Model training method and device based on sleep simulation, medium and equipment

Abstract

The invention discloses a model training method and device based on sleep simulation, a medium and equipment. The method comprises the steps of obtaining a trained target image recognition model, wherein the target image recognition model is a pulse neural network model; inputting a specific signal to enable the network to generate spindle wave rhythm oscillation so as to perform non-fast eye movement sleep simulation on the target image recognition model; according to the pulse distribution sequence of the neuron nodes with the connection relation in the target image recognition model in the non-fast eye movement sleep simulation process, the connection weight between the neuron nodes is updated, and the model is finely adjusted according to the updated connection weight between the neuron nodes. According to the scheme, knowledge learned by the model in the training process is strengthened by performing non-fast eye movement sleep simulation on the trained model, so that the model trained based on a small number of samples can obtain relatively high model performance under parameter strengthening in the sleep simulation process.

Description

Technical Field

[0001] The present specification relates to the field of computer technology, and in particular to a model training method, device, medium and equipment based on sleep simulation. Background Art

[0002] The rapid development of artificial intelligence technology has brought about tremendous changes to the world's development model and human lifestyle. In the training process of neural network models, it is usually necessary to use labeled training data to continuously optimize the model's parameters, and the performance of neural network models depends largely on a large amount of training data.

[0003] However, in many real-world fields, due to various restrictions such as privacy protection, it is extremely difficult to obtain specific types of data, and labeling large amounts of data is an expensive and time-consuming task. This situation greatly limits the improvement of model performance.

[0004] Therefore, how to improve model performance with a small number of samples and ensure the recognition accuracy of the model is an urgent problem to be solved. Summary of the invention

[0005] This specification provides a model training method, device, medium and equipment based on sleep simulation to partially solve the above-mentioned problems existing in the prior art.

[0006] This manual adopts the following technical solutions: This manual provides a model training method based on sleep simulation, including: Acquire a trained target image recognition model, wherein the target image recognition model is a pulse neural network model; Inputting a specific signal into the target image recognition model so that the target image recognition model simulates non-rapid eye movement sleep by generating spindle wave rhythm oscillations, wherein the specific signal is a composite signal formed by superposition of two periodic signals, the deviation between the frequencies corresponding to the two periodic signals is less than a preset threshold and is within the spindle wave rhythm range, and the target image recognition model generates response activities close to the spindle wave shape and rhythm as the specific signal is released; According to the pulse emission order of each neuron node having a connection relationship in the target image recognition model during the sleep simulation process, the connection weights between the neuron nodes are updated; The target image recognition model is adjusted according to the updated connection weights between the neuron nodes.

[0007] Optionally, obtaining a trained target image recognition model includes: Acquire a trained initial image recognition model, wherein the initial image recognition model is an artificial neural network model for image recognition; Based on the model data of the initial image recognition model, converting the initial image recognition model into a spiking neural network model to obtain the target image recognition model; The method further comprises: The target image recognition model is adjusted according to the updated connection weights between the neuron nodes, and the adjusted target image recognition model is converted into an artificial neural network model.

[0008] Optionally, issuing spindle pulses to the target image recognition model at preset time intervals specifically includes: For each neuron node except the neuron nodes in the input layer of the target image recognition model, determine the membrane potential of the neuron node in the current time unit according to the membrane potential of the neuron node in the previous time unit, the pulse value output by the neuron node of the upper layer that has an information transmission relationship with the neuron node, the connection weight between the neuron node and the neuron node of the upper layer in the previous time unit, and the preset scaling factor corresponding to the neuron node; If the membrane potential of the neuron node in the current time unit exceeds the preset pulse voltage threshold, a pulse is sent to the next layer of neuron nodes that have an information transmission relationship with the neuron node, and the membrane potential corresponding to the neuron node is reset to the initial membrane potential.

[0009] Optionally, updating the connection weights between the neuron nodes in the target image recognition model according to the pulse emission order of the neuron nodes having a connection relationship in the process of the non-rapid eye movement sleep simulation specifically includes: For each neuron node in the target image recognition model, if a neuron node in a next layer that has an information transmission relationship with the neuron node emits a pulse after the neuron node, then the connection weight between the neuron node and the neuron node in the next layer is increased, and If the next layer of neuron nodes that have an information transmission relationship with the neuron node emits a pulse before the neuron node, the connection weight between the neuron node and the next layer of neuron nodes is reduced.

[0010] Optionally, updating the connection weights between the neuron nodes according to the pulse emission order of the neuron nodes having a connection relationship in the target image recognition model during the non-rapid eye movement sleep simulation specifically includes: For each neuron node in the target image recognition model, a target time scale value is determined from among the preset time scale values ​​according to a pulse emission sequence between the neuron node and a next-layer neuron node having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation process; Determine a weight adjustment factor according to a time difference between a pulse emission time corresponding to the neuron node and a pulse emission time corresponding to a neuron node in a next layer and the target time scale value; According to the weight adjustment factor, an adjustment value of the connection weight between the neuron node and the neuron nodes in the next layer is determined.

[0011] Optionally, the preset time scale value includes: a fast time scale value and a slow time scale value, the fast time scale value includes a first scale value and a second scale value, the slow time scale value includes a third scale value and a fourth scale value, the fast time scale value is smaller than the slow time scale value, the first scale value is smaller than the second scale value, and the third scale value is smaller than the fourth scale value; For each neuron node in the target image recognition model, a target time scale value is determined according to the pulse emission sequence of the neuron node and the next layer of neuron nodes having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation, specifically including: Obtaining a preset time scale value, wherein the time scale value includes: a fast time scale value and a slow time scale value, the fast time scale value includes a first scale value and a second scale value, the slow time scale value includes a third scale value and a fourth scale value, the fast time scale value is smaller than the slow time scale value, the first scale value is smaller than the second scale value, and the third scale value is smaller than the fourth scale value; For each neuron node in the target image recognition model, if the neuron node in the next layer emits a pulse after the neuron node, the first scale value and the fourth scale value are used as the target time scale value; If the neuron node in the next layer emits a pulse after the neuron node, the second scale value and the third scale value are used as the target time scale value.

[0012] Optionally, determining an adjustment value of a connection weight between the neuron node and the neuron node of the next layer according to the weight adjustment factor specifically includes: For each neuron node in the target image recognition model, determining a target weight adjustment item parameter according to a pulse emission order between the neuron node and a next-layer neuron node having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation; The adjustment value of the connection weight between the neuron node and the neuron node of the next layer is determined according to the weight adjustment item parameter and the weight adjustment factor, wherein the larger the target weight adjustment item parameter is, the larger the adjustment value is.

[0013] Optionally, for each neuron node in the target image recognition model, a target weight adjustment item parameter is determined according to a pulse emission sequence between the neuron node and a next-layer neuron node having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation, specifically including: Obtaining preset weight adjustment item parameters, wherein the weight adjustment item parameters include an enhancement parameter and a suppression parameter, the enhancement parameter includes a first enhancement parameter value and a second enhancement parameter value, the suppression parameter includes a first suppression parameter value and a second suppression parameter value, and the values ​​of the parameters are increased in sequence of the first enhancement parameter value, the second suppression parameter value, the second enhancement parameter value, and the first suppression parameter value; For each neuron node in the target image recognition model, if the neuron node in the next layer emits a pulse after the neuron node, the first enhancement parameter value and the second enhancement parameter value are used as the target weight adjustment item parameters; If the neuron node in the next layer emits a pulse after the neuron node, the first inhibition parameter value and the second inhibition parameter value are used as the target weight adjustment item parameters.

[0014] Optionally, the size of the weight adjustment factor is negatively correlated with the absolute value of the adjustment value, the time difference is positively correlated with the size of the weight adjustment factor, and the target time scale value is negatively correlated with the size of the weight adjustment factor.

[0015] This specification provides a model training device based on sleep simulation, including: An acquisition module, used to acquire a trained target image recognition model, wherein the target image recognition model is a pulse neural network model; A simulation module, inputting a specific signal to the target image recognition model so that the target image recognition model simulates non-rapid eye movement sleep by generating spindle wave rhythm oscillations, wherein the specific signal is a composite signal formed by superposition of two periodic signals, the deviation between the frequencies corresponding to the two periodic signals is less than a preset threshold and is within the spindle wave rhythm range, and the target image recognition model generates a response activity close to the spindle wave shape and rhythm as the specific signal is released; An updating module, configured to update the connection weights between the neuron nodes in the target image recognition model according to the pulse emission order of the neuron nodes having a connection relationship in the sleep simulation process; The fine-tuning module is used to adjust the target image recognition model according to the updated connection weights between the neuron nodes.

[0016] This specification provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned model training method based on sleep simulation is implemented.

[0017] This specification provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned model training method based on sleep simulation when executing the program.

[0018] At least one of the above technical solutions adopted in this specification can achieve the following beneficial effects: In the model training method based on sleep simulation provided in the present specification, a trained target image recognition model is obtained, and the target image recognition model is a pulse neural network model; a specific signal is input to make the target image recognition model generate spindle rhythmic oscillations, thereby simulating non-rapid eye movement sleep; according to the pulse emission sequence of each neuron node that has a connection relationship in the target image recognition model during the non-rapid eye movement sleep simulation, the connection weights between each neuron node are updated; and the model is fine-tuned according to the updated connection weights between each neuron node.

[0019] It can be seen from the above method that this scheme inputs a specific signal to make the target image recognition model generate spindle rhythm oscillations, thereby simulating non-rapid eye movement sleep, and then determining the correlation between neurons based on the pulse emission sequence of each neuron node with a connection relationship during the sleep simulation process, and updating the connection weights between each neuron node to further enhance the knowledge learned by the model during the training process. Compared with the current method of training models through a large number of samples, this scheme can further update the model connection weights in the absence of training samples by introducing spindle rhythm, so that the model trained with a few samples can also achieve the model performance of the model trained with more samples by optimizing the weights during the sleep simulation process. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of this specification and constitute a part of this specification. The illustrative embodiments and descriptions of this specification are used to explain this specification and do not constitute an improper limitation on this specification. In the drawings: Figure 1 A flowchart of a model training method based on sleep simulation provided in this specification; Figure 2 A schematic diagram of a connection weight update process provided in this specification; Figure 3 A schematic diagram of the overall optimization process of an image recognition model provided in this specification; Figure 4 A schematic diagram of a model training device based on sleep simulation provided in this specification; Figure 5 A method corresponding to the Figure 1 Schematic diagram of electronic equipment. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solutions and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this specification.

[0022] The technical solutions provided by the embodiments of this specification are described in detail below in conjunction with the accompanying drawings.

[0023] Figure 1 A flowchart of a model training method based on sleep simulation provided in this specification includes the following steps: S101: Acquire a trained target image recognition model, where the target image recognition model is a pulse neural network model.

[0024] The success of current deep learning models relies heavily on a large amount of training data. When the available labeled data is limited, the learning performance is also greatly attenuated. In contrast, human learning usually requires only a small number of examples to learn to recognize new things or skills. Therefore, the concept of few-shot learning was proposed, which aims to learn from a small number of labeled instances and recognize unseen instances. However, the current development of this technology is still very limited, so it is crucial to propose advanced few-shot learning algorithms.

[0025] The biological brain completes memory consolidation through memory replay during non-rapid eye movement (NREM) sleep, converting new knowledge from short-term memory to long-term memory. The biological brain can learn and memorize efficiently through its powerful memory ability.

[0026] In reality, the human brain can still learn quickly under limited or few sample conditions, which shows that combining neural networks with brain-inspired or biologically constrained computing mechanisms can improve the learning efficiency of the network. Neurobiological research shows that human small sample learning ability comes from strong memory ability, and sleep plays an important role in memory consolidation in biological brains. During non-rapid eye movement sleep, the brain's neurons produce complex oscillatory activities and rhythmic patterns, which play a role in memory consolidation.

[0027] Based on this, this specification provides a model training method based on sleep simulation, which introduces the spindle wave oscillation rhythm into the trained pulse neural network model to simulate the memory consolidation process during non-rapid eye movement sleep of the human brain, and then adjusts the network parameters of the model based on the plasticity of synapses.

[0028] In this specification, the execution entity for implementing a model training method based on sleep simulation can be a designated device such as a server. For the sake of ease of description, the following will only take the server as the execution entity as an example to illustrate a model training method based on sleep simulation provided in this specification.

[0029] The server may obtain a trained initial image recognition model. If the initial image recognition model is an artificial neural network (ANNs) model, the server may convert the initial image recognition model into a spiking neural network (SNN) model based on the model data of the initial image recognition model, thereby obtaining a target image recognition model for non-rapid eye movement sleep simulation.

[0030] Among them, the server can construct each neuron node in the target image recognition model (SNN) through leaky integrate and fire (LIF) type neurons, and each neuron node in the target image recognition model corresponds to each computing node in the initial image recognition model (ANN).

[0031] Of course, the initial image recognition model can also be a trained SNN model. In this case, the server can directly use it as the target image recognition model for subsequent non-rapid eye movement sleep simulation.

[0032] Furthermore, in the process of training the target image recognition model (initial image recognition model), the server may obtain a small amount of sample image data, and then input the sample image data into the image recognition model, so as to perform image recognition on the sample image data through the image recognition model to obtain an image recognition result. The server may then determine the loss value of the image recognition model based on the deviation between the image recognition result and the actual label of the sample image data, and then train the image recognition model based on the loss value, thereby obtaining the trained target image recognition model.

[0033] S102: Input a specific signal to the target image recognition model so that the target image recognition model simulates non-rapid eye movement sleep by generating spindle rhythm oscillations, wherein the specific signal is a composite signal formed by the superposition of two periodic signals, the deviation between the frequencies corresponding to the two periodic signals is less than a preset threshold and is within the spindle rhythm range, and the target image recognition model generates response activities that are close to the spindle shape and rhythm as the specific signal is released.

[0034] The server can input a specific signal to make the target image recognition model generate spindle wave rhythm oscillations, thereby simulating non-rapid eye movement sleep. In this specification, the specific signal can be a composite signal formed by the superposition of two periodic signals, and the frequencies of the two periodic signals are similar, that is, the deviation between the corresponding frequencies is within the preset threshold and is within the spindle wave rhythm range. The target image recognition model generates a response activity that matches the spindle wave shape and rhythm as the specific signal is issued. Among them, the response activity that matches the spindle wave shape and rhythm can be a response activity that is the same as the spindle wave shape and rhythm, or it can be a response activity that is close to the spindle wave shape and rhythm. The above-mentioned preset threshold can be set according to the actual situation, and this specification does not make specific restrictions on this.

[0035] In biological concepts, the rhythm of the above spindle waves can range from 8 to 16 Hz, and the duration of the pulse emission phase does not exceed 3 seconds.

[0036] In one embodiment provided in this specification, the rhythm range of the spindle waves can be 8-16 Hz, with a center frequency of 12 Hz.

[0037] Among them, each pulse emission stage corresponds to a spindle wave window. Since the amplitude of the spindle wave gradually decays from the middle to the two sides, there are more pulses at the central peak position in the spindle wave window, and fewer pulses at the edge of the window. In other words, the emission frequency in the spindle wave window changes continuously over time within the preset frequency range.

[0038] The superposition signal for generating spindle rhythm can be generated as follows:

[0039] in, , These are two similar frequencies, ranging from [12-16] Hz, which can generate a composite signal with a spindle wave shape and rhythm after superposition, with a large center and small sides. Inputting this signal into the SNN can induce the SNN to generate a spindle wave rhythm response.

[0040] Of course, in this specification, the target image recognition model can also be used to simulate the awake state of the human brain and compare it with the above-mentioned non-rapid eye movement sleep state. In this awake state, each layer of neuron nodes emits pulses at a pulse emission frequency of 2.5 Hz.

[0041] In the process of sending spindle pulses to the target image recognition model, the neuron nodes in the input layer of the spiking neural network model send pulses based on the time series of the spindle pulses, and the neuron nodes in other network layers send pulses along with the network layers connected to them.

[0042] For each neuron node except the neuron nodes in the input layer of the target image recognition model, the server can determine the membrane potential of the neuron node in the current time unit based on the membrane potential of the neuron node in the previous time unit, the pulse value output by the neuron node in the previous layer that has an information transmission relationship with the neuron node, the connection weight between the neuron node and the neuron node in the previous layer in the previous time unit, and the preset scaling factor corresponding to the neuron node.

[0043] If the membrane potential of the neuron node in the current time unit exceeds the preset pulse voltage threshold, a pulse is sent to the next layer of neuron nodes that have an information transmission relationship with the neuron node, and the membrane potential corresponding to the neuron node is reset to the initial membrane potential.

[0044] Specifically, the relationship between the membrane potential of each neuron node and time can be expressed as:

[0045] in, is the initial membrane potential, For the moment The corresponding membrane potential, for The current input corresponding to the moment, To preset time parameters, For preset resistance parameters, in this manual, .

[0046] The server can assign an initial membrane potential value and a pulse voltage threshold to all neuron nodes in the target image recognition model. In this specification, the initial membrane potential and pulse voltage threshold Set to = -75 mV, = -55 mV. Of course, in practical applications, other initial membrane potentials and pulse voltage thresholds can also be set, and this specification does not specifically limit this.

[0047] The neuron nodes in other network layers except the input layer neuron nodes can receive the pulses emitted by the neuron nodes in the previous network layer, and scale the integrated network connection weights and the inner product emitted by the upper layer neuron nodes by scaling the scaling factor. The current input of the neuron node receiving the pulse integration can be expressed as:

[0048] in, is the current input of the neuron node receiving the pulse integration at time t, is the pulse time series (0 / 1) emitted by the neuron node in the previous layer at time t, is the network connection weight at time t, is the preset scaling factor for Scaling is performed so that the neuron node can receive all the information input by the upper neuron nodes.

[0049] In the feedforward process of the spiking neural network model, each neuron node can integrate the input it receives into its current membrane potential according to the weight, emit a pulse when the membrane potential exceeds the pulse threshold (-55 mV), and simultaneously reset the membrane potential to the initial membrane potential (-75 mV).

[0050] S103: updating the connection weights between the neuron nodes according to the pulse emission order of the neuron nodes having a connection relationship in the target image recognition model during the non-rapid eye movement sleep simulation.

[0051] Specifically, for each neuron node in the target image recognition model, if the neuron node in the next layer that has an information transmission relationship with the neuron node emits a pulse after the neuron node, the connection weight between the neuron node and the neuron nodes in the next layer is increased; if the neuron node in the next layer that has an information transmission relationship with the neuron node emits a pulse before the neuron node, the connection weight between the neuron node and the neuron nodes in the next layer is reduced.

[0052] Furthermore, the spiking neural network updates the connection weights between each neuron node based on the pulse sequence, and determines the presynaptic and postsynaptic neurons according to the information transmission direction of the weight in the feedforward according to the firing state of each neuron. Accordingly, two time scale values ​​are generated for the firing of each neuron, one fast time scale value and the other slow time scale value, for a total of four time scales. When the presynaptic neuron fires before the postsynaptic neuron, the fast time scale value of the presynaptic firing and the slow time scale value of the postsynaptic firing work together to increase the weight; on the contrary, when the presynaptic neuron fires after the postsynaptic neuron, the fast time scale value of the postsynaptic firing and the slow time scale value of the presynaptic firing work together to weaken the weight.

[0053] In this specification, for each neuron node in the target image recognition model, the server determines a target time scale value from each preset time scale value according to the pulse emission sequence of the neuron node and the next layer of neuron nodes that have an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation, and then determines a weight adjustment factor according to the time difference between the pulse emission time corresponding to the neuron node and the pulse emission time corresponding to the next layer of neuron nodes and the target time scale value. , and then determine the adjustment value of the connection weight between the neuron node and the neuron node of the next layer according to the weight adjustment factor.

[0054] In this specification, the preset time scale value may include: a fast time scale value and a slow time scale value, and the fast time scale value includes a first scale value and the second scale value , the slow time scale value includes the third scale value and the fourth scale value , wherein the fast time scale value is smaller than the slow time scale value, the first scale value is smaller than the second scale value, and the third scale value is smaller than the fourth scale value.

[0055] The above time scale values ​​can be set as: =14 ms, =101 ms, =30 ms, =125 ms.

[0056] For each neuron node in the target image recognition model, if the next layer of neuron nodes emits a pulse after the neuron node, the server can and the fourth scale value As the target time scale value; if the next layer of neuron nodes emits pulses after this neuron node, the second scale value and the third scale value as the target time scale value.

[0057] In addition, for each neuron node in the target image recognition model, the server can determine the target weight adjustment item parameter according to the pulse emission sequence of the neuron node and the next layer of neuron nodes that have an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation process. .

[0058] The server can then determine the adjustment value of the connection weight between the neuron node and the next layer of neuron nodes based on the weight adjustment item parameter and the weight adjustment factor, wherein the larger the target weight adjustment item parameter, the larger the adjustment value of the connection weight, and vice versa.

[0059] In this specification, the weight adjustment item parameters include enhancement parameters and suppression parameters, and the enhancement parameters include the first enhancement parameter value and the second enhancement parameter value , the suppression parameters include the first suppression parameter value and the second suppression parameter value The values ​​of the parameters are increased in the order of the first enhancement parameter value, the second suppression parameter value, the second enhancement parameter value, and the first suppression parameter value.

[0060] The above weight adjustment parameter values ​​can be set to, =0.00005, =0.007, =0.006, =0.0023.

[0061] For each neuron node in the target image recognition model, if the next layer of neuron nodes emits pulses after the neuron node, the first enhancement parameter value is and the second enhancement parameter value As the target weight adjustment parameter, if the next layer of neuron nodes emits pulses after this neuron node, the first inhibition parameter value is and the second suppression parameter value As the target weight adjustment parameter.

[0062] For ease of understanding, the adjustment value of the above connection weight can be expressed as:

[0063]

[0064]

[0065] in, A neuron node with a data transmission relationship and neuron nodes The weight adjustment value of the connection weight between The degree of weight change also depends on the weight itself. When the presynaptic neuron emits pulses before the postsynaptic neuron, the corresponding weight increase stage , when the postsynaptic neuron node emits pulses before the presynaptic neuron node, the corresponding weight decreases .

[0066] And, the weight adjustment factor Size and weight adjustment value The absolute value of is negatively correlated with the time difference It is positively correlated with the size of the weight adjustment factor, and the target time scale value It is negatively correlated with the size of the weight adjustment factor.

[0067] In this process, for each presynaptic neuron node, if the presynaptic neuron node emits a pulse before the postsynaptic neuron node connected to it, it means that the postsynaptic neuron node emits a pulse along with the presynaptic neuron node, and the correlation between the presynaptic neuron node and the postsynaptic neuron node is stronger than the correlation between other neuron nodes and the postsynaptic neuron node. In addition, the time difference between the two in emitting pulses is The smaller the value, the stronger the correlation between the two. At this time, the adjustment value (enhancement value) of the connection weight can be increased, and the time difference between the two pulses is The larger the value, the weaker the correlation between the two. In this case, the adjustment value (enhancement value) of the connection weight can be reduced.

[0068] If the postsynaptic neuron emits pulses before the presynaptic neuron connected to it, it means that the postsynaptic neuron emits pulses along with the other presynaptic neuron nodes connected to it, and the correlation between the presynaptic neuron node and the postsynaptic neuron node is weaker than the correlation between other neuron nodes and the postsynaptic neuron node. The smaller the value, the stronger the correlation between the two. At this time, the adjustment value (inhibition value) of the connection weight can be increased, and the time difference between the two pulses is The larger the value, the weaker the correlation between the two. In this case, the adjustment value (inhibition value) of the connection weight can be reduced.

[0069] For ease of understanding, this specification provides a schematic diagram of the update process of the connection weights, such as Figure 2 shown.

[0070] Figure 2 A schematic diagram of a connection weight update process provided in this specification.

[0071] Among them, in the training stage of the target image recognition model, a small number of digital "0" samples in the MNIST handwritten data set can be used as an illustration (the number is recorded as the proportion K of the original data set). In actual training, K can be adjusted to control the size of the training set. For the sake of clarity, 10% of the samples of the original data set of a single category (K=10%) are used as a demonstration. It should be noted that this algorithm can use a variety of open source data sets, such as ImageNet, CIFAR, etc. In this example, the network is trained with a small sample data set to complete the recognition of the number "0", and a "0" that has not been seen outside the training set is used for testing to determine whether the image is 0.

[0072] After completing the training of the target image recognition model, the server can update the connection weights between neuronal nodes based on the spindle rhythm response through non-rapid eye movement sleep simulation, thereby obtaining the updated connection weights between the nodes.

[0073] S104: Adjusting the target image recognition model according to the updated connection weights between the neuron nodes.

[0074] Each time the server sends a spindle pulse to the target image recognition model, it continuously updates and iterates the connection weights between each neuron node, thereby adjusting the target image recognition model without samples, that is, fine-tuning and consolidating the model.

[0075] When it is determined that the preset training conditions are met (such as reaching the preset training rounds or the model converges to the preset range), the server can convert the target image recognition model into an artificial neural network model and then test it. The server can use other unseen "0"s for classification / recognition tests. In addition, the number of samples K can be adjusted to control the size of the data set and further explore the performance of the algorithm on small sample tasks.

[0076] For ease of understanding, this specification provides a schematic diagram of the overall optimization process of an image recognition model, such as Figure 3 shown.

[0077] Figure 3 A schematic diagram of the overall optimization process of an image recognition model provided in this specification.

[0078] Among them, the training stage and the testing stage correspond to the brain's awake state, while the fine-tuning (consolidation) stage corresponds to the brain's non-rapid eye movement sleep state. Through the training-consolidation-testing process, the model parameters can be further optimized so that it can still achieve the performance of the model trained by a large number of training samples with fewer training samples.

[0079] After completing the testing of the image recognition model, the server can deploy it for subsequent image recognition tasks.

[0080] In practical applications, the server can obtain image data to be recognized, and input the image data into the image recognition model trained by the above method, so as to obtain the image recognition result of the image to be recognized through the image recognition model.

[0081] It should be pointed out that the above model training method can be applied not only to image recognition models, but also to other neural network models such as natural language processing, anomaly detection, instance segmentation, and corresponding practical tasks.

[0082] From the above method, it can be seen that this scheme achieves the learning ability of small sample tasks by simulating the spindle oscillation during non-rapid eye movement sleep of biological brains and combining the principle of neural plasticity of biological brains. In learning tasks, the learned content is consolidated by simulating sleep, reducing the dependence on large-scale labeled data, and achieving the performance of large-scale training set learning in fewer learning samples.

[0083] The above is one or more embodiments of the model training method based on sleep simulation in this specification. Based on the same idea, this specification also provides a corresponding model training device based on sleep simulation, such as Figure 4 shown.

[0084] Figure 4 A schematic diagram of a model training device based on sleep simulation provided in this specification includes: An acquisition module 401 is used to acquire a trained target image recognition model, wherein the target image recognition model is a pulse neural network model; A simulation module 402 is used to input a specific signal to the target image recognition model so that the target image recognition model performs non-rapid eye movement sleep simulation by generating spindle wave rhythm oscillations, wherein the specific signal is a composite signal formed by superposition of two periodic signals, the deviation between the frequencies corresponding to the two periodic signals is less than a preset threshold and is within the spindle wave rhythm range, and the target image recognition model generates a response activity close to the spindle wave shape and rhythm as the specific signal is released; An updating module 403 is used to update the connection weights between the neuron nodes according to the pulse emission order of the neuron nodes having a connection relationship in the target image recognition model during the sleep simulation process; The fine-tuning module 404 is used to adjust the target image recognition model according to the updated connection weights between the neuron nodes.

[0085] Optionally, the acquisition module 401 is specifically used to acquire a trained initial image recognition model, wherein the initial image recognition model is an artificial neural network model; based on the model data of the initial image recognition model, the initial image recognition model is converted into a pulse neural network model to obtain the target image recognition model; The fine-tuning module 404 is also used to adjust the target image recognition model according to the updated connection weights between the neuron nodes, and convert the adjusted target image recognition model into an artificial neural network model.

[0086] Optionally, the simulation module 402 is specifically used to determine the membrane potential of each neuron node except the neuron nodes in the input layer of the target image recognition model in the current time unit according to the membrane potential of the neuron node in the previous time unit, the pulse value output by the neuron node in the previous layer that has an information transmission relationship with the neuron node, the connection weight between the neuron node and the neuron node in the previous time unit, and a preset scaling factor corresponding to the neuron node; if the membrane potential of the neuron node in the current time unit exceeds a preset pulse voltage threshold, a pulse is sent to the neuron node in the next layer that has an information transmission relationship with the neuron node, and the membrane potential corresponding to the neuron node is reset to the initial membrane potential.

[0087] Optionally, the update module 403 is specifically used to, for each neuron node in the target image recognition model, increase the connection weight between the neuron node and the neuron nodes in the next layer if the neuron node in the next layer that has an information transmission relationship with the neuron node emits a pulse after the neuron node; and reduce the connection weight between the neuron node and the neuron nodes in the next layer if the neuron node in the next layer that has an information transmission relationship with the neuron node emits a pulse before the neuron node.

[0088] Optionally, the update module 403 is specifically used to determine, for each neuron node in the target image recognition model, a target time scale value from each preset time scale value according to the pulse firing order of the neuron node and the next layer of neuron nodes that have an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation; determine a weight adjustment factor according to the time difference between the pulse firing time corresponding to the neuron node and the pulse firing time corresponding to the next layer of neuron nodes and the target time scale value; and determine the adjustment value of the connection weight between the neuron node and the next layer of neuron nodes according to the weight adjustment factor.

[0089] Optionally, the preset time scale value includes: a fast time scale value and a slow time scale value, the fast time scale value includes a first scale value and a second scale value, the slow time scale value includes a third scale value and a fourth scale value, the fast time scale value is smaller than the slow time scale value, the first scale value is smaller than the second scale value, and the third scale value is smaller than the fourth scale value; The update module 403 is specifically used to, for each neuron node in the target image recognition model, if the next-layer neuron node emits a pulse after the neuron node, use the first scale value and the fourth scale value as the target time scale value; if the next-layer neuron node emits a pulse after the neuron node, use the second scale value and the third scale value as the target time scale value.

[0090] Optionally, the update module 403 is specifically used to determine, for each neuron node in the target image recognition model, a target weight adjustment item parameter according to the pulse emission order of the neuron node and the next layer of neuron nodes that have an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation; and determine the adjustment value of the connection weight between the neuron node and the next layer of neuron nodes according to the weight adjustment item parameter and the weight adjustment factor, wherein the larger the target weight adjustment item parameter, the larger the adjustment value.

[0091] Optionally, the update module 403 is specifically used to obtain preset weight adjustment item parameters, wherein the weight adjustment item parameters include enhancement parameters and suppression parameters, the enhancement parameters include a first enhancement parameter value and a second enhancement parameter value, the suppression parameters include a first suppression parameter value and a second suppression parameter value, and the size of each parameter value increases in the order of the first enhancement parameter value, the second suppression parameter value, the second enhancement parameter value, and the first suppression parameter value; for each neuron node in the target image recognition model, if the next layer of neuron nodes emits pulses after the neuron node, the first enhancement parameter value and the second enhancement parameter value are used as the target weight adjustment item parameters; if the next layer of neuron nodes emits pulses after the neuron node, the first suppression parameter value and the second suppression parameter value are used as the target weight adjustment item parameters.

[0092] Optionally, the size of the weight adjustment factor is negatively correlated with the absolute value of the adjustment value, the time difference is positively correlated with the size of the weight adjustment factor, and the target time scale value is negatively correlated with the size of the weight adjustment factor.

[0093] This specification also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 A model training method based on sleep simulation is provided.

[0094] This manual also provides Figure 5 The one shown corresponds to Figure 1 A schematic diagram of the electronic device. Figure 5 As mentioned above, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The model training method based on sleep simulation. Of course, in addition to the software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0095] For the improvement of a technology, it can be clearly distinguished whether it is a hardware improvement (for example, improvement of the circuit structure of diodes, transistors, switches, etc.) or a software improvement (improvement of the method flow). However, with the development of technology, many improvements of the method flow today can be regarded as direct improvements of the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that the improvement of a method flow cannot be implemented with 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's programming of the device. Designers can "integrate" a digital system on a PLD by programming themselves, without having to ask chip manufacturers to design and make dedicated integrated circuit chips. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one kind of HDL, but many kinds, 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 are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0096] The controller may be implemented in any suitable manner, for example, the controller may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (e.g., software or firmware) executable by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include but are not limited to the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, and the memory controller may also be implemented as part of the control logic of the memory. It is also known to those skilled in the art that, in addition to implementing the controller in a purely computer-readable program code manner, the controller may be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, such a controller may be considered as a hardware component, and the devices for implementing various functions included therein may also be considered as structures within the hardware component. Or even, the devices for implementing various functions may be considered as both software modules for implementing the method and structures within the hardware component.

[0097] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may 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.

[0098] For the convenience of description, the above device is described in 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.

[0099] Those skilled in the art will appreciate 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.

[0100] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0101] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0103] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0104] The memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. The memory is an example of a computer-readable medium.

[0105] 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 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.

[0106] 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.

[0107] 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.

[0108] 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.

[0109] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0110] The above description is only an embodiment of the present specification and is not intended to limit the present specification. For those skilled in the art, the present specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification shall be included in the scope of the claims of the present specification.

Claims

1. A model training method based on sleep simulation, characterized in that: include: Acquire a trained target image recognition model, wherein the target image recognition model is a pulse neural network model; Inputting a specific signal into the target image recognition model so that the target image recognition model simulates non-rapid eye movement sleep by generating spindle wave rhythm oscillations, wherein the specific signal is a composite signal formed by superposition of two periodic signals, the deviation between the frequencies corresponding to the two periodic signals is less than a preset threshold and is within the spindle wave rhythm range, and the target image recognition model generates response activities close to the spindle wave shape and rhythm as the specific signal is released; According to the pulse emission order of each neuron node having a connection relationship in the target image recognition model during the sleep simulation process, the connection weights between the neuron nodes are updated; The target image recognition model is adjusted according to the updated connection weights between the neuron nodes.

2. The method according to claim 1, characterized in that Get the trained target image recognition model, including: Acquire a trained initial image recognition model, wherein the initial image recognition model is an artificial neural network model; Based on the model data of the initial image recognition model, converting the initial image recognition model into a spiking neural network model to obtain the target image recognition model; The method further comprises: The target image recognition model is adjusted according to the updated connection weights between the neuron nodes, and the adjusted target image recognition model is converted into an artificial neural network model.

3. The method according to claim 1, characterized in that At a preset time interval, the spindle pulses are sent to the target image recognition model, specifically including: For each neuron node except the neuron nodes in the input layer of the target image recognition model, determine the membrane potential of the neuron node in the current time unit according to the membrane potential of the neuron node in the previous time unit, the pulse value output by the neuron node of the upper layer that has an information transmission relationship with the neuron node, the connection weight between the neuron node and the neuron node of the upper layer in the previous time unit, and the preset scaling factor corresponding to the neuron node; If the membrane potential of the neuron node in the current time unit exceeds the preset pulse voltage threshold, a pulse is sent to the next layer of neuron nodes that have an information transmission relationship with the neuron node, and the membrane potential corresponding to the neuron node is reset to the initial membrane potential.

4. The method according to claim 1, characterized in that According to the pulse emission order of each neuron node having a connection relationship in the target image recognition model during the non-rapid eye movement sleep simulation, the connection weights between the neuron nodes are updated, specifically including: For each neuron node in the target image recognition model, if a neuron node in a next layer that has an information transmission relationship with the neuron node emits a pulse after the neuron node, then the connection weight between the neuron node and the neuron node in the next layer is increased, and If the next layer of neuron nodes that have an information transmission relationship with the neuron node emits a pulse before the neuron node, the connection weight between the neuron node and the next layer of neuron nodes is reduced.

5. The method according to claim 1, characterized in that According to the pulse emission order of each neuron node having a connection relationship in the target image recognition model during the non-rapid eye movement sleep simulation, the connection weights between the neuron nodes are updated, specifically including: For each neuron node in the target image recognition model, a target time scale value is determined from among the preset time scale values ​​according to a pulse emission sequence between the neuron node and a next-layer neuron node having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation process; Determine a weight adjustment factor according to a time difference between a pulse emission time corresponding to the neuron node and a pulse emission time corresponding to a neuron node in a next layer and the target time scale value; According to the weight adjustment factor, an adjustment value of the connection weight between the neuron node and the neuron nodes in the next layer is determined.

6. The method according to claim 5, characterized in that The preset time scale value includes: a fast time scale value and a slow time scale value, the fast time scale value includes a first scale value and a second scale value, the slow time scale value includes a third scale value and a fourth scale value, the fast time scale value is smaller than the slow time scale value, the first scale value is smaller than the second scale value, and the third scale value is smaller than the fourth scale value; For each neuron node in the target image recognition model, a target time scale value is determined according to the pulse emission sequence of the neuron node and the next layer of neuron nodes having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation, specifically including: For each neuron node in the target image recognition model, if the neuron node in the next layer emits a pulse after the neuron node, the first scale value and the fourth scale value are used as the target time scale value; If the neuron node in the next layer emits a pulse after the neuron node, the second scale value and the third scale value are used as the target time scale value.

7. The method according to claim 5, characterized in that Determining the adjustment value of the connection weight between the neuron node and the neuron node of the next layer according to the weight adjustment factor specifically includes: For each neuron node in the target image recognition model, determining a target weight adjustment item parameter according to a pulse emission order between the neuron node and a next-layer neuron node having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation; The adjustment value of the connection weight between the neuron node and the neuron node of the next layer is determined according to the weight adjustment item parameter and the weight adjustment factor, wherein the larger the target weight adjustment item parameter is, the larger the adjustment value is.

8. The method according to claim 7, characterized in that For each neuron node in the target image recognition model, the target weight adjustment item parameter is determined according to the pulse emission sequence of the neuron node and the next layer of neuron nodes having an information transmission relationship with the neuron node during the non-rapid eye movement sleep simulation, specifically including: Obtaining preset weight adjustment item parameters, wherein the weight adjustment item parameters include an enhancement parameter and a suppression parameter, the enhancement parameter includes a first enhancement parameter value and a second enhancement parameter value, the suppression parameter includes a first suppression parameter value and a second suppression parameter value, and the values ​​of the parameters are increased in sequence of the first enhancement parameter value, the second suppression parameter value, the second enhancement parameter value, and the first suppression parameter value; For each neuron node in the target image recognition model, if the neuron node in the next layer emits a pulse after the neuron node, the first enhancement parameter value and the second enhancement parameter value are used as the target weight adjustment item parameters; If the neuron node in the next layer emits a pulse after the neuron node, the first inhibition parameter value and the second inhibition parameter value are used as the target weight adjustment item parameters.

9. The method according to claim 6, characterized in that The size of the weight adjustment factor is negatively correlated with the absolute value of the adjustment value, the time difference is positively correlated with the size of the weight adjustment factor, and the target time scale value is negatively correlated with the size of the weight adjustment factor.

10. A model training device based on sleep simulation, characterized in that: include: An acquisition module, used to acquire a trained target image recognition model, wherein the target image recognition model is a pulse neural network model; A simulation module, used for inputting a specific signal into the target image recognition model so that the target image recognition model simulates non-rapid eye movement sleep by generating spindle wave rhythm oscillations, wherein the specific signal is a composite signal formed by superposition of two periodic signals, the deviation between the frequencies corresponding to the two periodic signals is less than a preset threshold and is within the spindle wave rhythm range, and the target image recognition model generates a response activity close to the spindle wave shape and rhythm as the specific signal is released; An updating module, configured to update the connection weights between the neuron nodes in the target image recognition model according to the pulse emission order of the neuron nodes having a connection relationship in the sleep simulation process; The fine-tuning module is used to fine-tune the target image recognition model according to the updated connection weights between the neuron nodes.

11. 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 9 is implemented.

12. 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 9 is implemented.

Images