Classifier-assisted neural network architecture search multi-target multi-level learning optimization method

Abstract

The invention discloses a classifier-assisted neural network architecture search multi-target multi-level learning optimization method, which comprises the following steps of: modeling by taking a neural network architecture search problem as a multi-target optimization problem, and constructing a search space; constructing a test network for training; screening a Pareto optimal solution through non-dominated sorting, and constructing a special search space based on a Pareto leading edge reserved operator; performing a classifier-assisted hierarchical learning-based multi-objective evolutionary optimization algorithm, selecting an updated population after evolution based on a candidate solution screening strategy, performing objective function evaluation based on a screened solution, updating an archive library, outputting a Pareto optimal solution in the archive library after iterative training, and performing multi-objective evolutionary optimization on the basis of the Pareto optimal solution in the archive library. And selecting a network in the optimal Pareto frontier of the last generation of population as an optimal neural network architecture. According to the method, efficient search of the neural network architecture is realized through the classifier-assisted hierarchical particle swarm optimization algorithm, and a solution meeting the double targets of classification accuracy and reasoning delay is obtained.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, specifically to a multi-objective, multi-level learning optimization method for classifier-assisted neural network architecture search. Background Technology

[0002] The success of deep learning heavily relies on innovations in neural network architectures, such as breakthroughs in computer vision using convolutional neural networks (CNNs) (e.g., ResNet, Inception) and applications of recurrent neural networks (RNNs) in natural language processing (e.g., LSTM). However, the design of these architectures primarily depends on expert experience and trial-and-error, and thus has the following significant limitations:

[0003] First, the high cost of human resources means that designing a high-performance architecture typically requires months or even years of iterative optimization. For example, the residual connection design of ResNet underwent dozens of experimental verifications. Second, there is the difficulty of adapting to different scenarios. The architectural requirements for different tasks (such as image classification, object detection, and speech recognition) vary greatly, and manual design is difficult to quickly adapt to diverse scenarios. At the same time, classic architectures (such as VGG and ResNet) are approaching their theoretical performance limits, requiring the exploration of more complex structures (such as multi-branch and dynamic routing) to overcome bottlenecks.

[0004] Neural architecture search arose precisely to address these problems. Neural network search is a complex optimization problem, aiming to find a neural network that satisfies specific objective constraints from a large number of candidate network architectures. Existing neural architecture search optimization methods suffer from single-objective bias; most methods only optimize accuracy, ignoring constraints such as latency and energy consumption in practical deployments, leading to difficulties in model deployment.

[0005] In existing structure search methods, reinforcement learning-based methods use recurrent neural networks (RNNs) as controllers to predict the string representation sequence of the neural network, using the performance of the neural network as a reward to guide the controller to sample better network structures. Due to the need for additional controller training, reinforcement learning-based search is more costly than other strategies when exploring high-dimensional search spaces. Gradient optimization-based structure search methods, on the other hand, introduce a supernet containing all candidate networks, introduce an attention mechanism to operations during training, and then remove less important operations. Due to the continuity of the search space, this method can use gradient guidance, making the search process more efficient. However, due to the size limitation of the supernet, the final searched networks lack diversity. Summary of the Invention

[0006] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a multi-objective, multi-level learning optimization method for classifier-assisted neural network architecture search. This invention achieves efficient search of neural network architectures through a hierarchical particle swarm optimization algorithm assisted by the AdaBoost classifier. Specifically, it solves the neural network architecture search problem by establishing a multi-objective optimization mathematical model, proposes a hierarchical learning-based multi-objective optimization strategy, constructs a transition fitness function that integrates convergence and diversity to achieve hierarchical solutions, trains a hierarchical prediction model using the AdaBoost classifier, and enables lower-level particles to learn from higher-level high-quality solutions. A classifier-assisted population replacement strategy is combined to retain high-potential solutions. After iterative optimization, a solution that satisfies both classification accuracy and inference delay objectives is obtained.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a method for multi-objective, multi-level learning optimization of classifier-assisted neural network architecture search, comprising the following steps:

[0009] The neural network architecture search problem is modeled as a multi-objective optimization problem, and a search space is constructed.

[0010] Obtain convolutional neural network operators and construct a test network for training.

[0011] Pareto optimal solutions are selected by non-dominated sorting, and a dedicated search space is constructed based on operators that preserve the Pareto front.

[0012] A hierarchical learning-based multi-objective evolutionary optimization algorithm with classifier assistance is implemented.

[0013] Latin hypercube sampling is performed to generate an initial solution. The objective function is evaluated on the initial solution, and the results are stored in the archive.

[0014] Initialize model parameters, calculate the diversity and convergence indices for each individual in the initial population, generate transition fitness values, and sort the initial population according to the transition fitness values;

[0015] The fitness is transformed, the solution is hierarchically divided, the surrogate model AdaBoost classifier is trained, multi-objective particle swarm evolution is carried out based on hierarchical learning, and the population is updated with classifier assistance.

[0016] A candidate solution selection strategy is constructed to select from the updated population after evolution. The objective function is evaluated based on the selected solutions, the database is updated, and after iterative training, the Pareto optimal solution in the database is output. The network in the optimal Pareto front of the last generation population is selected as the optimal neural network architecture.

[0017] As a preferred technical solution, the neural network architecture search problem is modeled as a multi-objective optimization problem, expressed as:

[0018]

[0019] Where F(x) represents the target vector, f1 is the network validation set classification error rate, and f2 is the actual inference latency. w represents the cross-entropy loss of the network on the training set. * (x) represents the optimal weights that minimize this loss, where x is the candidate neural network, and Ω... x Ω represents the search space of the neural network. w Let w represent the weight vector space, and w represent the weight.

[0020] As a preferred technical solution, the fitness is transformed, specifically expressed as follows:

[0021] TF(p i ,P)=α·div(p i ,P)+β·conv(p i ,P)

[0022] Among them, div(p i ,P) represents the density estimation result based on shift-based estimation, conv(p i ,P) represents the solution p i The Euclidean distance to the ideal point, α and β are equilibrium coefficients, and P represents the solution p. i Other solutions in the surrounding area.

[0023] As a preferred technical solution, training the AdaBoost classifier surrogate model specifically includes:

[0024] The model's input and output are the decision variables and hierarchical labels of the solutions in the archive, respectively.

[0025] Initialize sample weights, iteratively train multiple weak classifiers, increase the weight of hard-to-classify samples in each iteration, and determine the solution level of the strong classifier through weighted voting.

[0026] As a preferred technical solution, the calculation formula for a strong classifier is expressed as follows:

[0027]

[0028] Where, β t h represents the weights of the weak classifier. t (x,y) represents the prediction of the t-th weak classifier that the solution x belongs to class y, where Y represents the number of classes and T represents the number of weak classifiers.

[0029] As a preferred technical solution, multi-objective particle swarm evolution is performed based on hierarchical learning, and the particle velocity is updated as follows:

[0030] v j,i =w·v j,i +r1·(z k1 -x j,i )+r2·(z k2 -x j,i )+r3·(z k1 -z k2 )

[0031] Among them, v j,i Let z represent the velocity of the i-th particle in the j-th layer. k1 ,z k2 This represents two guide particles randomly selected from a higher level, w is the inertia coefficient, and r1, r2, r3 ~ U(0,1) represent uniform random numbers.

[0032] As a preferred technical solution, a classifier-assisted population update is performed. After generating the offspring solution, the AdaBoost classifier is used to predict its hierarchical level. If the predicted level of the offspring solution is higher than that of the parent solution, the parent solution is replaced by the offspring solution.

[0033] As a preferred technical solution, a candidate solution screening strategy is constructed, specifically including:

[0034] Pre-screening is performed to obtain candidate solutions P′. The candidate solutions P′ are then compared with the top N solutions with the highest diversity in the archive. s The mutual information entropy of the solutions is calculated, and the N′ solutions with the smallest entropy values ​​are retained. K-means clustering is then performed to cluster these N′ solutions into Nn clusters. s For each class, one solution is selected for evaluation of the objective function.

[0035] The present invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the above-described classifier-assisted neural network architecture search multi-objective multi-level learning optimization method.

[0036] The present invention also provides a computer device, including a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described classifier-assisted neural network architecture search multi-objective multi-level learning optimization method.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] (1) This invention solves the problem of neural network architecture search by establishing a mathematical model for multi-objective optimization, proposes a multi-objective optimization strategy based on hierarchical learning, constructs a transformation fitness function that integrates convergence and diversity to realize the hierarchical solution, uses the AdaBoost classifier to train the hierarchical prediction model and enables low-level particles to learn from high-level high-quality solutions, and combines a population replacement strategy assisted by the classifier to retain high-potential solutions. After iterative optimization, a solution that satisfies the dual objectives of classification accuracy and inference delay is obtained.

[0039] (2) Based on the transformation fitness method, this invention takes into account the diversity and convergence of solutions. In order to treat diversity and convergence equally, the same importance is set when evaluating fitness. Then, the population is sorted according to the level of this comprehensive fitness and the population is divided into levels. Solutions of higher levels will provide guidance for the search direction of solutions of lower levels. Through this hierarchical learning method, the optimization efficiency of the entire population is improved.

[0040] (3) This invention adopts a hierarchical evolution strategy to realize a multi-objective hierarchical evolution optimizer. In terms of particle update, for particles at lower levels, their speed update is affected by several factors: on the one hand, there is the effect of inertia, which keeps the particles in a certain motion trend; on the other hand, two particles will be randomly selected from higher levels to guide their motion, one providing guidance to move closer to a better state, and the other increasing the randomness of the motion direction; at the same time, some random factors will be introduced to make the search process more exploratory. In terms of replacement rules, if the potential level of the offspring solution is higher than that of the parent solution as predicted by the classifier, the offspring solution will replace the parent solution, driving the entire population to evolve towards a better direction. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the multi-objective, multi-level learning optimization method for classifier-assisted neural network architecture search according to the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides a classifier-assisted neural network architecture search multi-objective, multi-level learning optimization method. It applies multi-objective optimization algorithms to neural network architecture search, reducing the number of expensive evaluations by predicting candidate architecture performance through a classifier. Specifically, it includes the following steps:

[0045] S1: Model the neural network architecture search problem as a multi-objective optimization problem, constructing the search space, specifically including:

[0046] S11: Treat the neural network architecture search problem as a multi-objective optimization problem, which can be modeled and expressed by the following formula:

[0047]

[0048] Among them, Ω x Let Ω represent the neural network search space, x be a candidate neural network, and Ω be the Ω space. w Let w represent the weight vector space, w∈Ω w The target vector F, associated with its weights, consists of two parts: f1 is the network validation set classification error rate, and f2 is the actual inference latency. w represents the cross-entropy loss of the network on the training set. * (x) is the optimal weight that minimizes the loss;

[0049] S12: Obtain mainstream convolutional neural network operators, including standard convolution, residual convolution, bottleneck convolution, inverted residual convolution, etc., and linearly stack these operators to build a test network;

[0050] S13: Train the test network for 100 generations on the CIFAR-10 dataset to obtain the classification accuracy on the validation set; at the same time, measure the inference latency to form an "accuracy-latency" performance pair.

[0051] S14: Select Pareto optimal solutions by non-dominated sorting, and construct a dedicated search space containing 8 operations based on the Pareto front-preserving operators to exclude inefficient operators;

[0052] S2: Based on the two objectives of network validation set classification error rate and actual inference latency mentioned above, a classifier-assisted hierarchical learning-based multi-objective evolutionary optimization algorithm is executed, specifically including:

[0053] S21: Initialization and Archive Construction: First, Latin hypercube sampling is performed to generate 11×(D-1) initial solutions (D being the dimension of the decision variables), ensuring that the solutions are uniformly distributed in the decision space and avoiding the clustering problem of random sampling. Then, expensive evaluation is performed to evaluate the initial solutions using the true objective function, and the results are stored in the archive Arc as the basic data for subsequent classifier training. At the same time, model parameters are initialized, the diversity and convergence indices of each individual in the initial population are calculated, a transition fitness value is generated, and the initial population is sorted according to this value, updating the archive. This process provides the initial dataset for subsequent hierarchical evolution and classifier training.

[0054] S22: The fitness is transformed. First, the multi-objective problem is transformed into a single-objective problem. Specifically, a transformation fitness function (TF) is designed to convert the multi-objective optimization problem into a single-objective ranking problem. This function combines the contributions of convergence (distance from the ideal point) and diversity (solution distribution density).

[0055] TF(p i ,P)=α·div(p i ,P)+β·conv(p i ,P)

[0056] Among them, div(p i The calculation of p is based on shift-based density estimation (SDE), specifically, it calculates the solution p. i The Euclidean distance to other solutions P in the surrounding area indicates that a solution has good diversity and can bring a wider search range to the algorithm. The larger the value, the more dispersed the solution is in space and the better the diversity.

[0057] conv(p i The calculation basis of P) is the solution p. i To the ideal point z * The Euclidean distance between the points (comprising the minimum values ​​of each objective) is such that a smaller value indicates better convergence. α and β are balance coefficients, and in this embodiment, α = 0.5 and β = 0.5 are preferred to balance diversity and convergence.

[0058] In this embodiment, based on a fitness transformation approach, the diversity and convergence of solutions are considered comprehensively. Specifically, diversity is measured by a method based on a specific density estimation, examining the dispersion of the solution's distribution relative to other solutions in the population; a more dispersed distribution indicates better diversity. Convergence is assessed by how close the solution is to the ideal state (represented by the minimum achievable value of each objective function); the closer to the ideal state, the better the convergence. Furthermore, to treat diversity and convergence equally, they are given the same importance when evaluating fitness. The population is then sorted according to this comprehensive fitness, dividing it into four levels. Solutions at higher levels provide search direction guidance for solutions at lower levels, improving the overall optimization efficiency of the population through this hierarchical learning approach.

[0059] S23: Implement stratification of solutions, sort them in descending order according to TF value, and divide the population into NL layers. Higher-level solutions guide the learning of lower-level solutions. In this embodiment, NL=4 is preferred, such as Level 1 being the best layer and Level 4 being the worst layer.

[0060] S24: Train the surrogate model AdaBoost classifier: The model's input and output are the decision variables x from the solutions in the archive.i ∈R D Reconciliation hierarchy label y i ∈{1,2,3,4} (corresponding to Level 1-4). First, initialize the sample weights D. t (i,j), iteratively train several weak classifiers (decision trees) (preferably 100 in this embodiment), increasing the weight of hard-to-classify samples in each iteration. Finally, the strong classifier determines the solution level through weighted voting. The calculation formula for the strong classifier is as follows:

[0061]

[0062] Where, β t h represents the weights of the weak classifier. t (x,y) represents the prediction of the t-th weak classifier that the solution x belongs to the category y;

[0063] S25: Perform multi-objective particle swarm evolution based on hierarchical learning. The designed velocity update formula introduces dual-guided particles to enhance search directionality and perturbation. Lower-level particles (such as Level 3-4) learn from higher-level particles (Level 1-2). The velocity update formula is as follows:

[0064] v j,i =w·v j,i +r1·(z k1 -x j,i )+r2·(z k2 -x j,i )+r3·(z k1 -z k2 )

[0065] Among them, v j,i z represents the velocity of the i-th particle in the j-th layer; k1 ,z k2 This represents two guide particles randomly selected from a higher layer, introducing dual guide directions to increase search perturbation; w is the inertia coefficient, preferably w = 0.4 in this embodiment; r1, r2, r3 ~ U(0,1) represent uniform random numbers. After updating the particle velocity, the particle position is updated according to the formula:

[0066] x j,i =x j,i +v j,i

[0067] S26: Perform classifier-assisted population update. After the offspring solution is generated, the AdaBoost classifier is used to predict its hierarchical level. If the predicted level of the offspring solution is higher than that of the parent solution, a replacement operation is triggered to replace the parent solution with the offspring solution. The purpose of this replacement strategy is to prioritize the retention of solutions with higher prediction potential and promote the population to evolve towards higher levels (i.e., better Pareto front regions).

[0068] In this embodiment, a hierarchical evolution strategy is employed to implement a multi-objective hierarchical evolutionary optimizer. Regarding particle updates, the velocity update of lower-level particles is influenced by several factors. Firstly, inertia maintains a certain motion trend. Secondly, two particles are randomly selected from higher levels to guide their motion; one provides guidance towards a better state, while the other increases the randomness of the motion direction. Simultaneously, some random factors are introduced to make the search process more exploratory. Regarding the replacement rule, if the classifier predicts that the potential level of the offspring solution is higher than that of the parent solution, the offspring solution replaces the parent solution, driving the entire population towards a better direction.

[0069] S27: Construct a candidate solution selection strategy to select solutions from the evolved, updated population. The selected solutions will be used for the actual, expensive evaluation. First, pre-screening is performed, retaining only solutions predicted by the classifier to be Level 1, denoted as P′. This pre-screening filters out low-potential solutions. Then, candidate solutions are filtered based on cross-information entropy. P′ is calculated against the "diversity leaders" (the top N most diverse solutions in the archive). s Mutual information entropy of solutions:

[0070]

[0071] Wherein, p(l i ,p′ j For leaders i With solution p′ j The joint probability, retaining the one with the minimum entropy. In this embodiment, λ = 0.7 is preferred as these solutions have the highest similarity. Finally, K-means clustering is performed to cluster the N′ solutions into Nn clusters. s Each class is used to select one solution for expensive evaluation to ensure uniform spatial distribution;

[0072] In this embodiment, the diversity-based candidate solution selection strategy consists of three stages: First, a pre-screening stage, where a classifier filters out solutions with low prediction potential, retaining only those with higher prediction levels to form a high-potential candidate set, thereby reducing the workload and cost of subsequent detailed evaluation of solutions; next, an entropy filtering stage, where the similarity between solutions in the candidate set and highly diverse solutions in the archive is calculated, retaining those solutions most similar to these highly diverse solutions to ensure that the selected solutions have good diversity; finally, a clustering selection stage, where the solutions retained after the first two stages are clustered, and one solution is randomly selected from each group for detailed objective function evaluation, thus ensuring that the evaluated solutions are spatially evenly distributed and focusing on the most promising regions.

[0073] S28: Perform iterations, terminating when the termination condition is met. Repeat the process of "classifier training → hierarchical learning-based multi-objective particle swarm optimization → candidate solution selection → actual expensive evaluation → updating the database" until the maximum number of evaluations (FE) is reached. max Output the Pareto optimal solution from the archive;

[0074] S29: Select the network in the optimal Pareto front of the last generation population as the optimal neural network architecture.

[0075] Example 2

[0076] This embodiment provides a storage medium, which may be a ROM, RAM, disk, optical disk, or other storage medium. The storage medium stores one or more programs. When the programs are executed by a processor, they implement the classifier-assisted neural network architecture search multi-objective multi-level learning optimization method of Embodiment 1.

[0077] Example 3

[0078] This embodiment provides a computing device, which may be a desktop computer, laptop computer, smartphone, PDA handheld terminal, tablet computer or other terminal device with display function. The computing device includes a processor and a memory. The memory stores one or more programs. When the processor executes the program stored in the memory, it implements the classifier-assisted neural network architecture search multi-objective multi-level learning optimization method of Embodiment 1.

[0079] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for multi-objective, multi-level learning optimization of a classifier-assisted neural network architecture search, characterized in that, Includes the following steps: The neural network architecture search problem is modeled as a multi-objective optimization problem, and a search space is constructed. Obtain convolutional neural network operators and construct a test network for training. Pareto optimal solutions are selected by non-dominated sorting, and a dedicated search space is constructed based on operators that preserve the Pareto front. A hierarchical learning-based multi-objective evolutionary optimization algorithm with classifier assistance is implemented. Latin hypercube sampling is performed to generate an initial solution. The objective function is evaluated on the initial solution, and the results are stored in the archive. Initialize model parameters, calculate the diversity and convergence indices for each individual in the initial population, generate transition fitness values, and sort the initial population according to the transition fitness values; The fitness is transformed, the solution is hierarchically divided, the surrogate model AdaBoost classifier is trained, multi-objective particle swarm evolution is carried out based on hierarchical learning, and the population is updated with classifier assistance. A candidate solution selection strategy is constructed to select from the updated population after evolution. The objective function is evaluated based on the selected solutions, the database is updated, and after iterative training, the Pareto optimal solution in the database is output. The network in the optimal Pareto front of the last generation population is selected as the optimal neural network architecture.

2. The classifier-assisted neural network architecture search multi-objective multi-level learning optimization method according to claim 1, characterized in that, The neural network architecture search problem can be modeled as a multi-objective optimization problem, represented as: Where F(x) represents the target vector, f1 is the network validation set classification error rate, and f2 is the actual inference latency. w represents the cross-entropy loss of the network on the training set. * (x) represents the optimal weights that minimize this loss, where x is the candidate neural network, and Ω... x Ω represents the search space of the neural network. w Let w represent the weight vector space, and w represent the weight.

3. The classifier-assisted neural network architecture search multi-objective multi-level learning optimization method according to claim 1, characterized in that, The fitness is transformed, specifically represented as follows: TF(p i ,P)sα·div(p i ,P)+β·conv(p). i ,P) Among them, div(p i ,P) represents the density estimation result based on shift-based estimation, conv(p i ,P) represents the solution p i The Euclidean distance to the ideal point, α and β are equilibrium coefficients, and P represents the solution p. i Other solutions in the surrounding area.

4. The classifier-assisted neural network architecture search multi-objective multi-level learning optimization method according to claim 1, characterized in that, Training the AdaBoost classifier surrogate model specifically includes: The model's input and output are the decision variables and hierarchical labels of the solutions in the archive, respectively. Initialize sample weights, iteratively train multiple weak classifiers, increase the weight of hard-to-classify samples in each iteration, and determine the solution level of the strong classifier through weighted voting.

5. The classifier-assisted neural network architecture search multi-objective multi-level learning optimization method according to claim 4, characterized in that, The formula for calculating a strong classifier is expressed as: Where, β t h represents the weights of the weak classifier. t (x,y) represents the prediction of the t-th weak classifier that the solution x belongs to class y, where Y represents the number of classes and T represents the number of weak classifiers.

6. The classifier-assisted neural network architecture search multi-objective multi-level learning optimization method according to claim 1, characterized in that, Multi-objective particle swarm evolution is performed based on hierarchical learning, and the particle velocity is updated as follows: in j,i =w·v j,i +r1·(z k1 -x j,i )+r2·(z k2 -x j,i )+r3·(z k1 -z k2 ) Among them, v j,i Let z represent the velocity of the i-th particle in the j-th layer. k1 ,z k2 This represents two guide particles randomly selected from a higher level, w is the inertia coefficient, and r1, r2, r3 ~ U(0,1) represent uniform random numbers.

7. The classifier-assisted neural network architecture search multi-objective multi-level learning optimization method according to claim 1, characterized in that, After performing classifier-assisted population updates and generating offspring solutions, the AdaBoost classifier is used to perform hierarchical predictions on them. If the predicted hierarchical level of the offspring solution is higher than that of the parent solution, the parent solution is replaced by the offspring solution.

8. The classifier-assisted neural network architecture search multi-objective multi-level learning optimization method according to claim 1, characterized in that, Constructing a candidate solution selection strategy, specifically including: Pre-screening is performed to obtain candidate solutions P′. The candidate solutions P′ are then compared with the top N solutions with the highest diversity in the archive. s The mutual information entropy of the solutions is calculated, and the N′ solutions with the smallest entropy values ​​are retained. Then, K-means P′ clustering is performed to cluster the N′ solutions into Nn clusters. s For each class, one solution is selected for evaluation of the objective function.

9. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the classifier-assisted neural network architecture search multi-objective multi-level learning optimization method as described in any one of claims 1-8.

10. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the classifier-assisted neural network architecture search multi-objective multi-level learning optimization method as described in any one of claims 1-8.

Images