A brain-inspired deep neural network enhancement method under limited data

Abstract

The application provides a deep neural network enhancement method based on brain inspiration under limited data, belongs to the technical field of deep learning, and refers to the asynchronous working mechanism of the hippocampus-frontal lobe circuit in the human brain to develop a general asynchronous learning framework. First, a given deep neural network model is divided into two small-scale submodels, and the submodels are caused to undertake the feature extraction and prediction functions of the original network respectively. Then, the submodels are used for model reconstruction to obtain the asynchronous architecture of the given deep neural network. Finally, the reconstructed model is trained by using the designed asynchronous training method, and the feature extractor and the predictor are independently optimized. The application overcomes the problems of parameter over-coupling and high dependence on a large amount of training data in the prior art without impairing the feature extraction capability of the deep neural network.

Description

Technical Field

[0001] This invention belongs to the field of deep learning technology, specifically, it relates to a brain-inspired deep neural network enhancement method under limited data. Background Technology

[0002] In recent years, deep neural network models, with their large number of parameters and complex architectures, have become a widely used machine learning method due to their powerful feature extraction and learning capabilities. However, deep neural networks suffer from several drawbacks in practical applications. The most typical of these is parameter overcoupling. Due to the inherent characteristics of deep neural network models, each parameter within the model is heavily constrained by a large number of related parameters during training, thus hindering in-depth optimization of those parameters. This problem severely limits the optimization of some key parameters (such as those in the last few network layers), resulting in a predictive performance ceiling for deep neural network models that is below their inherent potential and difficult to surpass. Moreover, this problem becomes more severe as the model size increases.

[0003] Furthermore, large-scale deep neural network models require training on massive datasets. However, in practical applications, it is often difficult to obtain sufficient training samples for deep neural networks. Especially in fields where data is scarce and privacy is sensitive, the cost of collecting and labeling large amounts of high-quality data is extremely high, which greatly limits the practical application of deep neural networks in many fields.

[0004] Intuitively, these challenges can be mitigated by simplifying the neural network model architecture, but this approach will impair the feature extraction capabilities of deep neural networks, thereby affecting the model's final predictive performance.

[0005] Unlike current deep neural network models, the human brain does not suffer from the aforementioned drawbacks such as over-coupling of parameters and high dependence on massive amounts of data. This is because the internal regions of the human brain operate asynchronously in certain situations. For example, the hippocampus-prefrontal cortex circuit employs an asynchronous working strategy, avoiding mutual constraints between its internal memory and decision-making processes, thus maintaining high efficiency under limited resources. The specific details of its asynchronous operation are as follows... Figure 1 As shown. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a brain-inspired deep neural network enhancement method for limited data. Referring to the asynchronous working mechanism of the hippocampus-prefrontal loop in the human brain, a general asynchronous learning framework is developed. This framework overcomes the problems of parameter over-coupling and high dependence on large amounts of training data without compromising the feature extraction capabilities of deep neural networks.

[0007] This invention is achieved through the following technical solution:

[0008] A brain-inspired deep neural network augmentation method for limited data:

[0009] The method specifically includes the following steps:

[0010] Step 1: Select the backbone model and divide the region;

[0011] Step 2: Reconstruct the model based on the regions defined in Step 1 to obtain the asynchronous architecture of the deep neural network;

[0012] Step 3: Design an asynchronous training method to train the model reconstructed in Step 2, and independently optimize its feature extractor and predictor respectively;

[0013] Step 4, Model Prediction: Use the model trained in Step 3 for practical application and output the prediction results.

[0014] Further, in step 1,

[0015] Choose a deep feedforward neural network (DNN) model with N layers as the backbone model of the asynchronous learning framework;

[0016] The selected DNN is divided into two parts: the feature extractor FE-net and the predictor IP-net;

[0017] The feature extractor FE-net includes an input layer and a network layer in front of the DNN, which is responsible for reading input data and extracting features.

[0018] The predictor IP-net contains the last few layers of the DNN network, which is responsible for integrating the features extracted by FE-net and completing the prediction.

[0019] Furthermore, in step 2,

[0020] An auxiliary output layer (AO) is introduced, with the same structure and settings as the output layer of a DNN.

[0021] The original sequential connection between the feature extractor FE-net and the predictor IP-net is modified so that the feature extractor FE-net is connected in parallel to the predictor IP-net and the auxiliary output layer AO, forming the asynchronous architecture DNN-A of DNN.

[0022] Furthermore, in step 3, the asynchronous training method includes two stages: feature extraction training and state prediction training.

[0023] The feature extraction training:

[0024] By combining the auxiliary output layer AO with the feature extractor FE-net, a primary implementation mode of DNN-A, DNN-AJ, is formed.

[0025] DNN-AJ is trained using training data (X,Y) to optimize the parameters in the feature extractor FE-net;

[0026] The state prediction training:

[0027] The trained feature extractor FE-net and predictor IP-net are combined to generate the advanced implementation mode of DNN-A, DNN-AS.

[0028] The advanced implementation pattern DNN-AS is trained using training data (X,Y), but only the parameters in IP-net are optimized, while the parameters in FE-net remain unchanged.

[0029] Furthermore, in step 4,

[0030] The trained DNN-AS is used to predict the input data and generate prediction results.

[0031] A brain-inspired deep neural network augmentation system for limited data:

[0032] The system includes a region partitioning module, a model reconstruction module, and an asynchronous training module.

[0033] The region division module is used to perform region division after selecting the backbone model;

[0034] The model reconstruction module reconstructs the model according to the regions divided by the region division module to obtain the asynchronous architecture of the given deep neural network.

[0035] The asynchronous training module designs an asynchronous training method to train the model reconstructed by the model reconstruction module, independently optimize its feature extractor and predictor, and output prediction results based on the reconstructed model in practical applications.

[0036] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0037] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0038] Beneficial effects of the invention

[0039] This invention separates the feature extractor (FE-net) and the predictor (IP-net) and optimizes them independently; this separation reduces the mutual influence between parameters, allowing each part to be optimized more flexibly to achieve better performance.

[0040] Compared to the original DNN, the FE-net and IP-net of this invention are smaller in scale and simpler in architecture, requiring less training data to achieve optimization. This reduces the dependence of deep neural network models on a large amount of training data, making model training more efficient. At the same time, the required computing resources are also relatively less, which reduces the computational cost during the training process to a certain extent.

[0041] This invention reduces the problem of parameter over-coupling through asynchronous training, rather than directly processing specific physical parameters or data, thus optimizing the training process of DNNs and improving the performance and efficiency of the model.

[0042] Because asynchronous training algorithms do not simplify the architecture and parameters of a given deep neural network model, this method does not impair the model's feature extraction and learning capabilities.

[0043] This invention, through asynchronous training, allows the model to focus on different optimization objectives at different stages, thereby optimizing feature extraction and state prediction capabilities, and is more efficient than traditional synchronous training methods.

[0044] The asynchronous learning framework proposed in this invention is a general method that can be applied to various classic deep neural networks and has good versatility. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the asynchronous operation of the hippocampus-prefrontal loop.

[0046] Figure 2 This is a diagram of the asynchronous learning framework of the present invention.

[0047] Figure 3 This chart compares the performance improvement of asynchronous learning methods on deep neural network models under different data scales.

[0048] Figure 4 A comparison chart showing the improvement in robustness of asynchronous learning frameworks on deep neural networks;

[0049] exist Figure 3 and Figure 4 middle,

[0050] (a) is MNIST & VGG16, (b) is Fashion-MNIST & VGG16.

[0051] (c) is CIFAR10 & VGG16, (d) is MNIST & MobileNet.

[0052] (e) is Fashion-MNIST & MobileNet, (f) is CIFAR10 & MobileNet;

[0053] Figure 5 This is a comparison chart showing the improvement in stability of asynchronous learning frameworks for deep neural networks.

[0054] exist Figure 5 In the image, (a) represents MNIST & VGG16, and (b) represents MNIST & MobileNet.

[0055] (c) is Fasten-MNIST & VGG16, (d) is Fasten-MNIST & MobileNet.

[0056] (e) is CIFAR10 & VGG16, (f) is CIFAR10 & MobileNet. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] like Figure 1 As shown, the hippocampo-prefrontal loop consists of memory and decision-making processes. During memory, the prefrontal cortex senses external information and transmits it to the hippocampus for storage. During decision-making, the hippocampus retrieves existing memory information and transmits it to the prefrontal cortex, which then makes the current decision based on this information. Generally, these two processes in the hippocampo-prefrontal loop work synchronously, enabling the simultaneous memorization of new knowledge and decision-making based on historical memory. However, in some cases, these processes are executed asynchronously to overcome resource constraints and maintain high efficiency.

[0059] Specifically, during emergency decision-making, the memory process within the aforementioned circuit is somewhat inhibited, while the decision-making process receives more attention and is strengthened. This reduces the attention and resources the brain allocates to the memory process, thus avoiding the restriction and competition between the memory process and the decision-making process within the circuit. Simultaneously, this operation also reduces interference from new memory storage in the hippocampus on historical memory retrieval. Conversely, during rote memorization, the decision-making process is inhibited to prevent interference and restriction from the decision-making process within the aforementioned circuit, allowing the brain to focus more on the memory process.

[0060] Based on the aforementioned asynchronous working mechanism, this invention proposes asynchronous learning as a general deep neural network enhancement framework to overcome the shortcomings of over-coupling of internal parameters and heavy reliance on large amounts of training data in deep neural networks. Specific details of the asynchronous learning method are as follows... Figure 2 As shown.

[0061] This invention selects a deep hierarchical feedforward neural network (DNN) with N layers as the backbone model of the proposed asynchronous learning framework. Furthermore, the designed asynchronous learning framework is a general method that can be applied to various classic deep neural networks; therefore, the choice of backbone model is arbitrary in practical applications.

[0062] A brain-inspired deep neural network augmentation method for limited data:

[0063] The method specifically includes the following steps:

[0064] Step 1: Select the backbone model and divide the region;

[0065] Choose a deep hierarchical feedforward neural network (DNN) with N layers as the backbone model of the asynchronous learning framework;

[0066] An arbitrary deep neural network model typically consists of two parts: a feature extractor and a predictor. In practice, the feature extractor first extracts features from the input data to generate abstract features, and then the predictor integrates and computes these abstract features to make subsequent predictions.

[0067] The selected DNN is divided into two small sub-models: the feature extractor (FE-net) and the predictor (IP-net).

[0068] In this model, FE-net serves as the feature extractor for the DNN, comprising an input layer and most of the network layers preceding the DNN, used to read the input data and extract relevant features. IP-net acts as the model's predictor, containing the last few network layers of the DNN, used to integrate the abstract features extracted by FE-net and complete the final prediction.

[0069] Step 2: Reconstruct the sub-models from Step 1 to obtain the asynchronous architecture of the given deep neural network;

[0070] After region partitioning, the DNN model architecture is reconstructed based on the resulting sub-models. Specifically, this invention first introduces an auxiliary output layer (AO), whose structure and settings are identical to the DNN's output layer. Then, this invention modifies the original sequential connection method between the FE-net and IP-net, enabling the FE-net to connect to the IP-net and AO in parallel, thereby obtaining the asynchronous architecture of the DNN (DNN-A).

[0071] The purpose of introducing an AO layer into the DNN in this invention is to assist FE-net in training, thereby overcoming the defect that FE-net cannot use supervised information to optimize its own parameters.

[0072] Step 3: Train the reconstructed model from Step 2 using the designed asynchronous training method, and optimize its feature extractor and predictor independently.

[0073] An asynchronous training method is proposed to independently optimize the feature extraction and state prediction capabilities of the reconstructed DNN-A. The specific details of the designed asynchronous training algorithm are shown in Algorithm 1.

[0074]

[0075] Specific details of Algorithm 1 asynchronous training algorithm

[0076] As shown in Algorithm 1, the proposed asynchronous training method includes two stages: feature extraction training and state prediction training, which independently optimize the feature extractor FE-net and the predictor IP-net of DNN-A.

[0077] In the feature extraction training phase, this invention first combines AO with FE-net to form a primary implementation mode of DNN-A, which is named DNN-AJ. Then, this invention trains DNN-AJ based on training data (X,Y) to optimize the parameters in FE-net (see lines 3-5 of Algorithm 1).

[0078] The feature extraction training phase optimizes the parameters in FE-net to enable it to extract features from the input data more accurately. Specifically, this invention compares the output of AO with the supervision information Y to calculate the error, and then backpropagates this error to FE-net to aid its training.

[0079] Subsequently, during the state prediction training phase, this invention combines the trained FE-net with IP-net to generate a high-level implementation of DNN, DNN-AS, and trains DNN-AS using X and Y. During the training of DNN-AS, this invention keeps the parameter values ​​in FE-net unchanged, optimizing only the parameters in IP-net (see lines 6-9 of Algorithm 1). Here, Y serves as the target output, helping IP-net learn how to predict the correct result based on the features extracted by FE-net.

[0080] Step 4, Model Prediction: Use the trained DNN-AS for practical application and output the prediction results.

[0081] After the above training process is completed, this invention is applied in practice based on the trained DNN-AS (see lines 10-11 of Algorithm 1). The trained DNN-AS is used to predict new input data X and generate prediction result Y'.

[0082] Y' is the result of a deep neural network (DNN) analyzing and processing data X after training using an asynchronous learning framework. This predicted output is generated by the model based on patterns and correlations learned from a large amount of labeled data (X,Y) during the training phase. During training, the model continuously adjusts its internal parameters, learning how to extract key information from input features and map it to corresponding output labels. As training progresses, the model gradually optimizes its feature extraction and prediction capabilities to achieve higher performance.

[0083] When faced with new input data, the model will use the knowledge gained during training to generate a predicted output Y'. This predicted output is the best estimate of the potential labels of the new data.

[0084] Furthermore, the core reason why the asynchronous learning framework proposed in this invention can enhance deep neural network models lies in its separation of the feature extractor FE-net and the predictor IP-net within the DNN, allowing them to be executed asynchronously and optimized independently during training. By doing so, the feature extractor and predictor within the DNN are freed from the constraints of each other's complex structures and numerous parameters during training, making their optimization more flexible and easier to train to the optimal state. This greatly alleviates the parameter over-coupling defect suffered by DNNs.

[0085] Meanwhile, FE-net and IP-net are relatively smaller in scale and simpler in architecture compared to DNNs, requiring only a small amount of training data for optimization. Therefore, the asynchronous learning framework proposed in this invention can greatly reduce the dependence of deep neural network models on massive amounts of training data.

[0086] Example:

[0087] (I) Experimental Setup

[0088] (1) Data set and evaluation metrics

[0089] This invention conducted experimental studies on three classic datasets: MNIST, Fashion-MNIST, and CIFAR-10. Furthermore, this invention selected accuracy, precision, recall, F1 score, and AUC as evaluation metrics for the algorithms. In the experiments, this invention performed five-fold cross-validation on each algorithm 10 times (i.e., 50 iterations), calculating the average of the above metrics as the final result of the algorithm.

[0090] (2) Specific implementation

[0091] This invention selects two classic deep convolutional neural networks of different sizes (VGG16 and the lightweight MobileNet) as the backbone models to verify the performance improvement effect of the designed asynchronous learning framework on deep neural networks. Specifically, to more fully verify the enhancement effect of the asynchronous learning framework on complex network models, this invention fine-tunes the structure of MobileNet, replacing the last hidden layer with three fully connected layers (containing 512, 256, and 128 neurons respectively).

[0092] This invention processes VGG16 and MobileNet based on the proposed asynchronous learning framework, and names the processed models VGG16-AL and MobileNet-AL, respectively. The original models VGG16 and MobileNet are used as corresponding benchmark algorithms for comparative analysis.

[0093] (II) Results of the Example:

[0094] (1) Complete results for all methods (i.e., comparison of prediction performance results)

[0095] To verify the performance improvement effect of the proposed asynchronous learning method on deep neural network models, we conducted comparative experiments based on VGG16-AL, MobileNet-AL, VGG16, and MobileNet, obtaining their accuracy, precision, recall, F1 score, and AUC on different datasets. Specific experimental results are shown in Table 1. The results of models incorporating the asynchronous learning framework, as well as the best results among all algorithms, are shown in bold.

[0096] Table 1 Comparison results of all methods on different datasets.

[0097]

[0098]

[0099] In Table 1, the improvement value for each metric is indicated by an upward-pointing black arrow, with the specific numerical value labeled to the right of the arrow. These values ​​represent the improvement values ​​of the deep neural network model on different metrics after processing with the designed asynchronous learning method.

[0100] Analysis of the experimental results in Table 1 reveals that, on the MNIST, Fashion-MNIST, and CIFAR-10 datasets, VGG16-AL and MobileNet-AL, which incorporate asynchronous learning frameworks, significantly outperform VGG16 and MobileNet, respectively, with most metrics differences between the two algorithms exceeding 0.04. MobileNet-AL, in particular, shows improvements of over 0.04 in accuracy, precision, recall, and F1 score across all three datasets compared to MobileNet, and an AUC improvement exceeding 0.013.

[0101] These experimental results demonstrate that the asynchronous learning framework designed in this invention can effectively improve the prediction performance of deep neural network models on different datasets.

[0102] (2) Experimental results under data-limited conditions:

[0103] To verify the impact of the designed asynchronous learning method on the performance of deep neural network models under limited training data, this invention selected VGG16-AL, MobileNet-AL, VGG16, and MobileNet for comparative experiments. Specifically, we selected subsets of different sizes from the MNIST, Fashion-MNIST, and CIFAR-10 datasets as training samples for each model algorithm to obtain its prediction performance on training datasets of different sizes. Due to space limitations, this invention only analyzes the prediction performance of each model algorithm on different datasets using accuracy as an example. Specific results are as follows... Figure 3 As shown.

[0104] from Figure 3As can be seen, in the MNIST, Fashion-MNIST, and CIFAR-10 datasets, the prediction accuracy of all models increases with the increase of dataset size. However, the prediction accuracy of VGG16-AL and MobileNet-AL, after asynchronous learning, improves significantly faster than that of VGG16 and MobileNet without asynchronous learning. Particularly in the MNIST and Fashion-MNIST datasets, when the number of training samples is 300, the prediction accuracy of MobileNet-AL using asynchronous learning exceeds 0.9 and 0.7 respectively, while the prediction accuracy of MobileNet is far below 0.2 and 0.4 respectively, and it only achieves relatively good prediction performance when the number of training samples reaches 700 or more.

[0105] Furthermore, in the MNIST, Fashion-MNIST, and CIFAR-10 datasets, the accuracy of the model algorithms incorporating the asynchronous learning framework consistently outperformed the original model algorithms, regardless of the amount of data. Particularly when training data was limited, the accuracy difference between the models incorporating the asynchronous learning framework and the original models was significant, and this difference gradually decreased as the data size increased. However, once the amount of training data reached a certain level, the accuracy difference between the two types of algorithms became relatively stable with changes in data volume.

[0106] Taking the Fashion-MNIST dataset as an example, when the data volume is 200, the accuracy differences between VGG16-AL and VGG16, and between MobileNet-AL and MobileNet, are 0.125 and 0.545, respectively. However, when the data volume is 700, the accuracy differences are 0.037 and 0.043, respectively. After the data volume reaches 700, the accuracy differences do not fluctuate significantly with the increase in data volume.

[0107] In other words, the asynchronous learning method designed in this invention can effectively reduce the dependence of deep neural network models on massive training data, thereby achieving the desired performance of the model when training data is limited.

[0108] (3) Robustness result verification:

[0109] To verify the robustness improvement effect of the designed asynchronous learning framework on deep neural networks, this invention introduces noise interference into the input data for robustness comparison analysis. Specifically, based on a given ratio parameter, this invention randomly selects some samples from the input data and modifies their labels to achieve noise interference. The ratio parameter values ​​are 0.1, 0.2, 0.3, 0.4, and 0.5. Due to space limitations and considering the actual characteristics of each metric, this section only selects accuracy as a representative metric to measure the performance changes of each algorithm. Detailed results are as follows... Figure 4 As shown.

[0110] analyze Figure 4 The experimental results show that the prediction performance of all algorithms on different datasets decreases with the increase of added noise. When the noise ratio parameter is 0.5, the accuracy of VGG16-AL and MobileNet-AL after asynchronous learning processing on the MNIST, Fashion-MNIST, and CIFAR-10 datasets decreased by approximately 0.202, 0.240, and 0.330 and 0.240, 0.224, and 0.200 respectively compared to the original state, which is significantly better than the performance of VGG16 and MobileNet. Moreover, the accuracy curves of VGG16-AL and MobileNet-AL show a smoother trend.

[0111] In other words, the asynchronous learning method proposed in this invention can significantly improve the robustness of classic deep neural network models, enabling them to better overcome noise interference in the input data.

[0112] (4) Stability results verification:

[0113] Five-fold cross-validation results of VGG16-AL, MobileNet-AL, VGG16, and MobileNet on the MNIST, Fashion-MNIST, and CIFAR-10 datasets were selected to examine the impact of the designed asynchronous learning method on the stability of deep neural network prediction results. Specific experimental results are shown below. Figure 5 As shown.

[0114] analyze Figure 5 The experimental results show that, in different datasets, the average values ​​of various indicators of the deep neural network model incorporating asynchronous learning, as well as the differences between the maximum and minimum values ​​of these indicators, are significantly better than the model algorithm that does not use asynchronous learning.

[0115] Taking CIFAR-10 as an example, the differences between the maximum and minimum values ​​of Accuracy, Precision, Recall, F1, and AUC for VGG16-AL and MobileNet-AL are 0.025, 0.029, 0.025, 0.026, 0.011 and 0.017, 0.023, 0.024, 0.024, 0.011, respectively, which are much smaller than the differences between the maximum and minimum values ​​for VGG16 and MobileNet.

[0116] The experimental results above demonstrate that the asynchronous learning method designed in this invention is of great significance for improving the stability of prediction results of deep neural network models.

[0117] In summary, this invention has conducted multiple comparative experiments on various classic datasets and popular deep neural networks. Experimental results show that the proposed asynchronous learning framework significantly improves the performance of deep neural network models in terms of prediction accuracy, robustness, and stability, and enables deep neural networks to achieve the desired performance even with limited training data.

[0118] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0119] A computer-readable storage medium for storing computer instructions that, when executed by a processor, implement the steps of the above-described method.

[0120] The memory in the embodiments of this application can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDRSDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DRRAM). It should be noted that the memory of the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0121] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means such as coaxial cable, optical fiber, digital subscriber line (DSL), or wireless means such as infrared, wireless, or microwave. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media such as floppy disks, hard disks, magnetic tapes, optical media such as high-density digital video discs (DVDs), or semiconductor media such as solid-state drives (SSDs).

[0122] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0123] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuits in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0124] The foregoing has provided a detailed description of the brain-inspired deep neural network enhancement method for limited data proposed in this invention, and has elucidated the principles and implementation methods of this invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A brain-inspired deep neural network augmentation method for limited data, characterized in that: The method specifically includes the following steps: The method is trained based on three datasets: MNIST, Fashion-MNIST, and CIFAR-10; Step 1: Select the backbone model and divide the region; Choose a deep feedforward neural network (DNN) model with N layers as the backbone model of the asynchronous learning framework; The selected DNN is divided into two parts: the feature extractor FE-net and the predictor IP-net; The feature extractor FE-net includes an input layer and a network layer in front of the DNN, which is responsible for reading input data and extracting features. The predictor IP-net contains the last few layers of the DNN network, which is responsible for integrating the features extracted by FE-net and completing the prediction. Step 2: Reconstruct the model based on the regions defined in Step 1 to obtain the asynchronous architecture of the deep neural network; An auxiliary output layer (AO) is introduced, with the same structure and settings as the output layer of a DNN. The original sequential connection between the feature extractor FE-net and the predictor IP-net is modified so that the feature extractor FE-net is connected in parallel to the predictor IP-net and the auxiliary output layer AO, forming the asynchronous architecture DNN-A of DNN; Step 3: Design an asynchronous training method to train the model reconstructed in Step 2, and independently optimize its feature extractor and predictor respectively; Step 4, Model Prediction; Use the model trained in Step 3 for practical application and output the prediction results; In step 3, the asynchronous training method includes two stages: feature extraction training and state prediction training. The feature extraction training: By combining the auxiliary output layer AO with the feature extractor FE-net, a primary implementation mode of DNN-A, DNN-AJ, is formed. DNN-AJ is trained using training data (X, Y), and the parameters in the feature extractor FE-net are optimized. The state prediction training: The trained feature extractor FE-net and predictor IP-net are combined to generate the advanced implementation mode of DNN-A, DNN-AS. The advanced implementation pattern DNN-AS is trained using training data (X, Y), but only the parameters in IP-net are optimized, while the parameters in FE-net remain unchanged.

2. The method of claim 1, wherein: In step 4, The trained DNN-AS is used to predict the input data and generate prediction results.

3. An enhancement system for a brain-inspired deep neural network enhancement method based on limited data according to any one of claims 1 to 2, characterized in that: The system includes a region partitioning module, a model reconstruction module, and an asynchronous training module. The region division module is used to perform region division after selecting the backbone model; The model reconstruction module reconstructs the model according to the regions divided by the region division module to obtain the asynchronous architecture of the given deep neural network. The asynchronous training module designs an asynchronous training method to train the model reconstructed by the model reconstruction module, independently optimize its feature extractor and predictor, and output prediction results based on the reconstructed model in practical applications.

4. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 2.

5. A computer readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 2.

Images