Semi-supervised machine learning method based on incremental learning and non-training evaluation

Abstract

The invention is suitable for the field of deep learning algorithm optimization and automatic machine learning method design, and particularly relates to a semi-supervised machine learning method based on incremental learning and non-training evaluation. Compared with the prior art, the adaptive extensible unified coding strategy oriented to different deep learning network architectures and hyper-parameters thereof is provided, the problem of unified coding of different deep learning network architectures and hyper-parameters thereof is solved, and adaptive network architecture extension is realized according to different problems and different types of deep learning network architectures. An evolutionary network architecture search algorithm based on simplified reinforcement learning is designed, and the optimization process of the evolutionary network architecture search algorithm is guided by using feedback information in a deep learning network architecture search process. A semi-supervised agent evaluation model based on incremental width learning and non-training evaluation is established, and the learning efficiency of the agent evaluation model is improved by means of a non-training evaluation mode.

Description

Technical Field

[0001] The present invention is applicable to the fields of deep learning algorithm optimization and automatic machine learning method design, and in particular relates to a semi-supervised machine learning method based on incremental learning and training-free evaluation. Background Art

[0002] With the continuous development of big data and artificial intelligence algorithms, as well as the continuous enrichment and improvement of deep learning frameworks such as PyTorch, TensorFlow, and PaddlePaddle, deep learning has achieved great success and has been widely used in various industries, including smart security, finance, autonomous driving, smart manufacturing, and medical research and development. However, facing a large number of big data application scenarios, designing and optimizing deep learning algorithms is a very complex and time-consuming process. Especially in the face of different application scenarios, how to design an automatic machine learning method similar to chatGPT function to realize the automatic learning and optimization of deep learning network architecture and parameters is even more challenging. Considering that different deep learning network architectures have different representation potentials, deep learning network architectures with strong inductive bias can even obtain better modeling effects without training network parameters. It can be seen that studying efficient automatic machine learning methods has important theoretical and application value.

[0003] At present, there are many methods and tools for automatic machine learning. For example, automatic machine learning methods based on reinforcement learning, Bayesian optimization and evolutionary algorithms, as well as open source automatic machine learning libraries such as Auto-Keras and VEGA. However, existing methods still have shortcomings. For example, different deep learning algorithms often have different network architectures, but existing work often studies automatic machine learning methods for the learning and optimization problems of a certain class or several classes of deep learning network architectures. Obviously, for different deep learning network architectures, establishing a more representative and extensive neural architecture search space and designing an adaptive automatic machine learning method are more valuable in theoretical and applied research. In addition, in practical applications, when there is new data, the cost of training the deep learning algorithm from scratch, the cost of evaluating the deep learning network architecture, and the cost of annotating the data are all very high. The research on existing automatic machine learning methods mainly includes the coding design, search optimization and model evaluation of the deep learning network architecture search space. This is also the main technical content of the present invention. In these three aspects, the implementation schemes most similar to the present invention and the existing problems and shortcomings are summarized as follows.

[0004] 1) In terms of coding design of deep learning network architecture search space, a binary coding strategy is designed for the cell-based convolutional neural network search space to achieve unified coding of different convolutional neural network architectures and realize adaptive adjustment according to different problems.

[0005] Although this scheme can achieve adaptive encoding of different convolutional neural network architectures, this encoding strategy is biased towards certain convolutional neural network architectures and cannot achieve unified and fair selection and optimization of all different network architectures, thereby introducing network architecture bias and affecting the performance of deep learning algorithms.

[0006] 2) In terms of search optimization algorithms, reinforcement learning is used to assist the learning of evolutionary algorithms, thereby achieving search optimization of deep learning network architectures. Although this solution can use reinforcement learning to improve the search efficiency of evolutionary algorithms, it will also restrict the search efficiency of deep learning network architectures due to the high computational cost of reinforcement learning itself.

[0007] 3) In terms of model evaluation, a semi-supervised evaluation model is established through a feedforward neural network and used to evaluate different deep learning network architectures. Although this solution can achieve semi-supervised evaluation learning, it cannot achieve incremental learning of the semi-supervised agent evaluation model; instead, it can only use new and old data to re-learn and establish a semi-supervised agent evaluation model at a high learning cost. The high learning cost of this model is also reflected in the fact that the model needs to be trained and learned with a high time cost before it can be used for evaluation after the model learning optimization is achieved. Although the non-training evaluation method can effectively improve the efficiency of the semi-supervised agent evaluation model, this non-training evaluation method is not directly used for the learning and establishment of the semi-supervised agent evaluation model.

[0008] Therefore, it is of great theoretical and practical value to design efficient semi-supervised automatic machine learning methods using a small amount of labeled data and to realize automatic machine learning with incremental learning capabilities when new data is available. Summary of the invention

[0009] The present invention provides a semi-supervised machine learning method based on incremental learning and training-free evaluation, aiming to solve the problem that existing automatic machine learning methods can only learn and optimize one or several types of deep learning network architectures, as well as the problem of high cost of training and evaluating deep learning algorithms.

[0010] The method comprises the following steps:

[0011] S1. Obtain multiple deep learning network architectures and corresponding hyperparameters for selection;

[0012] S2. Encoding the multiple deep learning network architectures and the corresponding optional ranges of hyperparameters through a preset encoding strategy to obtain a neural architecture search space;

[0013] S3, obtaining an evolutionary network architecture search algorithm, and optimizing the evolutionary network search algorithm according to a preset method to obtain an optimized search algorithm;

[0014] S4, searching the neural architecture search space by using the optimization search algorithm to obtain multiple network architectures;

[0015] S5, establishing a semi-supervised agent evaluation model, and evaluating the plurality of network architectures by using the semi-supervised agent evaluation model to obtain evaluation results;

[0016] S6. Iteratively update the multiple network architectures according to the evaluation results and return to step S4 until the preset conditions are met to obtain the optimal deep learning network.

[0017] Preferably, the preset encoding strategy is:

[0018] Chaining the binary codes of each neural unit in the deep learning network architecture; wherein the neural unit includes a plurality of nodes, and the nodes represent feature information;

[0019] Connecting different nodes in a multi-branch manner based on a hierarchical structure to form neural units with different network topological connections;

[0020] The nodes are numbered in sequence, and only the node with a smaller sequence number can be connected to the node with a larger sequence number.

[0021] Preferably, the binary code of the neural unit includes a first part, a second part and a third part; the first part is used to indicate whether the neural unit exists; the second part is used to indicate the selectable deep learning network architecture and corresponding hyperparameters; and the third part is used to indicate the connection relationship with other nodes.

[0022] Preferably, the preset method is:

[0023] The search operators in the evolutionary network architecture search algorithm are classified according to their emphasis into convergence operators and diversity operators; wherein the evolutionary network architecture search algorithm includes at least one of the convergence operators and one of the diversity operators;

[0024] According to a preset calculation formula and the search result of the search operator, a positive reward or a reverse penalty is given to the search operator; the positive reward is used to increase the execution probability, and the reverse penalty is used to reduce the execution probability.

[0025] Preferably, the population number of the evolutionary network architecture search algorithm is defined as N p , Represents individual i(i∈[1,N p ]) At time t, the state of different search operators is selected; then the preset calculation formula is as follows:

[0026]

[0027] p i (t) = p i (t-1)+Δp i (t)

[0028]

[0029] Among them, the parameter w i The learning algorithm of (t) is as follows:

[0030] w i (t) = w i (t-1)+Δw i (t)

[0031]

[0032] in, Indicates that individual i takes The probability of i (t) represents the probability value of individual i at time t; Δp i (t) represents the probability change of individual i at time t compared to time t-1; η represents the algorithm hyperparameter greater than 0; b i represents the initial value of the feedback reward obtained by individual i; w i (t) represents the parameters to be learned by individual i at time t; Represents feedback reward.

[0033] Preferably, in step S5, establishing the semi-supervised agent evaluation model includes the following sub-steps:

[0034] S51, obtaining a plurality of data samples, randomly inputting the plurality of data samples into different network architectures, performing calculations according to the order of the number of layers in the input network architectures and the total number of activation functions in the corresponding network architectures, and calculating to obtain a plurality of binary strings;

[0035] S52, calculating the distance between different binary strings according to a distance measurement formula to obtain a comparison matrix, wherein the comparison matrix is ​​used to measure the similarity of the data samples after being processed by the network architecture;

[0036] S53, calculating the performance of each of the network architectures to obtain a performance result corresponding to each of the network architectures;

[0037] S54, randomly selecting a part of the network architectures from the plurality of network architectures for training and calculating the performance of the trained network architectures; correcting the performance results of the network architectures before training by the performance results of the trained network architectures, and taking all the network architectures as a network architecture set;

[0038] S55. Learning the data pairs consisting of the network architecture set and the corresponding performance results through a width learning algorithm to obtain the semi-supervised agent evaluation model.

[0039] Preferably, the performance calculation formula of the network architecture is as follows:

[0040] Y=log|A|

[0041] Wherein, Y represents the performance of the network architecture, and A represents the comparison matrix.

[0042] Preferably, the semi-supervised agent evaluation model performs incremental learning through an incremental width learning algorithm and the newly added data.

[0043] Compared with the prior art, the present invention proposes an adaptive and scalable unified coding strategy for different deep learning network architectures and their hyperparameters, solves the problem of unified coding of different deep learning network architectures and their hyperparameters, and realizes adaptive network architecture expansion according to different problems and different types of deep learning network architectures. An evolutionary network architecture search algorithm based on simplified reinforcement learning is designed, and the optimization process of the evolutionary network architecture search algorithm is guided by utilizing the feedback information in the deep learning network architecture search process, thereby improving the optimization efficiency and accuracy of the evolutionary network architecture search algorithm and solving the search optimization problem of the deep learning network architecture. A semi-supervised agent evaluation model based on incremental width learning and non-training evaluation is established. Through the width learning algorithm, a semi-supervised agent evaluation model of the deep learning network architecture and its hyperparameters is established, and the learning efficiency of the agent evaluation model is improved by means of the non-training evaluation method. At the same time, through the incremental width learning algorithm, when the new network architecture and its performance data are added, the incremental update of the semi-supervised agent evaluation model is realized, reducing the high incremental learning cost of model training from scratch. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention will be described in detail below in conjunction with the accompanying drawings. The above and other aspects of the present invention will become clearer and easier to understand through the detailed description made in conjunction with the following drawings. In the accompanying drawings:

[0045] Figure 1 is a flow chart of a semi-supervised machine learning method based on incremental learning and non-training evaluation provided by an embodiment of the present invention;

[0046] Figure 2It is a structural schematic diagram of a semi-supervised machine learning method based on incremental learning and no training evaluation provided by an embodiment of the present invention;

[0047] Figure 3 It is a schematic diagram of a preset encoding strategy for a semi-supervised machine learning method based on incremental learning and training-free evaluation provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] Please refer to Figure 1-Figure 3 The present invention provides a semi-supervised machine learning method based on incremental learning and training-free evaluation, comprising the following steps:

[0050] S1. Obtain multiple deep learning network architectures and corresponding hyperparameters for selection;

[0051] S2. Encoding the multiple deep learning network architectures and the corresponding optional ranges of hyperparameters through a preset encoding strategy to obtain a neural architecture search space;

[0052] In the embodiment of the present invention, the preset encoding strategy is:

[0053] Chaining the binary codes of each neural unit in the deep learning network architecture; wherein the neural unit includes a plurality of nodes, and the nodes represent feature information;

[0054] Connecting different nodes in a multi-branch manner based on a hierarchical structure to form neural units with different network topological connections;

[0055] The nodes are numbered in sequence, and only the node with a smaller sequence number can be connected to the node with a larger sequence number.

[0056] In an embodiment of the present invention, the binary code of the neural unit includes a first part, a second part and a third part; the first part is used to indicate whether the neural unit exists; the second part is used to indicate the selectable deep learning network architecture and corresponding hyperparameters; the third part is used to indicate the connection relationship with other nodes.

[0057] In an embodiment of the present invention, the present invention proposes to connect different cells (i.e., neural units composed of different nodes) in a chain manner to form various deep learning network architectures. Specifically, each node in each cell represents feature information. These different nodes are connected in a multi-branch manner based on a hierarchical structure to form cells with different network topological connections. Here, the connection between the nodes represents a certain operation or set of operations. For example, convolution operations, normalization operations, pooling operations, attention learning, linear or nonlinear operations, etc. All these cells are connected in a chain manner to form a complete deep learning network architecture. Specifically, in order to realize the encoding of this network architecture, the present invention proposes the following preset encoding strategy.

[0058] The encoding of the entire deep learning network architecture and its hyperparameters is composed of the binary codes of each cell in a chain connection order. Without loss of generality, assume that the entire network architecture consists of at most m (m ≥ 1) cells, where the i-th (i∈[1, m]) cell has n i (n i ≥1) nodes. Here m and n i All need to be set in advance. Figure 2 As shown in the figure, the encoding of each cell consists of three parts. The first part consists of a binary digit, indicating whether the current cell exists. 0 means that the current cell does not exist, and 1 means that the current cell exists. The second part consists of a string of s j Bit binary, total Bit binary. j Bit binary coded representation The third part consists of 0.5n i (n i -1) bits of binary represent the interconnectedness of each feature node. Since the deep learning network architecture can be viewed as a directed acyclic graph, each node is numbered sequentially, and only nodes with smaller numbers can be connected to nodes with larger numbers. Starting from the node with the smallest number, it is connected to each node with larger numbers in turn. If there is a connection between nodes, the corresponding binary bit is 1, otherwise it is 0.

[0059] Since not every cell exists, in order to ensure the legitimacy of the entire deep learning network architecture, additional operations may need to be added between each cell to keep the input and output of the previous and next cells consistent. For example, the output of the previous cell must be consistent with the input dimension of the next cell.

[0060] Existing deep learning network architecture encoding methods based on cells and hierarchical structures can be integrated into the encoding scheme of the present invention. The entire deep learning network architecture and its hyperparameters are given by Whether each cell and its internal nodes exist, whether there is a connection relationship between nodes, and which network architecture and its hyperparameters are selected are all adaptively adjusted according to the performance of the deep learning network on specific problems, thereby realizing adaptive and scalable deep learning network architecture encoding.

[0061] Exemplarily, in practical applications, the coding strategy proposed in the present invention is applicable to different deep learning network architectures such as Convolutional Neural Networks (CNN), Long Short-term Memory Networks (LSTM), Recurrent Neural Networks (RNN), Deep Belief Networks (DBN) and Transformer. Here, the coding strategy is explained by taking Convolutional Neural Network CNN as an example.

[0062] Since not every cell exists, in order to ensure the legitimacy of the entire neural architecture, the present invention stipulates that a convolutional layer is placed in front of each cell, followed by a convolutional layer and a pooling layer. If the i-th cell does not exist, the convolutional layer in front of it and the pooling layer behind it are directly connected. In this way, it can be ensured that the output dimension of the previous cell is consistent with the input dimension of the next cell. The codes of all cells are arranged and spliced ​​together in the order of the cells, which constitutes the code of the entire neural architecture.

[0063] Please refer to Figure 3 , Figure 3 This is a specific implementation example. This example contains at most 3 cells, each of which consists of 3, 4 and 5 nodes respectively. For this specific deep learning network architecture, the 3rd cell does not exist because the first bit of the 3rd cell is encoded as "0". That is, only the 1st and 2nd cells constitute the final network architecture.

[0064] Taking the first cell as an example, the second part is encoded as "0110". The first two bits represent the encoding of the output channel number parameter, and the last two bits represent the encoding of the convolution kernel size parameter. There are 4 options for the two binary digits, representing 4 options for each of the two parameters. In fact, other hyperparameters and multiple binary bits (more than 2 binary bits) that can represent more options can also be selected. This is just an example, and does not mean that the present invention is only applicable to the encoding of these two hyperparameters and 4 options. Here, "01" represents the second option, that is, the number of channels is selected as 128; "10" represents the third option, that is, the convolution kernel size is selected as 5*5. The last part of the code represents the connection relationship between the nodes in the cell. Since only nodes with small serial numbers can be connected to nodes with large serial numbers, and starting from the node with the smallest serial number, they are connected to the nodes with large serial numbers in turn; therefore, the first bit is encoded as "1" to indicate that node 1 is connected to node 2; the second bit is encoded as "0" to indicate that node 1 is not connected to node 3. The third bit is encoded as "1" to indicate that node 2 is connected to node 3.

[0065] By analogy, all deep learning network architectures with different connection relationships and hyperparameters with different numbers of channels and convolution kernel sizes can be encoded. These ultimately constitute a neural architecture search space composed of different deep learning network architecture search spaces.

[0066] S3, obtaining an evolutionary network architecture search algorithm, and optimizing the evolutionary network search algorithm according to a preset method to obtain an optimized search algorithm;

[0067] In the embodiment of the present invention, the optimization search algorithm obtained is as follows Figure 2 The evolutionary network architecture search algorithm based on simplified reinforcement learning shown in.

[0068] In the embodiment of the present invention, the preset method is:

[0069] The search operators in the evolutionary network architecture search algorithm are classified according to their emphasis into convergence operators and diversity operators; wherein the evolutionary network architecture search algorithm includes at least one of the convergence operators and one of the diversity operators;

[0070] According to a preset calculation formula and the search result of the search operator, a positive reward or a reverse penalty is given to the search operator; the positive reward is used to increase the execution probability, and the reverse penalty is used to reduce the execution probability.

[0071] In the embodiment of the present invention, the population number of the evolutionary network architecture search algorithm is defined as N p , Represents individual i(i∈[1,N p ]) At time t, the state of different search operators is selected; then the preset calculation formula is as follows:

[0072]

[0073] p i (t) = p i (t-1)+Δp i (t)

[0074]

[0075] Parameter w i The learning algorithm of (t) is as follows:

[0076] w i (t) = w i (t-1)+Δw i (t)

[0077]

[0078] in, Indicates that individual i takes The probability of i (t) represents the probability value of individual i at the tth moment; Δp i (t) represents the probability change of individual i at time t compared to time t-1; η represents the algorithm hyperparameter greater than 0; b i represents the initial value of the feedback reward obtained by individual i; w i (t) represents the parameters to be learned by individual i at time t; Represents feedback reward.

[0079] Specifically, in an evolutionary algorithm with multiple search operators (i.e., an evolutionary network architecture search algorithm), different search operators often have different search performances. In theory, operators with better search performance should be executed first. If different search operators are executed at the same time without selection, or are executed with the same probability, it will inevitably affect the optimization efficiency of the algorithm. In order to make full use of the feedback information in the deep learning network architecture search process, the present invention proposes an evolutionary network architecture search algorithm based on simplified reinforcement learning, namely, an optimization search algorithm. By utilizing feedback information, learning to adaptively adjust the execution probability of different operators ensures that operators with better search performance are more likely to be selected for execution, thereby improving the optimization efficiency and accuracy of the evolutionary network architecture search algorithm. The principles of the algorithm proposed in the present invention are summarized as follows.

[0080] Evolutionary algorithms with at least two search operators (such as genetic algorithms, differential evolution algorithms, chicken swarm optimization algorithms, and bird swarm optimization algorithms, etc.) are divided into two categories (convergent operators and diverse operators) according to the different convergence and diversity of their operators. According to the optimization effects of these two types of operators, learn the frequency of their use. For operators with good optimization effects, the positive reward r is given, and the probability g of being selected for execution will also be greater; for operators with poor optimization effects, the reverse penalty r is given, and the probability g of being selected for execution will also be smaller. By controlling the execution of the diversity search operator and the convergence search operator, the balance between the diversity and convergence of the evolutionary algorithm is achieved, thereby improving the optimization efficiency of the algorithm.

[0081] In order to learn the selection and execution probability of convergence and diversity operators and improve the optimization efficiency of evolutionary algorithms, the present invention proposes a feedback learning strategy based on simplified reinforcement learning. Different from the existing reinforcement learning that uses complex algorithms to learn execution-evaluation (actor-critic) operations, the present invention uses a certain function g=f(r, y) to describe the correlation between the size of the reward value, the selection state y of different operators (i.e., whether a certain operator is selected for execution or not) and the probability g of different operators being selected for execution. As long as it is ensured that the reward of feedback is proportional to the probability of the corresponding operator being selected for execution g∝r, the above-mentioned feedback learning based on the search process can be realized. That is, the greater the positive reward, the greater the probability of the corresponding operator being selected for execution; on the contrary, the greater the reverse reward, the smaller the probability of the corresponding operator being selected for execution. Since only a simple function is added and no complex learning calculation process is involved, this simplified reinforcement learning algorithm does not add additional computational overhead to the evolutionary algorithm, but can make full use of the feedback information in the deep learning network architecture search process to guide the optimization process of the evolutionary algorithm, and ultimately improve the optimization efficiency and accuracy of the evolutionary network architecture search algorithm.

[0082] For the evolutionary network architecture search algorithm based on simplified reinforcement learning proposed in the present invention, the key to algorithm design is to design a function that can effectively describe the correlation between the size of the reward value, the selection state of different operators, and the probability of different operators being selected for execution. The present invention proposes an implementation example, which is as follows.

[0083] Assume that the population size of the entire evolutionary algorithm is Represents individual i(i∈[1,N p ]) At the tth moment (or the number of iterations), the states of different search operators (search operators focusing on convergence and search operators focusing on diversity, i.e., convergence operators and diversity operators) are selected. When , individual i chooses a search operator that focuses on convergence; when When , individual i chooses a search operator that focuses on diversity. The quality of fitness obtained by individuals using different search operators is used as feedback reward The basis for setting. It refers to the individual i taking The probability of and p i The function of (t) is specifically expressed as follows:

[0084]

[0085] p i (t) = p i (t-1)+Δp i (t)

[0086]

[0087] Parameter w i The learning algorithm of (t) is as follows:

[0088] w i (t) = w i (t-1)+Δw i (t)

[0089]

[0090] Where η is an algorithm hyperparameter greater than 0, b i is the initial value of the feedback reward for individual i, which is a hyperparameter. For a certain problem (certain input) and w i For (t), b i and Independent of each other. The size depends on how good the fitness value obtained by the individual using different search operators is compared to its historical value. If it is better than the historical optimal value, then If the change is bad, set it to a positive value; conversely, if it is bad, set it to a negative value.

[0091] The above function meets the requirement that the probability of an operator being selected for execution is proportional to its feedback reward value. That is, the greater the feedback reward, the greater the probability of the corresponding operator being selected, which can realize feedback learning based on search process information, thereby improving the optimization efficiency and accuracy of the evolutionary network architecture search algorithm. Since only a few simple steps are added, it will not increase the computational overhead of the evolutionary algorithm. The entire algorithm flow is as follows:

[0092] S4, searching the neural architecture search space by using the optimization search algorithm to obtain multiple network architectures;

[0093] S5, establishing a semi-supervised agent evaluation model, and evaluating the plurality of network architectures by using the semi-supervised agent evaluation model to obtain evaluation results;

[0094] In an embodiment of the present invention, establishing the semi-supervised agent evaluation model includes the following sub-steps:

[0095] S51, obtaining a plurality of data samples, randomly inputting the plurality of data samples into different network architectures, performing calculations according to the order of the number of layers in the input network architectures and the total number of activation functions in the corresponding network architectures, and calculating to obtain a plurality of binary strings;

[0096] S52, calculating the distance between different binary strings according to a distance measurement formula to obtain a comparison matrix, wherein the comparison matrix is ​​used to measure the similarity of the data samples after being processed by the network architecture;

[0097] S53, calculating the performance of each of the network architectures to obtain a performance result corresponding to each of the network architectures;

[0098] S54, randomly selecting a part of the network architectures from the plurality of network architectures for training and calculating the performance of the trained network architectures; correcting the performance results of the network architectures before training by the performance results of the trained network architectures, and taking all the network architectures as a network architecture set;

[0099] S55. Learning the data pairs consisting of the network architecture set and the corresponding performance results through a width learning algorithm to obtain the semi-supervised agent evaluation model.

[0100] In the embodiment of the present invention, the calculation formula of the performance of the network architecture is as follows:

[0101] Y=log|A|

[0102] Wherein, Y represents the performance of the network architecture, and A represents the comparison matrix.

[0103] In the embodiment of the present invention, the semi-supervised agent evaluation model performs incremental learning through an incremental width learning algorithm and the newly added data.

[0104] It should be noted that after encoding the deep learning network architecture and its hyperparameters to obtain a search space containing different network architectures, and designing an optimized search algorithm, the key to achieving automatic machine learning is to evaluate the performance of different network architectures in the search space and select the optimal network architecture from it. After selecting the optimal network architecture, overall training and learning are performed to obtain the optimal deep learning network and achieve automatic machine learning. Considering that the deep learning network itself has a huge learning time cost, the time cost of searching and evaluating the optimal deep learning network suitable for a certain problem or a certain type of problem from different deep learning network architecture search spaces will be even greater. To solve this problem, the present invention proposes solutions from three aspects.

[0105] First, from the perspective of the evaluation of deep learning network architecture, an evaluation strategy based on no training is proposed. The existing deep learning network uses continuous training to learn to obtain an optimized network, and then tests the network to obtain the final network performance. Although this method can ensure the accuracy of the network evaluation results, it will undoubtedly increase the time cost. However, under the conditions of given input and untrained network, although only the initial performance of different deep learning network architectures can be obtained, based on these preliminary evaluation results, it is sufficient to obtain the ranking of the performance of different network architectures. Considering that the focus of deep learning network architecture search is to find the optimal network architecture. Therefore, this non-training evaluation method is sufficient to meet the task of network architecture search to find the optimal network architecture. The principle of the non-training evaluation strategy proposed in the present invention is summarized as follows.

[0106] The training-free evaluation strategy proposed in the present invention is derived from two observations on the principles of existing deep learning algorithms. On the one hand, the existing deep learning algorithms mainly learn layer by layer by first linear mapping and then processing the activation function; on the other hand, the activation function of the existing deep learning algorithm is mainly ReLU and its variants, and the essence of activation functions such as ReLU and its variants is to divide the input of the linear mapping into the positive region and the negative region of the input (or directly 0). Combining these two observations, it can be seen that for the deep learning network architecture with ReLU and its variants as the activation function, the input is divided into positive and negative regions in turn after layer-by-layer learning of the network. The task of the final network is equivalent to establishing the correlation between these positive and negative regions and the expected output of the network. In theory, different inputs should be divided into different positive and negative regions after layer-by-layer learning of the network; otherwise, these different inputs cannot be effectively distinguished by the network. It can be seen that for different inputs, the combination of positive and negative regions divided by layer-by-layer learning of the network can be used to judge the performance of the deep learning network. If the combinations of positive and negative regions corresponding to different inputs are more different, the corresponding deep learning network will have better performance; on the contrary, if the combinations of positive and negative regions corresponding to different inputs are more similar, the corresponding deep learning network will have worse performance. The main principles and steps are as follows:

[0107] 1) Given different input data (i.e., data samples, the number of samples is denoted as N) and a deep learning network obtained by random encoding based on the encoding strategy proposed in the present invention.

[0108] 2) For each input data, the positive and negative areas of each layer of input are calculated in sequence according to the order of the number of input layers of the deep learning network, and marked as 1 and 0 respectively. In this way, a binary string S with a total length of L can be obtained. Here, L represents the total number of activation functions in the entire deep learning network.

[0109] 3) By comparing the binary strings S corresponding to the N different inputs one by one, an N*N matrix is ​​obtained. This matrix reflects the modeling performance of the corresponding deep learning network for the N different input data.

[0110] 4) For another deep learning network, an N*N matrix is ​​also calculated. By comparing the differences between the two matrices, the network that calculates the N*N matrix that is more different has better performance. In this way, the performance of different deep learning networks can be evaluated.

[0111] Secondly, from the perspective of semi-supervised learning under conditions of insufficient data labels, a semi-supervised agent evaluation model based on width learning is proposed. The time cost of learning and evaluating deep learning network architectures is very high. Although the untrained evaluation strategy can effectively alleviate the time cost of learning and evaluating the network architecture, this method is not accurate enough after all, and the generalization performance may not be good enough. By learning the correlation between the two from the data pairs of real deep learning network architectures and their performance, and establishing a semi-supervised agent evaluation model, the above problems can be effectively compensated. In particular, the width learning algorithm has the ability to learn quickly. Establishing a semi-supervised agent evaluation model based on width learning is more conducive to quickly learning the correlation between the network architecture and its performance, and using it to predict the performance of unknown network architectures, thereby realizing semi-supervised learning. The main principle steps of the semi-supervised agent evaluation model proposed in the present invention are as follows:

[0112] 1) Based on the encoding strategy proposed in the present invention, a certain number of deep learning network architectures and their hyperparameters are obtained by random encoding.

[0113] 2) According to the non-training evaluation strategy proposed in the present invention, the performance of the deep learning networks corresponding to these network architectures and their hyperparameters is evaluated. At the same time, a portion of the networks is selected from these deep learning networks according to a certain ratio, and training and learning are first performed according to the existing method. After the network completes the learning optimization, the performance of this portion of the network is evaluated. The network performance results obtained after training are used to correct the errors in the network performance evaluation without training.

[0114] 3) Using the deep learning network architecture and its hyperparameters obtained above and the corresponding network performance data pairs, a semi-supervised agent evaluation model based on width learning is trained.

[0115] 4) Use the learned semi-supervised proxy evaluation model to predict the performance of unknown deep learning network architectures and achieve semi-supervised learning under conditions of insufficient data labels.

[0116] Finally, from the perspective of using new data to achieve incremental learning of a semi-supervised proxy evaluation model, a semi-supervised proxy evaluation model based on incremental width learning is proposed. When there is new data, the existing learning method is to merge the new and old data together and retrain the proxy evaluation model. This method completely discards the previously learned model, and the time cost of repeated training and learning is definitely higher. The model learning method proposed in the present invention can achieve incremental updates of the proxy model only through incremental learning of new data on the basis of making full use of the existing model, thereby greatly reducing the learning time cost of the model. The main principle steps of the incremental learning of the semi-supervised proxy evaluation model proposed in the present invention are as follows.

[0117] 1) Given the learned semi-supervised proxy evaluation model and its corresponding parameters, preprocess (e.g., normalize) the newly added deep learning network architecture and its hyperparameters and their corresponding network performance data pairs.

[0118] 2) Using an incremental width learning algorithm for input data (which can also be combined with an incremental width learning algorithm for newly added mapping feature nodes and / or newly added enhancement nodes), learn the newly added data pairs and update the parameters of the semi-supervised agent evaluation model to obtain an updated model and its parameters.

[0119] 3) The new network architecture is searched by the evolutionary network architecture search algorithm based on simplified reinforcement learning, and the semi-supervised agent evaluation model obtained by incremental learning is used to evaluate the performance of this new unknown deep learning network architecture.

[0120] 4) Based on the comparison between the new deep learning network architecture performance and the historical network performance, and combined with the iterative update formula of the specific evolutionary algorithm, the entire population in the evolutionary network architecture search algorithm is updated to achieve continuous iterative optimization. Finally, the optimal deep learning network architecture is obtained when the established stopping conditions are met (such as the maximum number of iterations, the maximum number of function evaluations, or the maximum floating point number, etc.).

[0121] 5) When there are new data pairs, repeat the above step 1); when a new unknown deep learning network architecture and its hyperparameters are generated, repeat the above step 3).

[0122] 6) After obtaining the optimal deep learning network architecture, the optimal deep learning network can be obtained according to the overall training and learning method to realize automatic machine learning.

[0123] Exemplarily, given the deep learning network architecture and its hyperparameters X obtained by encoding based on the encoding strategy of the present invention, the newly added deep learning network architecture and its hyperparameters X a , and N sample data for a specific problem. Considering X and X aThe existence of discrete values ​​in the data is not convenient for learning. The present invention does not directly use these discrete values, but the continuous values ​​converted after normalization. A specific implementation example of the semi-supervised agent evaluation model proposed in the present invention is as follows:

[0124] 1) For each deep learning network composed of a deep learning network architecture and its hyperparameters X, calculate the binary string S obtained when all N sample data are input into the deep learning network, and then use the Hamming distance (other distance measurement formulas can also be used) to describe the difference d (S) between these N different binary strings S when they are compared with each other. i ,S j )(ij∈[1,2,…,N]). For a given deep learning network, the N*N matrix A formed by comparing all binary strings S pairwise can be used to measure the similarity of these different input data. The lower the similarity, the closer A is to the diagonal.

[0125]

[0126] 2) Use the following formula to calculate the performance Y of each deep learning network without training. The larger the Y value, the more effectively the corresponding deep learning network can process different input data, and the better its performance.

[0127] Y=log|A|

[0128] 3) For some deep learning networks, they are first trained and learned according to the existing methods. When the model training reaches the predetermined stop condition (such as the number of iterations), they are used for actual data prediction and learning. The deep learning network performance obtained based on this evaluation method is used to correct the performance of the network under the untrained condition. When the performance of two networks is judged wrongly under the untrained condition, the Y values ​​of the two networks are exchanged according to the performance results of the two networks during training.

[0129] 4) Using the wide learning algorithm, learn the data pairs (X, Y) consisting of the deep learning network and its performance, and obtain a preliminary semi-supervised agent evaluation model

[0130] 5) For the newly added deep learning network architecture and its hyperparameters X a , using steps 1) and 2) to calculate X a The performance of the deep learning network Y a , and form a new data pair (X a , Y a ).

[0131] 6) Adopt the incremental width learning algorithm based on the newly added input data to learn the newly added data pairs (X a , Y a ), update the semi-supervised agent evaluation model This enables incremental learning of the semi-supervised agent evaluation model.

[0132] S6. Iteratively update the multiple network architectures according to the evaluation results and return to step S4 until the preset conditions are met to obtain the optimal deep learning network.

[0133] Compared with the prior art, the present invention proposes an adaptive and scalable unified coding strategy for different deep learning network architectures and their hyperparameters, solves the problem of unified coding of different deep learning network architectures and their hyperparameters, and realizes adaptive network architecture expansion according to different problems and different types of deep learning network architectures. An evolutionary network architecture search algorithm based on simplified reinforcement learning is designed, and the optimization process of the evolutionary network architecture search algorithm is guided by utilizing the feedback information in the deep learning network architecture search process, thereby improving the optimization efficiency and accuracy of the evolutionary network architecture search algorithm and solving the search optimization problem of the deep learning network architecture. A semi-supervised agent evaluation model based on incremental width learning and non-training evaluation is established. Through the width learning algorithm, a semi-supervised agent evaluation model of the deep learning network architecture and its hyperparameters is established, and the learning efficiency of the agent evaluation model is improved by means of the non-training evaluation method. At the same time, through the incremental width learning algorithm, when the new network architecture and its performance data are added, the incremental update of the semi-supervised agent evaluation model is realized, reducing the high incremental learning cost of model training from scratch.

[0134] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article 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, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0135] The embodiments of the present invention are described above in conjunction with the accompanying drawings. What is disclosed is only the preferred embodiment of the present invention. However, the present invention is not limited to the above-mentioned specific implementation manner. The above-mentioned specific implementation manner is only illustrative rather than restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms and equivalent changes without departing from the scope of protection of the purpose of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A semi-supervised machine learning method based on incremental learning and training-free evaluation, characterized in that: The following steps are involved: S1. Obtain multiple deep learning network architectures and corresponding hyperparameters for selection; S2. Encoding the multiple deep learning network architectures and the corresponding optional ranges of hyperparameters through a preset encoding strategy to obtain a neural architecture search space; S3, obtaining an evolutionary network architecture search algorithm, and optimizing the evolutionary network search algorithm according to a preset method to obtain an optimized search algorithm; S4, searching the neural architecture search space by using the optimization search algorithm to obtain multiple network architectures; S5, establishing a semi-supervised agent evaluation model, wherein the semi-supervised agent evaluation model is used to evaluate the performance of the network architecture, and evaluating multiple network architectures through the semi-supervised agent evaluation model to obtain evaluation results; S6. Determine whether the preset conditions are met. If not, iteratively update the multiple network architectures according to the evaluation results and return to step S4; if so, output the current network architecture as the optimal deep learning network.

2. The semi-supervised machine learning method based on incremental learning and training-free evaluation as claimed in claim 1, characterized in that: The preset encoding strategy is: Chaining the binary codes of each neural unit in the deep learning network architecture; wherein the neural unit includes a plurality of nodes, and the nodes represent feature information; Connecting different nodes in a multi-branch manner based on a hierarchical structure to form neural units with different network topological connections; The nodes are numbered in sequence, and only the node with a smaller sequence number can be connected to the node with a larger sequence number.

3. The semi-supervised machine learning method based on incremental learning and training-free evaluation as claimed in claim 2, characterized in that: The binary code of the neural unit includes a first part, a second part and a third part; the first part is used to indicate whether the neural unit exists; the second part is used to indicate the selectable deep learning network architecture and corresponding hyperparameters; the third part is used to indicate the connection relationship with other nodes.

4. The semi-supervised machine learning method based on incremental learning and training-free evaluation as claimed in claim 1, characterized in that: The preset method is: The search operators in the evolutionary network architecture search algorithm are classified according to their emphasis into convergence operators and diversity operators; wherein the evolutionary network architecture search algorithm includes at least one of the convergence operators and one of the diversity operators; According to a preset calculation formula and the search result of the search operator, a positive reward or a reverse penalty is given to the search operator; the positive reward is used to increase the execution probability, and the reverse penalty is used to reduce the execution probability.

5. The semi-supervised machine learning method based on incremental learning and training-free evaluation as claimed in claim 4, characterized in that: Define the population number of the evolutionary network architecture search algorithm as N p , Represents individual i(i∈[1,N p ]) At time t, the state of different search operators is selected; then the preset calculation formula is as follows: p i (t)=p i (t-1)+Δp i (t) Among them, the parameter w i The learning algorithm of (t) is as follows: w i (t)=w i (t-1)+Δw i (t) in, Indicates that individual i takes The probability of i (t) represents the probability value of individual i at the tth moment; Δp i (t) represents the probability change of individual i at time t compared to time t-1; η represents the algorithm hyperparameter greater than 0; b i represents the initial value of the feedback reward obtained by individual i; w i (t) represents the parameters to be learned by individual i at time t; Represents feedback reward.

6. The semi-supervised machine learning method based on incremental learning and training-free evaluation as claimed in claim 1, characterized in that: In step S5, establishing the semi-supervised agent evaluation model includes the following sub-steps: S51, obtaining a plurality of data samples, inputting the plurality of data samples into different network architectures, performing calculations according to the order of the number of layers in the input network architectures and the total number of activation functions in the corresponding network architectures, and calculating to obtain a plurality of binary strings; S52, calculating the distance between different binary strings according to a distance measurement formula to obtain a comparison matrix, wherein the comparison matrix is ​​used to measure the similarity of the data samples after being processed by the network architecture; S53, calculating the performance of each of the network architectures to obtain a performance result corresponding to each of the network architectures; S54, randomly selecting a part of the network architectures from the plurality of network architectures for training and calculating the performance of the trained network architectures; correcting the performance results of the network architectures before training by the performance results of the trained network architectures, and taking all the network architectures as a network architecture set; S55. Learning the data pairs consisting of the network architecture set and the corresponding performance results through a width learning algorithm to obtain the semi-supervised agent evaluation model.

7. The semi-supervised machine learning method based on incremental learning and no training evaluation as claimed in claim 6, characterized in that: The performance of the network architecture is calculated as follows: Y=log|A| Wherein, Y represents the performance of the network architecture, and A represents the comparison matrix.

8. The semi-supervised machine learning method based on incremental learning and no training evaluation as claimed in claim 6, characterized in that: The semi-supervised agent evaluation model performs incremental learning through an incremental width learning algorithm and the newly added data pairs.

Images