according to the adjusted connection weights.

[0099] For the convenience of understanding, the following takes the above-mentioned training sample sets for different tasks as two training sample sets as an example to elaborate on the above-mentioned continuous learning method based on memory consolidation in detail, specifically as Figure 2 、 Figure 3 shown.

[0100] Figure 2 This is a schematic diagram of the first-round parameter adjustment process provided in this specification.

[0101] Combined with Figure 2 it can be seen that during the first-round parameter adjustment process, the server can use the first training sample set of task one (that is, Figure 2The image data containing the digit "0" and the image data containing the digit "1" in it) are used to adjust the parameters of the sensory cortex module, and the sensory cortex module after the first round of parameter adjustment is obtained. Furthermore, the module parameters of the sensory cortex module after the first round of parameter adjustment can be transferred to the hippocampal module and the prefrontal cortex module to be adjusted.

[0102] Among them, the hippocampal module is used to simulate the hippocampus in the biological brain, and the prefrontal cortex module is used to simulate the cerebral cortex in the biological brain. In Figure 2 each dot in the hippocampal module is used to represent each neuron node in the hippocampal module. Similarly, each dot in the prefrontal cortex module is used to represent each neuron node in the prefrontal cortex module to be adjusted.

[0103] Furthermore, the server can load the connection weights of the hippocampal module into the prefrontal cortex module to be adjusted, and input a preset slow-wave rhythm signal into the prefrontal cortex module to be adjusted, so as to replay the interaction relationship between each neuron node when performing the task corresponding to the first training sample set during the first round of parameter adjustment by the sensory cortex module to the prefrontal cortex module.

[0104] Furthermore, each neuron node in the prefrontal cortex module fires pulses under the action of the slow-wave rhythm signal, so as to achieve the effect of consolidating the interaction relationship between the neuron nodes that play a key role in obtaining the task result.

[0105] Combined with Figure 2 it can be seen that after inputting the slow-wave rhythm signal into the prefrontal cortex module to be adjusted and going through the slow-wave sleep stage, the adjusted prefrontal cortex module is obtained. At this time, Figure 2 the thick black lines between the neuron nodes in the adjusted prefrontal cortex module in

[0106] Furthermore, the server can test the adjusted prefrontal cortex module after the first round of parameter adjustment through the test set corresponding to the first training sample set, obtain the test result, and determine whether to use the adjusted prefrontal cortex module after the first round of parameter adjustment as the prefrontal cortex module in the second round of parameter adjustment according to the test result.

[0107] Figure 3 This is a schematic diagram of the second round of parameter adjustment process provided in this specification.

[0108] Combined with Figure 3 it can be seen that the server can, select a new sensory cortex module for the second round of training, and in the second round of training, the second training sample set of task 2 can be used (that is,Figure 2 The image data containing the digit "7" and the image data containing the digit "8" in are used to train the somatosensory cortex module in the second-round parameter adjustment, obtaining the somatosensory cortex module after the second-round parameter adjustment. Subsequently, the module parameters of the somatosensory cortex module after the second-round parameter adjustment can be transferred to the hippocampal module used in the second-round parameter adjustment. Note that the hippocampal module in the second round is completely different from the hippocampal module in the first round, which corresponds to the hippocampus only retaining short-term memory.

[0109] Furthermore, the server can load the connection weights of the hippocampal module used in the second-round parameter adjustment into the adjusted prefrontal cortex module obtained from the previous-round parameter adjustment (i.e., the prefrontal cortex module to be adjusted in the second-round parameter adjustment), and input the preset slow-wave rhythm signal into the prefrontal cortex module to be adjusted in the second-round parameter adjustment, so as to strengthen the interaction relationship between the neuron nodes that play a key role in obtaining the task result of Task 2 in the prefrontal cortex module to be adjusted.

[0110] It should be noted that in combination with Figure 2 and Figure 3 it can be seen that the content of the task corresponding to the training sample set used in each round of parameter adjustment is encoded by different neurons and connection patterns. Subsequently, according to the order of firing pulses of the neuron nodes included in the prefrontal cortex module under the action of the slow-wave rhythm signal, the neuron nodes and the connection patterns between the neuron nodes used to represent the new task content and the old task content can be adjusted, so that the prefrontal cortex module can retain the content of the learned new task and the old task.

[0111] In this specification, when the server determines that all the training sample sets for different tasks have been used, it can be regarded as meeting the above termination condition. Subsequently, the adjusted prefrontal cortex module after the last round of parameter adjustment can be used as the target prefrontal cortex module, and the task can be executed through the target prefrontal cortex module.

[0112] Among them, the above tasks can be determined according to the actual application scenario, such as: tasks such as image recognition, natural language processing, and information recommendation.

[0113] For example, each of the above training sample sets can be a training sample set containing different types of images. The first training sample set contains images of various bird animals, the second training sample set contains images of various feline animals, etc. Thus, in the first-round parameter adjustment, the first training sample set can be input into the sensory cortex module used in the first-round parameter adjustment, and then the prefrontal cortex module after parameter adjustment can be obtained through the above method. In the second-round parameter adjustment, the second training sample set can be input into the sensory cortex module used in the second-round parameter adjustment, and then the prefrontal cortex module after the second-round parameter adjustment can be obtained through the above method as the target prefrontal cortex module.

[0114] In practical applications, after receiving an image of a bird animal or an image of a feline animal input by the user, the target model can be used to identify the image of the bird animal or the image of the feline animal input by the user.

[0115] In the above content, the target model can be a spiking neural network model that makes up the target prefrontal cortex module, or an artificial neural network model obtained by converting the spiking neural network model that makes up the target prefrontal cortex module.

[0116] From the above content, it can be seen that every time the server needs to use the prefrontal cortex module to learn new task content, it can train the sensory cortex module with the training sample set corresponding to the new task, and transfer the module parameters learned by the sensory cortex module to the hippocampal module. Then, the weights of the hippocampal module can be extracted and loaded into the prefrontal cortex module, so that the prefrontal cortex module superimposes the weights of the hippocampal module. Furthermore, by inputting slow-wave rhythm signals into the prefrontal cortex module, each neuron node included in the prefrontal cortex module can emit pulses under the action of the slow-wave rhythm signals. Then, according to the order of the pulses emitted by each neuron node included in the prefrontal cortex module under the action of the slow-wave rhythm signals, the neuron nodes and the connection patterns between the neuron nodes used to represent the new task content and the old task content can be adjusted, so that the prefrontal cortex module can retain the learned new task content and old task content, alleviate the "catastrophic forgetting" problem in the continuous learning process of the neural network model, and further improve the accuracy of task execution by the neural network model.

[0117] For the sake of easy understanding, the process of task execution by the target model trained by the above method is described in detail below, specifically as Figure 4 shown.

[0118] Figure 4 It is a schematic flowchart of a task execution method provided in this specification, including the following steps:

[0119] S401: Obtain the task data of the task to be executed, where the task data is at least one of image data, text data, and audio data;

[0120] S402: Input the task data into a pre-trained target model to process the task data through the target model to execute the task to be executed and obtain an execution result, where the target model is obtained based on the target prefrontal cortex module trained by the above continuous learning method based on memory consolidation.

[0121] The above is the method for deploying one or more implementation models of this specification. Based on the same idea, this specification also provides a corresponding continuous learning device based on memory consolidation, such as Figure 5 、 Figure 6 shown.

[0122] Figure 5 The following is a schematic diagram of a continuous learning device based on memory consolidation provided by this specification, including:

[0123] An acquisition module 501, configured to acquire each training sample set for different tasks, where each training sample set includes at least one type of data among image data, text data, and audio data;

[0124] A training module 502, configured to perform multiple rounds of parameter adjustment on the module parameters of the prefrontal cortex module according to each training sample set to obtain a target prefrontal cortex module, and perform task execution through the target prefrontal cortex module; wherein, for each round of parameter adjustment, select the target training sample set used in this round of parameter adjustment from each training sample set, and use the prefrontal cortex module after the previous round of parameter adjustment as the prefrontal cortex module to be adjusted used in this round of parameter adjustment; train the sensory cortex module according to the target training sample set to obtain a trained sensory cortex module, transfer the module parameters of the trained sensory cortex module to the hippocampal module, and extract the connection weights between each neuron node included in the hippocampal module as target weights, load the target weights into the prefrontal cortex module to be adjusted, and input a preset slow-wave rhythm signal into the prefrontal cortex module to be adjusted, so as to adjust the connection weights between each neuron node in the prefrontal cortex module to be adjusted according to the spike firing order between each neuron node when the prefrontal cortex module to be adjusted receives the slow-wave rhythm signal, and obtain the prefrontal cortex module after this round of adjustment; when it is determined that a preset termination condition is met, then use the prefrontal cortex module after this round of adjustment as the target prefrontal cortex module.

[0125] Optionally, each network layer included in the hippocampal module corresponds one-to-one with each network layer included in the sensory cortex module. For each network layer included in the hippocampal module, each neuron node included in this network layer corresponds one-to-one with each neuron node included in the corresponding network layer of this network layer in the sensory cortex module.

[0126] Optionally, for each neuron node included in the prefrontal cortex module to be adjusted, this neuron node updates the membrane potential corresponding to this neuron node according to the received electrical signal, the connection weight between this neuron node and the neuron nodes of the previous network layer, and the slow wave phase of the slow wave cycle in which this neuron node is currently located. And if the membrane potential corresponding to this neuron node exceeds the preset pulse threshold, a pulse is sent to the neuron nodes of the next network layer that have a connection relationship with this neuron node; the slow wave phase of the slow wave cycle in which it is currently located includes: the rising phase, the falling phase; wherein, if the slow wave phase of the slow wave cycle in which it is currently located is the rising phase, the update amplitude of the membrane potential corresponding to this neuron node each time is greater than the update amplitude of the membrane potential corresponding to this neuron node when it is currently in the falling phase; the electrical signal includes: at least one of a slow wave rhythm signal and a pulse output by the neuron nodes of the previous network layer that have a connection relationship with this neuron node.

[0127] Optionally, for each neuron node included in the prefrontal cortex module to be adjusted, this neuron node updates the membrane potential corresponding to this neuron node according to the received electrical signal, the connection weight between this neuron node and the neuron nodes of the previous network layer, and the scaling factor corresponding to the slow wave phase of the slow wave cycle in which this neuron node is currently located.

[0128] Optionally, the prefrontal cortex module and the hippocampal module are composed of a spiking neural network model, and the sensory cortex module is composed of a spiking neural network model or an artificial neural network model.

[0129] Figure 6 The figure is a schematic diagram of a task execution device provided in this specification, including:

[0130] A task data acquisition module 601, configured to acquire task data of a task to be executed, where the task data is at least one of image data, text data, and audio data;

[0131] A task execution module 602, configured to input the task data into a pre-trained target model, so as to process the task data through the target model to execute the task to be executed, and obtain an execution result, where the target model is obtained based on the target prefrontal cortex module trained by the above-mentioned continuous learning method based on memory consolidation.

[0132] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-mentioned Figure 1 continuous learning and task execution method based on memory consolidation.

[0133] This specification also provides Figure 7 a schematic structural diagram of an electronic device corresponding to Figure 1 as shown. As Figure 7 described, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the above-mentioned Figure 1 continuous learning and task execution method based on memory consolidation. Of course, in addition to the software implementation method, this specification does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or a logic device.

[0134] Improvements to a technology can be clearly distinguished as either hardware improvements (e.g., improvements to circuit structures such as diodes, transistors, switches, etc.) or software improvements (improvements to method flows). However, with the development of technology, many improvements to method flows today can be regarded as direct improvements to hardware circuit structures. Almost all designers obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that an improvement to a method flow cannot be implemented using a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user programming the device. The designer can program by themselves to "integrate" a digital system on a single PLD, without having to ask a chip manufacturer to design and fabricate a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly implemented using "logic compiler" software, which is similar to the software compiler used in program development and writing. The original code before compilation also has to be written in a specific programming language, which is called a hardware description language (HDL), and there is not just one type of HDL, but many types, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones currently are VHDL (Very-High-Speed Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also be aware that by simply performing a little logical programming on the method flow using the above-mentioned several hardware description languages and programming it into the integrated circuit, it is easy to obtain the hardware circuit that implements the logical method flow.

[0135] The controller can be implemented in any suitable manner. For example, the controller can take the form of, for example, a microprocessor or a processor and a computer-readable medium storing computer-readable program code (such as software or firmware) executable by the (micro)processor, logic gates, switches, an application specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of the controller include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art also know that, in addition to implementing the controller in the form of pure computer-readable program code, it is entirely possible to make the controller implement the same function in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be regarded as the structures within the hardware component. Or even, the devices for implementing various functions can be regarded as either software modules for implementing the method or structures within the hardware component.

[0136] The systems, devices, modules, or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0137] For the convenience of description, when describing the above devices, they are described separately as various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0138] Those skilled in the art should understand that the embodiments of this specification can be provided as a method, a system, or a computer program product. Therefore, this specification can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program code.

[0139] This specification is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the specification. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processors of general purpose computers, special purpose computers, embedded processors, or other programmable data processing devices to produce a machine, such that the instructions executed by the processors of the computer or other programmable data processing devices produce means for implementing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 means for implementing the functions specified in one or more blocks or multiple blocks.

[0140] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means that implement the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 means for implementing the functions specified in one or more blocks or multiple blocks.

[0141] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operational steps are performed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in the Figure 1 one or more flows and / or blocks Figure 1 means for implementing the functions specified in one or more blocks or multiple blocks.

[0142] In a typical configuration, a computing device includes one or more processors (CPUs), an input / output interface, a network interface, and memory.

[0143] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash RAM. The memory is an example of computer-readable media.

[0144] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0145] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0146] It should be understood by those skilled in the art that the embodiments of this specification may be provided as methods, systems or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] This specification may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0148] The various embodiments in this specification are described in a progressive manner. For the parts that are the same or similar among the various embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, reference can be made to the corresponding descriptions in the method embodiments.

[0149] The above description is only for the embodiments of this specification and is not intended to limit this specification. For those skilled in the art, various modifications and changes can be made to this specification. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of this specification shall be included within the scope of the claims of this specification.

Claims

1. A continuous learning method based on memory consolidation, characterized in that: The method is applied to a multi-task continuous learning system, which includes: a prefrontal cortex module, a hippocampus module, and a sensory cortex module, wherein the prefrontal cortex module, the hippocampus module, and the sensory cortex module include network layers, and for each network layer, the network layer includes neuron nodes, and the method includes: Acquire training sample sets for different tasks, wherein each training sample set includes: at least one of image data, text data, and audio data; According to each training sample set, multiple rounds of parameter adjustment are performed on the module parameters of the prefrontal cortex module to obtain a target prefrontal cortex module, and the task is executed through the target prefrontal cortex module; wherein, For each round of parameter adjustment, a target training sample set used in the current round of parameter adjustment is selected from each training sample set, and the prefrontal cortex module after the previous round of parameter adjustment is used as the prefrontal cortex module to be adjusted in the current round of parameter adjustment; The sensory cortex module is trained according to the target training sample set to obtain a trained sensory cortex module, the module parameters of the trained sensory cortex module are transferred to the hippocampus module, and the connection weights between the neuronal nodes contained in the hippocampus module are extracted as the target weights, the target weights are loaded into the prefrontal cortex module to be adjusted, and the preset slow wave rhythm signal is input into the prefrontal cortex module to be adjusted, so as to adjust the connection weights between the neuronal nodes in the prefrontal cortex module to be adjusted according to the pulse emission sequence between the neuronal nodes when the prefrontal cortex module to be adjusted receives the slow wave rhythm signal, so as to obtain the prefrontal cortex module after this round of adjustment; when it is determined that the preset termination condition is met, the prefrontal cortex module after this round of adjustment is used as the target prefrontal cortex module.

2. The method according to claim 1, characterized in that There is a one-to-one correspondence between each network layer included in the hippocampus module and each network layer included in the sensory cortex module. For each network layer included in the hippocampus module, there is a one-to-one correspondence between each neuron node included in the network layer and each neuron node included in the network layer corresponding to the network layer in the sensory cortex module.

3. The method according to claim 1, characterized in that For each neuron node included in the prefrontal cortex module to be adjusted, the neuron node updates the membrane potential corresponding to the neuron node according to the received electrical signal, the connection weight between the neuron node and the neuron node of the previous network layer, and the slow wave stage of the slow wave cycle in which the neuron node is currently located, and if the membrane potential corresponding to the neuron node exceeds a preset pulse threshold, a pulse is issued to the neuron node of the next network layer that is connected to the neuron node; The slow wave stage of the current slow wave cycle includes: a rising stage and a falling stage; wherein, if the slow wave stage of the current slow wave cycle is the rising stage, the update amplitude of the membrane potential corresponding to the neuron node each time is greater than the update amplitude of the membrane potential corresponding to the neuron node when it is currently in the falling stage; the electrical signal includes: a slow wave rhythmic signal, and at least one of the pulses output by the neuron node of the upper network layer that is connected to the neuron node.

4. The method according to claim 3, characterized in that For each neuron node included in the prefrontal cortex module to be adjusted, the neuron node updates the membrane potential corresponding to the neuron node based on the received electrical signal, the connection weight between the neuron node and the neuron node of the previous network layer, and the scaling factor corresponding to the slow wave stage of the slow wave cycle in which the neuron node is currently located.

5. The method according to claim 1, characterized in that The prefrontal cortex module and the hippocampus module are composed of pulse neural network models, and the sensory cortex module is composed of a pulse neural network model or an artificial neural network model.

6. A task execution method, characterized in that: include: Acquire task data of the task to be executed, wherein the task data is at least one of image data, text data, and audio data; The task data is input into a pre-trained target model so that the task data is processed by the target model to execute the task to be executed and obtain an execution result. The target model is obtained based on a target prefrontal cortex module trained by the method described in any one of claims 1 to 5 above.

7. A continuous learning device based on memory consolidation, characterized in that: include: An acquisition module, used to acquire each training sample set for different tasks, wherein each training sample set includes: at least one of image data, text data, and audio data; A training module is used to perform multiple rounds of parameter adjustment on the module parameters of the prefrontal cortex module according to the training sample sets to obtain a target prefrontal cortex module, and perform tasks through the target prefrontal cortex module; wherein, for each round of parameter adjustment, a target training sample set used in the current round of parameter adjustment is selected from the training sample sets, and the prefrontal cortex module after the previous round of parameter adjustment is used as the prefrontal cortex module to be adjusted in the current round of parameter adjustment; the sensory cortex module is trained according to the target training sample set to obtain a trained sensory cortex module, and the module parameters of the trained sensory cortex module are transferred to the hippocampal module block, and extract the connection weights between the neuron nodes contained in the hippocampus module as the target weight, load the target weight into the prefrontal cortex module to be adjusted, and input the preset slow wave rhythm signal into the prefrontal cortex module to be adjusted, so as to adjust the connection weights between the neuron nodes in the prefrontal cortex module to be adjusted according to the pulse emission sequence between the neuron nodes when the prefrontal cortex module to be adjusted receives the slow wave rhythm signal, so as to obtain the prefrontal cortex module after this round of adjustment; when it is determined that the preset termination condition is met, the prefrontal cortex module after this round of adjustment is used as the target prefrontal cortex module.

8. A task execution device, characterized in that: include: A task data acquisition module, used to acquire task data of the task to be executed, wherein the task data is at least one of image data, text data, and audio data; A task execution module, used for inputting the task data into a pre-trained target model so as to process the task data through the target model to execute the task to be executed and obtain an execution result, wherein the target model is obtained based on a target prefrontal cortex module trained by the method described in any one of claims 1 to 5 above.

9. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method described in any one of claims 1 to 6 is implemented.

Images