CN110766072A - Automatic generation method of computational graph evolution AI model based on structural similarity - Google Patents
Automatic generation method of computational graph evolution AI model based on structural similarity Download PDFInfo
- Publication number
- CN110766072A CN110766072A CN201911006315.6A CN201911006315A CN110766072A CN 110766072 A CN110766072 A CN 110766072A CN 201911006315 A CN201911006315 A CN 201911006315A CN 110766072 A CN110766072 A CN 110766072A
- Authority
- CN
- China
- Prior art keywords
- model
- graph
- models
- calculation
- similarity
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 57
- 238000004364 calculation method Methods 0.000 claims abstract description 139
- 238000013528 artificial neural network Methods 0.000 claims description 34
- 239000011159 matrix material Substances 0.000 claims description 31
- 230000002068 genetic effect Effects 0.000 claims description 22
- 238000013461 design Methods 0.000 claims description 16
- 210000002569 neuron Anatomy 0.000 claims description 16
- 108090000623 proteins and genes Proteins 0.000 claims description 13
- 238000012549 training Methods 0.000 claims description 11
- 239000013598 vector Substances 0.000 claims description 11
- 238000004519 manufacturing process Methods 0.000 claims description 8
- 230000000717 retained effect Effects 0.000 claims description 8
- 230000035772 mutation Effects 0.000 claims description 5
- 238000009826 distribution Methods 0.000 claims description 4
- 239000002689 soil Substances 0.000 claims description 3
- 238000002360 preparation method Methods 0.000 claims 1
- 230000015556 catabolic process Effects 0.000 abstract description 2
- 238000006731 degradation reaction Methods 0.000 abstract description 2
- 230000009191 jumping Effects 0.000 abstract description 2
- 230000002401 inhibitory effect Effects 0.000 abstract 1
- 239000013604 expression vector Substances 0.000 description 4
- 238000005070 sampling Methods 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 230000002787 reinforcement Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000002028 premature Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
Abstract
The invention provides a computation graph evolution AI model automatic generation method based on structure similarity, which adopts graph similarity technology-a method combining multiple graph similarity calculations in the AI model automatic generation method based on computation graph evolution, and is used for calculating the similarity calculation between a generated new network model and a known network model in the automatic generation process of an AI model, and inhibiting the occurrence of repeated models, similar models and performance similar models in a certain similarity threshold range. The diversity of model samples in the model searching process can be effectively ensured, and the success rate of model network searching is improved; the situation of performance degradation of the search network is obviously reduced. Meanwhile, the similarity threshold is dynamically adjusted according to the searching efficiency, so that the function of jumping out of local optimum can be realized according to the model sample, and the model searching is accelerated. Thereby improving the efficiency of automatic generation of the AI model.
Description
Technical Field
The invention relates to the technical field of AI model (AI model, namely artificial intelligence model) correlation, in particular to a calculation chart evolution AI model automatic generation method based on structural similarity.
Background
AI model auto-generation is a leading-edge area of research. Automatic model generation can generate a simpler and more efficient neural network model from the distribution of data. The search space automatically generated by the AI model is fn×2n(n-1)/2Wherein f is the number of operators of different neurons, and n is the maximum depth of the neural network. It can be seen that in the generation process, as the number of supported neural network operators increases and the network model deepens, the complexity of the problem may become a problem approaching an infinite search space, thereby causing an inability to solve.
At present, the main search methods include reinforcement learning (i.e., reinforcement learning), monte carlo tree search (random sampling or statistical experimental method), and the like. However, these methods need to accumulate certain statistical information first, and a good neural network model structure may be searched in a limited period after a prior probability effective for model design is generated. After the traditional algorithm fits the selected network, it can only run completely to find further search directions. However, in the deep learning field, the actual time of one training may be tens of minutes, even tens of hours. And in many cases, when the search approaches towards the optimal solution, the network difference is also reduced, and the training results of similar networks are very similar, so that the whole model search process is very long. At present, one complete training of deep learning is different from hours to weeks, and the optimal solution can be found only on the basis of a large amount of training required by automatic neural network design, and the problem of no solution is almost achieved under the condition of the existing computing power along with the deepening of the network.
In the process of searching and designing the network model, a large number of repeated, similar-structure or similar-performance network models can appear, and due to the fact that the time consumption of the calculation process of the model performance is serious, the diversity of samples in the process of the evolutionary algorithm is reduced due to the large number of repeated-structure and similar models, the problem that the searched model performance is poor and the like is caused under the condition that the searching efficiency of the whole evolutionary algorithm is low.
Disclosure of Invention
The purpose of the invention is:
in order to overcome the defects of the prior art, in the AI model automatic generation method based on computational graph evolution, a graph similarity technology-a method combining multiple graph similarity calculations is adopted for calculating the similarity calculation between a generated new network model and a known network model in the automatic generation process of an AI model, and the occurrence of repeated models, similar models and performance similar models is greatly inhibited within a certain similarity threshold range. Compared with the method of only removing repeated models, the method of removing similar models adopted in the initial stage of evolution search can effectively ensure the diversity of model samples in the process of model search, has the probability of obviously searching the optimal model for the network model search calculation with smaller scale, and obviously improves the success rate of model network search; for network model search with large scale, the model sampling space can be fully sampled in a short time, so that search in the local optimal range is realized, and the condition of performance degradation of the search network is obviously reduced. Meanwhile, the similarity threshold is dynamically adjusted according to the searching efficiency, so that the function of jumping out of local optimum can be realized according to the model sample, and the model searching is accelerated. Thereby improving the efficiency of automatic generation of the AI model.
The purpose of the invention can be realized by the following technical scheme:
a computation graph evolution AI model automatic generation method based on structural similarity comprises the following steps:
step (1): according to data preset by a user, preparing data, setting production parameters of a model design platform, and setting a new model configuration threshold, wherein the new model configuration threshold comprises the following steps: the model is automatically designed for a distance threshold, or a similarity threshold. The data preset by the user in this step includes statistical distribution of the data, correlation coefficients among data dimensions, and/or statistical correlations between each dimension of the data and the tags. The model design platform production parameters set in this step include computing resources, job run time, job targets and/or genetic algorithm parameters. The operation target comprises a fitness threshold value of the calculation graph model: a fitness threshold that is considered to satisfy an evolution termination condition and a fitness threshold that is considered to be an invalid model are included.
Step (2): and generating a first generation computational graph model by using a genetic algorithm operator. The genetic algorithm operators include random operators, crossover operators and/or mutation operators. The random operator is used for randomly selecting the number of the neurons, randomly selecting the types of the neurons and/or randomly determining the connection relation of the neurons.
And (3): judging whether all new calculation map models generated in the previous step are repeated or not or are calculation map models with approximate performance, wherein the judgment method for judging whether the new calculation map models in the step (3) are the calculation map models with the approximate performance comprises the following steps: judging the structural similarity of the models; if the model is judged to have the models which are mutually repeated or the models with similar performance, one of the models is reserved, and the rest repeated models and the similar models are removed;
and (4): according to the model retained in step (3):
a. calculating the performance of each computational graph model;
b. and calculating the fitness of each computational graph model according to the performance and the complexity of the computational graph. The complexity refers to the complexity calculated according to the number of nodes and the number of edges of the calculation graph.
And (5): removing the useless model according to the model fitness, taking the rest models as alternative models, and reserving the alternative models as next generation seeds;
and (6): picking out a plurality of optimal models according to the next generation of seeds reserved in the step (5);
and (7): and (4) generating a new calculation chart model by using a genetic algorithm operator according to the alternative model selected in the step (6) and used as the next generation of seeds. The genetic algorithm operators include random operators, crossover operators and/or mutation operators. The random operator is used for randomly selecting the number of the neurons, randomly selecting the types of the neurons and/or randomly determining the connection relation of the neurons.
And (8): judging whether the new model of the computational graph generated in the last step is the generated computational graph model or not, or whether the new model of the computational graph is the computational graph model with performance similar to the generated computational graph model, and if not, entering the step (9); if so, returning to the step (7); in the step (8), the method for determining whether the model is a model with performance similar to that of the generated computation graph model includes: and judging the structural similarity of the models.
And (9): saving the retained computational graph model of the step (8) as a new generation computational graph model;
step (10): judging whether the number of the new generation of calculation graph models in the step (9) meets the number of the preset population in the step (1), if yes, entering the next step; if not, returning to the step (7);
step (11): calculating to obtain the performance of all the new generation of calculation graph models stored in the last step;
step (12):
a. and searching for the optimal solution or the suboptimal solution close to the optimal solution by using the hyperparametric search for the model with the life cycle exceeding three generations, wherein the definition of the life cycle exceeding three generations is the same as the definition of the generation number in the genetic algorithm, and the model structure is calculated as the first generation from the first occurrence in the process. The hyper-parameter in the hyper-parameter search refers to control parameters of a neural network inside the AI, including a learning rate, namely parameters and/or weight attenuation parameters. And (4) entering the model reserved after the super-parameter search into step (13), and calculating the fitness of the new calculation graph model according to the performance of the calculation graph calculated after the super-parameter search and the complexity of the model. The complexity refers to the complexity calculated according to the number of nodes and the number of edges of the calculation graph.
b. For the model with life cycle not exceeding three generations, calculating model performance according to the structure of the computation graph, calculating the fitness of each computation graph model according to the performance and the complexity of the computation graph, and then entering step (13). The complexity refers to the complexity calculated according to the number of nodes and the number of edges of the calculation graph.
Step (13): judging whether the new model of the calculation graph meets the evolution ending condition preset in the step (1), and if so, entering the step (14); if not, returning to the step (5);
step (14): and summarizing the evolution calculation results, carrying out comprehensive scoring according to the complexity and the accuracy of the model, and selecting the optimal model.
The method for judging the structural similarity of the models in the steps (3) and (8) comprises the following steps:
the first judgment method comprises the following steps:
a. presetting a new model configuration threshold, namely a distance threshold, wherein the distance threshold can be dynamically set;
b. constructing a distance vector according to a connection mode of the calculation map model, and calculating the distance between the calculation map model to be judged and all the generated models according to the distance vector, wherein the distance can be an editing distance, a soil moving distance, a cosine distance or a Euclidean distance;
c. and if the distance value between the newly generated computational graph model and all the generated models is lower than the preset new model configuration threshold, judging the newly generated computational graph model as a model with the performance similar to that of the generated computational graph model.
The second judgment method comprises the following steps:
a. presetting a new model configuration threshold, namely a similarity threshold, wherein the similarity threshold can be dynamically set; preparing a neural network for judging the similarity in advance, wherein the establishment and training method of the neural network is as described in the steps b-e;
b. expressing the structure of the computation graph model by using a computation graph expression matrix: in a system of graph node types with N types (N is the number of gene types) of different genes, for a computation graph with the number of nodes being M (M is the number of nodes in the computation graph), expressing the connection relation of the computation graph by using a computation graph expression matrix of M, wherein the value in the computation graph expression matrix is the gene connection type, and the value range of the gene connection type is 1 to N;
c. preparing a group of known data sets, wherein the known data sets comprise a computation graph expression matrix and a similarity score, the computation graph expression matrix comprises node information (namely the number and the type of nodes) of a graph, edge connection information (edge connection condition, namely whether edges are connected or not), and the similarity score ranges from 0 to 1.0 (the similarity score is 0, namely completely dissimilar, and the similarity score is 1.0, namely 100%);
d. establishing a neural network, wherein the neural network comprises a graph embedding layer, a graph volume layer, a volume layer and a full connection layer, the output is a scalar, the network output value is defined as a similarity score, and the value range of the similarity score is 0.0-1.0 (the score is 0, namely completely dissimilar, and the score is 1.0, namely the similarity is 100%); the input of the neural network is a calculation graph expression matrix of two calculation graphs;
e. training the neural network of step d to stable convergence using the data set of step c, with the input being the computational graph expression matrix of step b;
f. and e, calculating the similarity between the newly generated computational graph model and all the generated models through the trained neural network in the step e, inputting the similarity into the newly generated computational graph model and a computational graph expression matrix of any generated model, and outputting the similarity of the two models, wherein the similarity is higher than a preset new model configuration threshold, and the newly generated computational graph model is judged to be a model with the performance similar to that of the generated computational graph model.
The invention has the beneficial effects that:
1. the invention improves the efficiency of model design, improves the success rate of searching the optimal model and reduces the time of model design: compared with the same model and the similar model, the method can avoid the times of the same model and the structural similar model which are calculated in the model searching calculation, and can design more different models under the condition of designing the same number of models, thereby improving the model design efficiency. In order to increase the probability of searching for the optimal model, the number of models to be calculated is reduced and the time for designing the models is also reduced when an equal number of different models are designed.
2. The diversity in the model design is improved, the fully sampled probability of a sampling space is improved, and the optimal search in the global scope is realized: because the appearance of the same and similar models is inhibited, the model with low similarity can be ensured to exist in the design process of each generation of model, so that the diversity of the model is improved, the probability of fully sampling the whole model search space is improved, the global search in the model search process is ensured, and the premature condition that the model with high similarity still appears after the model with high similarity is subjected to cross variation design is prevented.
3. Different model distance similarity calculation methods (editing distance, earth moving distance, cosine distance or Euclidean distance is used for similarity calculation) and neural network similarity calculation methods are flexibly adopted according to different scenes, so that the models which are applied to different scales and different fields can be flexibly used in the neural network similarity calculation method, and the application range of model similarity calculation is enlarged.
4. The method can control the trend of model design and realize the conditions of local optimum, low diversity, low search efficiency and the like in the jumping-out search process by dynamically setting the setting of the similarity threshold.
5. The judgment of the structure similarity model in the invention is not influenced by other algorithms, can be independent, and can flexibly combine different algorithms to form different algorithm combinations.
Drawings
Fig. 1 is a general flow chart of an implementation of the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific examples.
As shown in fig. 1, the steps of the present invention are as follows:
a computation graph evolution AI model automatic generation method based on structural similarity comprises the following steps:
step (1): according to data preset by a user, preparing data, setting production parameters of a model design platform, and setting a new model configuration threshold, wherein the new model configuration threshold comprises the following steps: the model is automatically designed for a distance threshold, or a similarity threshold. The data preset by the user in this step includes statistical distribution of the data, correlation coefficients among data dimensions, and/or statistical correlations between each dimension of the data and the tags. The model design platform production parameters set in this step include computing resources, job run time, job targets and/or genetic algorithm parameters. The operation target comprises a fitness threshold value of the calculation graph model: a fitness threshold that is considered to satisfy an evolution termination condition and a fitness threshold that is considered to be an invalid model are included.
Step (2): and generating a first generation computational graph model by using a genetic algorithm operator. The genetic algorithm operators include random operators, crossover operators and/or mutation operators. The random operator is used for randomly selecting the number of the neurons, randomly selecting the types of the neurons and/or randomly determining the connection relation of the neurons.
And (3): judging whether all new calculation map models generated in the previous step are repeated or not or are calculation map models with approximate performance, wherein the judgment method for judging whether the new calculation map models in the step (3) are the calculation map models with the approximate performance comprises the following steps: judging the structural similarity of the models; if the model is judged to have the models which are mutually repeated or the models with similar performance, one of the models is reserved, and the rest repeated models and the similar models are removed;
and (4): according to the model retained in step (3):
a. calculating the performance of each computational graph model;
b. and calculating the fitness of each computational graph model according to the performance and the complexity of the computational graph. The complexity refers to the complexity calculated according to the number of nodes and the number of edges of the calculation graph.
And (5): removing the useless model according to the model fitness, taking the rest models as alternative models, and reserving the alternative models as next generation seeds;
and (6): picking out a plurality of optimal models according to the next generation of seeds reserved in the step (5);
and (7): and (4) generating a new calculation chart model by using a genetic algorithm operator according to the alternative model selected in the step (6) and used as the next generation of seeds. The genetic algorithm operators include random operators, crossover operators and/or mutation operators. The random operator is used for randomly selecting the number of the neurons, randomly selecting the types of the neurons and/or randomly determining the connection relation of the neurons.
And (8): judging whether the new model of the computational graph generated in the last step is the generated computational graph model or not, or whether the new model of the computational graph is the computational graph model with performance similar to the generated computational graph model, and if not, entering the step (9); if so, returning to the step (7); in the step (8), the method for determining whether the model is a model with performance similar to that of the generated computation graph model includes: and judging the structural similarity of the models.
And (9): saving the retained computational graph model of the step (8) as a new generation computational graph model;
step (10): judging whether the number of the new generation of calculation graph models in the step (9) meets the number of the preset population in the step (1), if yes, entering the next step; if not, returning to the step (7);
step (11): calculating to obtain the performance of all the new generation of calculation graph models stored in the last step;
step (12):
a. and searching for the optimal solution or the suboptimal solution close to the optimal solution by using the hyperparametric search for the model with the life cycle exceeding three generations, wherein the definition of the life cycle exceeding three generations is the same as the definition of the generation number in the genetic algorithm, and the model structure is calculated as the first generation from the first occurrence in the process. The hyper-parameter in the hyper-parameter search refers to control parameters of a neural network inside the AI, including a learning rate, namely parameters and/or weight attenuation parameters. And (4) entering the model reserved after the super-parameter search into step (13), and calculating the fitness of the new calculation graph model according to the performance of the calculation graph calculated after the super-parameter search and the complexity of the model. The complexity refers to the complexity calculated according to the number of nodes and the number of edges of the calculation graph.
b. For the model with life cycle not exceeding three generations, calculating model performance according to the structure of the computation graph, calculating the fitness of each computation graph model according to the performance and the complexity of the computation graph, and then entering step (13). The complexity refers to the complexity calculated according to the number of nodes and the number of edges of the calculation graph.
Step (13): judging whether the new model of the calculation graph meets the evolution ending condition preset in the step (1), and if so, entering the step (14); if not, returning to the step (5);
step (14): and summarizing the evolution calculation results, carrying out comprehensive scoring according to the complexity and the accuracy of the model, and selecting the optimal model.
The method for judging the structural similarity of the models in the steps (3) and (8) comprises the following steps:
the first judgment method comprises the following steps:
a. presetting a new model configuration threshold, namely a distance threshold, wherein the distance threshold can be dynamically set;
b. constructing a distance vector according to a connection mode of the calculation map model, and calculating the distance between the calculation map model to be judged and all the generated models according to the distance vector, wherein the distance can be an editing distance, a soil moving distance, a cosine distance or a Euclidean distance;
c. and if the distance value between the newly generated computational graph model and all the generated models is lower than the preset new model configuration threshold, judging the newly generated computational graph model as a model with the performance similar to that of the generated computational graph model. The second judgment method comprises the following steps:
g. presetting a new model configuration threshold, namely a similarity threshold, wherein the similarity threshold can be dynamically set; preparing a neural network for judging the similarity in advance, wherein the establishment and training method of the neural network is as described in the steps b-e;
h. expressing the structure of the computation graph model by using a computation graph expression matrix: in a system of graph node types with N types (N is the number of gene types) of different genes, for a computation graph with the number of nodes being M (M is the number of nodes in the computation graph), expressing the connection relation of the computation graph by using a computation graph expression matrix of M, wherein the value in the computation graph expression matrix is the gene connection type, and the value range of the gene connection type is 1 to N;
i. preparing a group of known data sets, wherein the known data sets comprise a computation graph expression matrix and a similarity score, the computation graph expression matrix comprises node information (namely the number and the type of nodes) of a graph, edge connection information (edge connection condition, namely whether edges are connected or not), and the similarity score ranges from 0 to 1.0 (the similarity score is 0, namely completely dissimilar, and the similarity score is 1.0, namely 100%);
j. establishing a neural network, wherein the neural network comprises a graph embedding layer, a graph volume layer, a volume layer and a full connection layer, the output is a scalar, the network output value is defined as a similarity score, and the value range of the similarity score is 0.0-1.0 (the score is 0, namely completely dissimilar, and the score is 1.0, namely the similarity is 100%); the input of the neural network is a calculation graph expression matrix of two calculation graphs;
k. training the neural network of step d to stable convergence using the data set of step c, with the input being the computational graph expression matrix of step b;
and l, calculating the similarity of the newly generated computation graph model and all the generated models through the trained neural network in the step e, inputting the similarity of the newly generated computation graph model and the computation graph expression matrix of any one generated model, outputting the similarity of the two models, and judging the newly generated computation graph model as a model with the performance similar to that of the generated computation graph model when the similarity value is higher than a preset new model configuration threshold.
The first embodiment is as follows:
as described in step (1), a user prepares numerical calculation data, namely platform production parameters (in a csv format or a picture format), and the data comprises a tag column; setting the maximum evolution algebra of the models as 2 generations, and setting the population number of the models of each generation as 3; presetting a new model structure similarity configuration threshold, wherein the required editing distance is not less than 1; the smaller the preset fitness, the better the model performance; calculating a fitness threshold value of the graph model, namely if the fitness of the optimal model is less than 40, considering that an evolution end condition is met, stopping calculation, and when the evolution algebra is more than or equal to 2 generations, considering that the evolution condition is met even if the fitness does not meet the numerical value, and stopping calculation; models with a preset fitness exceeding 800 are considered as invalid models. The fitness in this embodiment is calculated by using the formula F +10 × N.
As described in step (2), randomly generating the first generation of 3 models by using a genetic random operator, wherein the first generation of 3 models respectively comprises: randomly generating 3 first generation models, which are respectively: a computation graph model 1, a computation graph model 2 and a computation graph model 3.
And (4) as described in the step (3), if the models which are repeated mutually or the models with similar performance exist, one of the models is reserved, and the rest repeated models and similar models are removed. For the calculation graph model 1, the calculation graph model 2, and the calculation graph model 3:
computational graph model 1 [ op1-op2, op2-op3]
Calculation graph model 2 [ op1-op2]
Calculation graph model 3 [ op1-op2, op1-op3, op2-op4, op3-op4]
The distance vector of the calculation map model 1 is [1,0,1], the distance vector of the calculation map model 2 is [1], the distance vector of the calculation map model 3 is [1,1,0,0,1,1], and the calculation is performed using the edit distance (eidtdstance), so that the edit distance between the calculation map model 1 and the calculation map model 2 is 1, the edit distance between the calculation map model 1 and the calculation map model 3 is 3, and the edit distance between the calculation map model 2 and the calculation map model 3 is 4. Because the preset new model structure similarity configuration threshold needs the editing distance not less than 1, and the editing distance between all known models is not less than 1, the generated model is known to have no model or model with similar performance which is repeated with each other, and no repeated or similar model which needs to be removed.
As described in step (4), in this embodiment, the performance of the computation graph is measured by the accuracy of the computation graph model. The accuracy of each of the computation graph model 1, the computation graph model 2, and the computation graph model 3 is calculated (the accuracy is hereinafter denoted by P), and in the present embodiment: calculating P of graph model 11Calculate P for graph model 2, 902Calculate P of graph model 3 as 1753＝720。
Computational complexity (in the following, N is used to represent complexity), in this embodiment: calculating N of graph model 11N of graph model 2 is calculated as 32Calculate N of graph model 3 as 23＝4。
The fitness of each computational graph model is calculated (hereinafter, the fitness is denoted by F), and a formula F +10 × N is used (other formulas may be used for the fitness formula), in this embodiment: calculating F of graph model 11＝120，Calculating F of graph model 22Calculate F of graph model 3 as 1953＝760。
Removing the invalid model, and determining that the model is an invalid model when the fitness exceeds 800 according to the preset value of the embodiment and no model is an invalid model; the remaining models serve as alternative models for the computation graph model 1, the computation graph model 2 and the computation graph model 3 and are reserved as next generation seeds.
And (5) as described in the step (6), the minimum F in the computational graph model 1 is used as an optimal model and reserved as a new generation computational graph model.
As described in step (7), a computation graph model a is generated by using a genetic operator, and the model expression vector is as follows:
[op1-op2，op1-op3，op2-op3]
as described in the step (8), the distance vector of the model a is [1,1,1], and the edit distance between the model a and the calculation map model 1 is 1; calculating to obtain the edit distance between the model a and the calculation chart model 2 as 2; calculating the distance vector of the graph model 3 to be [1,1,0,0,1,1], and calculating the editing distance between the model a and the graph model 3 to be 3; the edit distance between all the known models and the model a is not less than 1, so that the model a and the existing models are not repeated or have dissimilar performances, and are new models.
As described in step (9), the model generated in step (8) is a new model, and the computation graph model a is saved as a new generation computation graph model 4.
As shown in step (10), since the number of the preset population of models in each generation is 3, and only one model, namely, the computational graph model 4, is currently available, the preset condition is not satisfied, and the step (7) should be returned to continue generating the computational graph model.
And (5) after repeating the steps (7) - (9), obtaining a new generation of computational graph models 5 meeting the conditions, and if the population quantity of the computational graph models 6 reaches 3, meeting the preset data, and entering the next step.
Calculating graph model 4 accuracy P as described in step (11)4225 complexity N43; calculating graph model 5 accuracy P56, complexity N56,; calculating graph model 6 accuracy P6290, complexity N6＝5；
As described in step (12), the currently retained computational graph model does not exceed three generations. Known calculation of P of graph model 41＝90， N1＝3，F1120. Calculating the fitness of the computation graph model 4, the computation graph model 5, and the computation graph model 6 (the fitness in this embodiment is calculated by using the formula F + 10N.) -which is generated in the step (9), the accuracy P of the computation graph model 44225 complexity N4With a fitness of F ═ 34255; calculating the accuracy P of the graph model 556, complexity N5With a fitness of F ═ 6566; calculating the accuracy P of the graph model 66290, complexity N6Fitness of F ═ 56＝340。
As described in step (13), since all newly generated calculated graph models have fitness greater than 40 but are already the second generation, the evolution stopping condition is satisfied, and the model generation is not continued.
As described in step (14), "scoring according to model complexity and performance": in step (1) of this embodiment, it is preset that the smaller the fitness, the better the model performance, and the formula F +10 × N is used for calculating the fitness in this embodiment. Through comparison, the fitness of the computational graph model 5 is the minimum, and the computational graph model 5 is the optimal model. And the automatic generation of the model is finished.
Example two:
as described in step (1), a user prepares numerical calculation data, namely platform production parameters (in a csv format or a picture format), and the data comprises a tag column; defining the algebra of each generation as G, and setting the maximum evolution generation G of the modelmaxThe number of model populations of each generation is 5 as 3 generations; the smaller the preset fitness, the better the model performance; presetting a dynamic new model similarity threshold value, wherein the similarity score is not more than G x 0.1+0.5 (the similarity score is not more than 0.8, namely the similarity is not more than 80 percent) required by a configuration threshold; calculating a fitness threshold value of the graph model, namely if the fitness of the optimal model is less than 100, considering that an evolution end condition is met, stopping calculation, and when the evolution algebra is more than or equal to 3 generations, considering that the fitness meets the requirement even if the fitness does not meet the numerical valueChanging the conditions, and stopping calculation; (ii) a A model with a preset fitness exceeding 1000 is considered as an ineffective model. The fitness in this embodiment is calculated by using the formula F +10 × N.
Using a known dataset (containing 2000 different computational graph structures, with a total of 10 different genes to construct a computational graph model), 2000 computational graph expression matrices were generated from 2000 different computational graph models. The second judging method according to claim 9, wherein the structure of the computation graph model is expressed by a computation graph expression matrix as follows: for example, in one of the computational graph models of this embodiment, there are 5 nodes, so the computational graph model uses a computational graph expression matrix of 5 × 5 (i.e., M × M) to represent the connection relationship of the computational graph, and the value range of the computational graph expression matrix is 1 to 10 (i.e., 1 to N). And simultaneously, combining similarity scores between data, training a 4-layer neural network (1 graph embedding layer, 3 graph convolution, 2 convolution layers and a full-connection layer) for calculating the similarity network between the two graphs. The neural network is trained to converge steadily.
As described in step (2), randomly generating the first generation of 5 models by using a genetic random operator, wherein the first generation of 5 models respectively comprises: randomly generating 5 first-generation models, which are respectively: a computation graph model 1, a computation graph model 2, a computation graph model 3, a computation graph model 4, and a computation graph model 5.
And (4) as described in the step (3), if the models which are repeated mutually or the models with similar performance exist, one of the models is reserved, and the rest repeated models and similar models are removed.
Encoding model expression vectors of the computation graph model 1, the computation graph model 2, the computation graph model 3, the computation graph model 4, and the computation graph model 5:
computational graph model 1 [ op1-op2, op2-op3, op3-op4]
Calculation graph model 2 [ op1-op2, op1-op2, op1-op3, …, op4-op5]
Calculation graph model 3 [ op1-op2, op1-op3, op2-op4, …, op6-op7]
Calculation graph model 4 [ op1-op2, op1-op2, op2-op3]
Calculation graph model 5 [ op1-op2, op1-op3, op2-op3, …, op3-op4]
Judging whether all new calculation map models generated in the last step have repeated calculation map models or performance-similar calculation map models, regarding the first generation G being 1, according to a model similarity threshold value G0.1 +0.5 and the similarity threshold of the first generation new models being 0.6 (namely, the similarity score being 0.6), respectively taking the calculation map model 1 and the calculation map models except the calculation map model 1 in the step (2) as input, and obtaining the predicted similarity score through the neural network calculation, wherein the similarity scores of the calculation map model 1 and the calculation map model 2, the calculation map model 3, the calculation map model 4 and the calculation map model 5 are 0.35, 0.57, 0.25 and 0.47, and if all the similarity scores are found to be less than 0.6, the models which are not repeated and have performance-similar models with the model 1 are not available; and (3) similarly, calculating a similarity score between the calculation graph model 2, the calculation graph model 3, the calculation graph model 4, the calculation graph model 5 and all other models in the step (2), and finding that the similarity score between all models is less than 0.6, so that the calculation graph model 1, the calculation graph model 2, the calculation graph model 3, the calculation graph model 4 and the calculation graph model 5 have no repeated or structurally similar models and have no models which need to be removed.
As described in step (4), in this embodiment, the performance of the computation graph is measured by the accuracy of the computation graph model. The accuracy of each of the computation graph model 1, the computation graph model 2, the computation graph model 3, the computation graph model 4, and the computation graph model 5 is calculated (the accuracy is hereinafter denoted by P), and in this embodiment: calculating P of graph model 11Calculating P of graph model 2 as 1242Calculate P for graph model 3 at 763Calculate P for graph model 4 as 5314P of graph model 5 is calculated 2215＝69。
Computational complexity (in the following, N is used to represent complexity), in this embodiment: calculating N of graph model 11Calculate N of graph model 2 as 42Calculate N of graph model 3 as 53Calculate N of graph model 4 as 74N of graph model 5 is calculated as 35＝4。
The fitness of each computation graph model is calculated (hereinafter, the fitness is denoted by F), and the present embodiment uses a formula F +10 × N (fitness metric)Other formulas can be adopted for the formula), in this embodiment: calculating F of graph model 11Calculating F of graph model 2 as 1642Calculating F of graph model 3, 1263F of graph model 4 is calculated 6014251, F of the graph model 5 is calculated5＝109。
As described in step (5), according to the preset of the embodiment, when the fitness exceeds 1000, the model is regarded as an invalid model, so all models are valid models; the computation graph model 1, the computation graph model 2, the computation graph model 3, the computation graph model 4 and the computation graph model 5 are used as alternative models and reserved as next generation seeds.
And (5) as described in the step (6), calculating the minimum F in the graph model 5, and keeping the minimum F as an optimal model for a new generation of the graph model.
As described in step (7), for the second generation, a computational graph model a is generated using a genetic random operator, and the model expression vector is:
[op1-op2，op1-op3，op2-op4，…，op6-op7]
as described in step (8), for the second generation G ═ 2, according to the model similarity threshold G × 0.1+0.5 and the similarity threshold of the first generation new model is 0.7, pairwise pairing the newly generated computation graph expression matrix of the computation graph model a and the already generated computation graph expression matrix of each model as inputs, and obtaining a similarity score through the aforementioned neural network computation: the similarity score of the model a and the calculation graph model 1 is 0.51, the similarity score of the model a and the calculation graph model 2 is 0.33, the similarity score of the model a and the calculation graph model 3 is 0.85, the similarity score of the model a and the calculation graph model 4 is 0.23, and the similarity score of the model a and the calculation graph model 5 is 0.66, wherein the similarity score of the model a and the calculation graph model 3 is greater than 0.7, so the model a is the same as or similar in performance to an existing model and is not a new model. Therefore, the model is regenerated by returning to the step (7).
Generating a computational graph model a by using a genetic random operator as described in the step (7), wherein the model expression vector is as follows:
[op1-op2，op1-op3，op2-op4，…，op11-op12]
as described in step (8), pairwise pairing the newly generated computation graph expression matrix of the computation graph model b and the already generated computation graph expression matrix of each model as inputs, and obtaining a similarity score through the neural network calculation: the similarity score of the model b and the calculation graph model 1 is 0.11, the similarity score of the model b and the calculation graph model 2 is 0.49, the similarity score of the model b and the calculation graph model 3 is 0.50, the similarity score of the model b and the calculation graph model 4 is 0.38, the similarity score of the model b and the calculation graph model 5 is 0.32, and the similarity score of the model b and all known models is not more than 0.7, so the model b and the existing model are not the same or similar in performance and are new models.
As described in step (9), the generated model is a new model, and the model b is saved as the new calculation graph model 6.
As shown in step (10), since the number of the preset population of each generation of models is 5, and the number of the current new models does not reach 5, the preset condition is not satisfied, and the step (7) should be returned to continue generating the computation graph model.
And (5) after repeating the steps (7) - (9) to obtain a new generation of calculation graph model 7, a calculation graph model 8, a calculation graph model 9 and a calculation graph model 10 which meet the conditions, and the population number reaches 5, the preset data of the population number is met, and the next step is carried out.
As described in step (10), the currently retained computational graph model does not exceed three generations. Known calculation of P of graph model 55＝69， N5＝4，F5106. Calculating the fitness of the computation graph model 6, the computation graph model 7, the computation graph model 8, the computation graph model 9 and the computation graph model 10 generated in the step (9), namely the accuracy P of the computation graph model 66631, complexity N612 with fitness F6751; calculating accuracy P of graph model 77324, complexity N712 with fitness F7444; calculating accuracy P of graph model 88Complexity N ═ 76812 with fitness F8196 parts by weight; calculating accuracy P of graph model 9935, complexity N915 with fitness F9185 (r); computational graph modelAccuracy P of 101012, complexity N10With a fitness of F, 810＝92。
As described in step (11), the fitness F of the graph model 10 is calculated1092, the evolution condition with fitness less than 100 is satisfied.
As described in step (12), "scoring according to model complexity and performance": in step (1) of this embodiment, it is preset that the smaller the fitness, the better the model performance, and the formula F +10 × N is used for calculating the fitness in this embodiment. By comparison, the fitness of the computational graph model 10 is minimal, and the computational graph model 10 is an optimal model. And the automatic generation of the model is finished.
Although the invention has been described and illustrated herein with reference to a particular arrangement or configurations, it is not intended to be limited to the details shown, since various modifications and structural changes may be made within the scope and spirit of the claims.
The parts involved in the invention are the same as or can be implemented by the prior art.
Claims (9)
1. A calculation graph evolution AI model automatic generation method based on structure similarity is characterized by comprising the following steps:
step (1): according to data preset by a user, data preparation is carried out, production parameters of a model design platform are set, a new model configuration threshold is set, and automatic model design is started;
step (2): generating a first generation computational graph model by utilizing a genetic algorithm operator;
and (3): judging whether all new calculation map models generated in the previous step are repeated or not or are calculation map models with approximate performance, wherein the judgment method for judging whether the new calculation map models in the step (3) are the calculation map models with the approximate performance comprises the following steps: judging the structural similarity of the models; if the model is judged to have the models which are mutually repeated or the models with similar performance, one of the models is reserved, and the rest repeated models and the similar models are removed;
and (4): according to the model retained in step (3):
a. calculating the performance of each computational graph model;
b. calculating the fitness of each calculation graph model according to the performance and the complexity of the calculation graph;
and (5): removing the useless model according to the model fitness, taking the rest models as alternative models, and reserving the alternative models as next generation seeds;
and (6): picking out a plurality of optimal models according to the next generation of seeds reserved in the step (5);
and (7): generating a new calculation chart model by using a genetic algorithm operator according to the alternative model selected in the step (6) and used as the next generation of seeds;
and (8): judging whether the new model of the computational graph generated in the last step is the generated computational graph model or not, or whether the new model of the computational graph is the computational graph model with performance similar to the generated computational graph model, and if not, entering the step (9); if so, returning to the step (7); in the step (8), the method for determining whether the model is a model with performance similar to that of the generated computation graph model includes: and judging the structural similarity of the models.
And (9): saving the retained computational graph model of the step (8) as a new generation computational graph model;
step (10): judging whether the number of the new generation of calculation graph models in the step (9) meets the number of the preset population in the step (1), if yes, entering the next step; if not, returning to the step (7);
step (11): calculating to obtain the performance of all the new generation of calculation graph models stored in the last step;
step (12):
a. if the life cycle of the model exceeds the third generation, then carrying out super-parameter search for searching the optimal solution or the suboptimal solution close to the optimal solution, wherein the definition of the life cycle exceeding the third generation is the same as the definition of the generation number in the genetic algorithm, the model structure is calculated as the first generation from the first occurrence in the process, the model reserved after the super-parameter search enters the step (13), and the fitness of a new calculation map model is calculated according to the performance of the calculation map calculated after the super-parameter search and the complexity of the model;
b. calculating the performance of the model according to the structure of the calculation graph and the fitness of each calculation graph model according to the performance and the complexity of the calculation graph for the model with the life cycle not exceeding three generations, and then entering the step (13);
step (13): judging whether the new model of the calculation graph meets the evolution ending condition preset in the step (1), and if so, entering the step (14); if not, returning to the step (5);
step (14): and summarizing the evolution calculation results, carrying out comprehensive scoring according to the complexity and the accuracy of the model, and selecting the optimal model.
2. The method for automatically generating the computational graph evolution AI model based on the structural similarity as claimed in claim 1, wherein the data preset by the user in the step (1) comprises the statistical distribution of the data, the correlation coefficient between the data dimensions and/or the statistical correlation between each dimension of the data and the label.
3. The method for automatically generating the computational graph evolution AI model based on the structural similarity as claimed in claim 1, wherein the model design platform production parameters set in step (1) comprise computing resources, operation running time, operation targets and/or genetic algorithm parameters.
4. The method according to claim 3, wherein the operation target comprises a fitness threshold of the computation graph model: a fitness threshold that is considered to satisfy an evolution termination condition and a fitness threshold that is considered to be an invalid model are included.
5. The method according to claim 1, wherein the genetic algorithm operators in the steps (2) and (7) include random operators, crossover operators and/or mutation operators.
6. The method according to claim 5, wherein the random operator is selected from a group consisting of randomly selecting the number of neurons, randomly selecting the type of neurons, and/or randomly determining the connection relationship between neurons.
7. The method for automatically generating the computational graph evolution AI model based on the structural similarity as claimed in claim 1, wherein the complexity in the steps (4) and (12) refers to the complexity calculated according to the number of nodes and the number of edges of the computational graph.
8. The method according to claim 1, wherein the hyperparameters in the hyperparameter search in step (12) refer to control parameters of an intra-AI neural network, including a learning rate, i.e. parameters, and/or weight decay parameters.
9. The method according to claim 1, wherein the step (3) (8) of determining the structural similarity of the model comprises:
the first judgment method comprises the following steps:
a. presetting a new model configuration threshold, namely a distance threshold;
b. constructing a distance vector according to a connection mode of the calculation map model, and calculating the distance between the calculation map model to be judged and all the generated models according to the distance vector, wherein the distance can be an editing distance, a soil moving distance, a cosine distance or a Euclidean distance;
c. and if the distance value between the newly generated computational graph model and all the generated models is lower than the preset new model configuration threshold, judging the newly generated computational graph model as a model with the performance similar to that of the generated computational graph model.
The second judgment method comprises the following steps:
a. presetting a new model configuration threshold, namely a similarity threshold; preparing a neural network for judging the similarity in advance, wherein the establishment and training method of the neural network is as described in the steps b-e;
b. expressing the structure of the computation graph model by using a computation graph expression matrix: in a system of graph node types with N types (N is the number of gene types) of different genes, for a computation graph with the number of nodes being M (M is the number of nodes in the computation graph), expressing the connection relation of the computation graph by using a computation graph expression matrix of M, wherein the value in the computation graph expression matrix is the gene connection type, and the value range of the gene connection type is 1 to N;
c. preparing a group of known data sets, wherein the known data sets comprise a computation graph expression matrix and a similarity score, the computation graph expression matrix comprises node information (namely the number and the type of nodes) of a graph, edge connection information (edge connection condition, namely whether edges are connected or not), and the similarity score ranges from 0 to 1.0 (the similarity score is 0, namely completely dissimilar, and the similarity score is 1.0, namely 100%);
d. establishing a neural network, wherein the neural network comprises a graph embedding layer, a graph volume layer, a volume layer and a full connection layer, the output is a scalar, the network output value is defined as a similarity score, and the value range of the similarity score is 0.0-1.0 (the score is 0, namely completely dissimilar, and the score is 1.0, namely the similarity is 100%); the input of the neural network is a calculation graph expression matrix of two calculation graphs;
e. training the neural network of step d to stable convergence using the data set of step c, with the input being the computational graph expression matrix of step b;
f. and e, calculating the similarity between the newly generated computational graph model and all the generated models through the trained neural network in the step e, inputting the similarity into the newly generated computational graph model and a computational graph expression matrix of any generated model, and outputting the similarity of the two models, wherein the similarity is higher than a preset new model configuration threshold, and the newly generated computational graph model is judged to be a model with the performance similar to that of the generated computational graph model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911006315.6A CN110766072A (en) | 2019-10-22 | 2019-10-22 | Automatic generation method of computational graph evolution AI model based on structural similarity |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911006315.6A CN110766072A (en) | 2019-10-22 | 2019-10-22 | Automatic generation method of computational graph evolution AI model based on structural similarity |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110766072A true CN110766072A (en) | 2020-02-07 |
Family
ID=69332512
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911006315.6A Pending CN110766072A (en) | 2019-10-22 | 2019-10-22 | Automatic generation method of computational graph evolution AI model based on structural similarity |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110766072A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111394778A (en) * | 2020-03-13 | 2020-07-10 | 江西景旺精密电路有限公司 | Automatic generation and loading method and device for VCP (virtual video protocol) line production parameters |
| WO2022127603A1 (en) * | 2020-12-14 | 2022-06-23 | 华为技术有限公司 | Model processing method and related device |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1450350A1 (en) * | 2003-02-20 | 2004-08-25 | Sony International (Europe) GmbH | Method for Recognizing Speech with attributes |
| CN108764029A (en) * | 2018-04-16 | 2018-11-06 | 南京维拓科技股份有限公司 | Model intelligence inspection method |
| US20180357152A1 (en) * | 2016-01-27 | 2018-12-13 | Bonsai AI, Inc. | Visual aids for debugging |
| CN109359303A (en) * | 2018-12-10 | 2019-02-19 | 枣庄学院 | A kind of Word sense disambiguation method and system based on graph model |
| CN109784497A (en) * | 2019-01-15 | 2019-05-21 | 探智立方(北京)科技有限公司 | Based on the method for calculating the AI auto-building model that figure is evolved |
| CN110168495A (en) * | 2016-01-27 | 2019-08-23 | 伯尼塞艾公司 | It can be re-used, reconfigure and be reassembled as the housebroken artificial intelligence object of artificial intelligence model |
| JP2019144639A (en) * | 2018-02-16 | 2019-08-29 | 株式会社日立製作所 | Method for training model outputting vector indicating tag set corresponding to image |
| JP2019148940A (en) * | 2018-02-26 | 2019-09-05 | パナソニック株式会社 | Learning processing method, server device, and reflection detection system |
| CN110310345A (en) * | 2019-06-11 | 2019-10-08 | 同济大学 | A kind of image generating method generating confrontation network based on hidden cluster of dividing the work automatically |
-
2019
- 2019-10-22 CN CN201911006315.6A patent/CN110766072A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1450350A1 (en) * | 2003-02-20 | 2004-08-25 | Sony International (Europe) GmbH | Method for Recognizing Speech with attributes |
| US20180357152A1 (en) * | 2016-01-27 | 2018-12-13 | Bonsai AI, Inc. | Visual aids for debugging |
| CN110168495A (en) * | 2016-01-27 | 2019-08-23 | 伯尼塞艾公司 | It can be re-used, reconfigure and be reassembled as the housebroken artificial intelligence object of artificial intelligence model |
| JP2019144639A (en) * | 2018-02-16 | 2019-08-29 | 株式会社日立製作所 | Method for training model outputting vector indicating tag set corresponding to image |
| JP2019148940A (en) * | 2018-02-26 | 2019-09-05 | パナソニック株式会社 | Learning processing method, server device, and reflection detection system |
| CN108764029A (en) * | 2018-04-16 | 2018-11-06 | 南京维拓科技股份有限公司 | Model intelligence inspection method |
| CN109359303A (en) * | 2018-12-10 | 2019-02-19 | 枣庄学院 | A kind of Word sense disambiguation method and system based on graph model |
| CN109784497A (en) * | 2019-01-15 | 2019-05-21 | 探智立方(北京)科技有限公司 | Based on the method for calculating the AI auto-building model that figure is evolved |
| CN110310345A (en) * | 2019-06-11 | 2019-10-08 | 同济大学 | A kind of image generating method generating confrontation network based on hidden cluster of dividing the work automatically |
Non-Patent Citations (1)
| Title |
|---|
| 李春磊: "零件工序模型几何演变序列生成方法及应用" * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111394778A (en) * | 2020-03-13 | 2020-07-10 | 江西景旺精密电路有限公司 | Automatic generation and loading method and device for VCP (virtual video protocol) line production parameters |
| WO2022127603A1 (en) * | 2020-12-14 | 2022-06-23 | 华为技术有限公司 | Model processing method and related device |
| EP4246375A4 (en) * | 2020-12-14 | 2024-04-24 | Huawei Tech Co Ltd | Model processing method and related device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Jiang et al. | Efficient network architecture search via multiobjective particle swarm optimization based on decomposition | |
| CN108334949B (en) | Image classifier construction method based on optimized deep convolutional neural network structure fast evolution | |
| CN109945881B (en) | Mobile robot path planning method based on ant colony algorithm | |
| CN108734202A (en) | A kind of Fault Diagnosis for HV Circuit Breakers method based on improved BP | |
| Esmin et al. | Hybrid evolutionary algorithm based on PSO and GA mutation | |
| CN109784497B (en) | AI model automatic generation method based on computational graph evolution | |
| CN109214503B (en) | Power transmission and transformation project cost prediction method based on KPCA-LA-RBM | |
| CN114022693B (en) | Single-cell RNA-seq data clustering method based on double self-supervision | |
| CN110458187A (en) | A kind of malicious code family clustering method and system | |
| Bakhshi et al. | Fast automatic optimisation of CNN architectures for image classification using genetic algorithm | |
| CN106778826A (en) | Based on the hereditary Hybrid Clustering Algorithm with preferred Fuzzy C average of self adaptation cellular | |
| CN103810261A (en) | K-means clustering method based on quotient space theory | |
| CN107092812B (en) | Method for identifying key protein based on genetic algorithm in PPI network | |
| CN110766072A (en) | Automatic generation method of computational graph evolution AI model based on structural similarity | |
| CN108399105B (en) | Software and hardware partitioning method based on improved brainstorming algorithm | |
| CN115952424A (en) | Graph convolution neural network clustering method based on multi-view structure | |
| CN111353534B (en) | Graph data category prediction method based on adaptive fractional order gradient | |
| CN114091650A (en) | Searching method and application of deep convolutional neural network architecture | |
| CN112819192A (en) | Method for predicting short-term power load of RF _ GRU network based on swarm algorithm optimization | |
| CN110929851A (en) | AI model automatic generation method based on computational graph subgraph | |
| CN110852435A (en) | Neural evolution calculation model | |
| CN109840551B (en) | Method for optimizing random forest parameters for machine learning model training | |
| Ghoneimy et al. | A new hybrid clustering method of binary differential evolution and marine predators algorithm for multi-omics datasets | |
| CN110580252A (en) | Space object indexing and query method under multi-objective optimization | |
| CN109074348A (en) | For being iterated the equipment and alternative manner of cluster to input data set |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20220801 Address after: 100080 unit 1-43, 17th floor, block B, No. 3 Danling street, Haidian District, Beijing Applicant after: Laiye Technology (Beijing) Co.,Ltd. Address before: 247, 2 / F, building 1, Tiandi Linfeng, No.1, yongtaizhuang North Road, Haidian District, Beijing 100192 Applicant before: INTELLIGENCE QUBIC (BEIJING) TECHNOLOGY Co.,Ltd. |
|
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200207 |