WO2022176196A1 - Dispositif et procédé d'apprentissage, et programme - Google Patents

Dispositif et procédé d'apprentissage, et programme Download PDF

Info

Publication number
WO2022176196A1
WO2022176196A1 PCT/JP2021/006627 JP2021006627W WO2022176196A1 WO 2022176196 A1 WO2022176196 A1 WO 2022176196A1 JP 2021006627 W JP2021006627 W JP 2021006627W WO 2022176196 A1 WO2022176196 A1 WO 2022176196A1
Authority
WO
WIPO (PCT)
Prior art keywords
latent
learning
input
variables
output data
Prior art date
Application number
PCT/JP2021/006627
Other languages
English (en)
Japanese (ja)
Inventor
雅人 宮原
大祐 佐藤
匡人 福田
成宗 松村
嘉樹 西川
Original Assignee
日本電信電話株式会社
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 日本電信電話株式会社 filed Critical 日本電信電話株式会社
Priority to PCT/JP2021/006627 priority Critical patent/WO2022176196A1/fr
Publication of WO2022176196A1 publication Critical patent/WO2022176196A1/fr

Links

Images

Classifications

    • 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

Definitions

  • One aspect of the present invention relates to a learning device, a learning method, and a program.
  • NP Neural Process
  • NP is known as one of the probabilistic function approximation methods based on deep generative models.
  • NP is a technique of learning probabilistic representations of original functions as latent variables from known input/output data, and estimating unknown outputs based on the learning results. Techniques of this kind can be applied, for example, to imputing missing values for high-dimensional data. Examples of applications include reconstruction of partial images and parameter optimization of an objective function whose shape is difficult to evaluate.
  • NP assumes that the latent variables are distributed according to a simple Gaussian distribution. For this reason, it has been difficult to apply NP to data whose latent variables follow functions expressed by complex (non-Gaussian) distributions.
  • the present invention has been made in view of the above circumstances, and aims to provide a technique that enables effective imputation of missing values for data whose latent variables follow functions represented by complex distributions. It is.
  • a learning device learns a probabilistic expression of an object function from input/output data as a latent variable.
  • This learning device includes a latent expression estimating unit, a latent expression aggregating unit, a first layer latent variable estimating unit, and an upper layer latent variable estimating unit.
  • the latent expression estimator stratifies the latent variables and applies input/output data to the neural network to estimate latent expressions expressing the latent variables.
  • the latent expression aggregation unit aggregates the estimated latent expressions for each input/output data.
  • the first layer latent variable estimating unit uses the aggregated latent expressions to obtain latent variable distribution parameters for the first layer of the layered latent variables.
  • the upper layer latent variable estimator obtains latent variable distribution parameters for layers higher than the first layer.
  • FIG. 1 is a diagram for briefly explaining missing value imputation based on function approximation.
  • FIG. 2 is a block diagram showing an example of a missing value imputation device according to one embodiment of the present invention.
  • FIG. 3 is a schematic diagram for explaining the first layer latent variable estimation unit 24 and the upper layer latent variable estimation unit 25.
  • FIG. 4 is a schematic diagram for explaining the estimation device 3.
  • FIG. 5 is a flow chart of the learning phase in one embodiment of the invention.
  • FIG. 6 is a flowchart of the estimation phase in one embodiment of the invention.
  • FIG. 7 is a block diagram showing an example of a computer that executes the program of the learning device.
  • FIG. 1 is a diagram for briefly explaining missing value imputation based on function approximation.
  • the Neural Process (NP) will be described with reference to FIG.
  • NP is a probabilistic function approximation technique based on deep generative models, which allows estimation taking into account the uncertainty of prediction.
  • NP obtains a probabilistic representation z of the function f by learning based on a known input-output data set according to some function f whose shape is unknown (Fig. 1(b) ). This processing corresponds, so to speak, to approximation of a function. An unknown output is then predicted based on the obtained representation z (FIG. 1(c)).
  • NP includes a learner that obtains from input/output data C a probabilistic representation z of the function of interest, and an estimator that makes predictions of y t given z and new inputs x t .
  • the learning device is also called the encoder and the estimating device is also called the decoder.
  • the names of the learning device and the estimation device are unified and explained.
  • the NP learning device has the function of executing the following three steps.
  • Step A1 Obtain a latent expression rc corresponding to each learning data.
  • Step A2 Aggregate latent expressions.
  • Step A3) Obtain the distribution parameter of the latent variable z.
  • step A1 the learning device obtains a latent expression r i from each learning input data set x i and y i using a neural network (NN).
  • NN neural network
  • the processing in which symmetry is maintained is processing in which the finally obtained aggregation r * of latent representations does not change even if the order of r c is changed.
  • the learning device uses a neural network to obtain distribution parameters ⁇ z and ⁇ z of the desired latent variable z.
  • Any neural network may be used as the neural network used in steps A1 and A3.
  • MLP multilayer perceptron
  • CNN convolutional neural network
  • Step B1 Obtain a latent variable z.
  • Step B2 Obtain the desired output distribution parameters.
  • Step B3 A desired output is obtained.
  • the estimation device obtains the latent variable z from the Gaussian distribution parameters ⁇ z and ⁇ z obtained by the learning device.
  • the estimator inputs the latent variable z and the desired input xt to the neural network, and obtains the distribution parameters ⁇ y and ⁇ y of the output y.
  • step B3 the estimator uses reparameterization tricks as in step B1 to obtain the desired output yt.
  • NP estimation device ⁇ About the NP estimation device>
  • the learning method of NP will be explained.
  • the right-hand side of the variational lower bound equation (1) is the objective function to be maximized.
  • a representative algorithm such as the stochastic gradient descent method can be applied to update equation (1), but it is of course not limited to this method.
  • the first step is to improve the expressiveness of the function by layering the latent variables.
  • the second stage is a contrivance to enable learning to proceed effectively even for the latent variable z in the higher hierarchy.
  • the latent variable z is hierarchized by preparing and superimposing a latent variable as a prior distribution at a higher level. This makes it possible to express a function with a latent variable having a complex non-Gaussian distribution, which is difficult to handle with the original NP.
  • the implementation changes from the neural network include adding neural network paths corresponding to the learning device and the estimating device, and changing the lower limit of variation, which is the objective function.
  • the inventors devised the following (1) to (3) in order to effectively advance the learning of all the latent variables of the hierarchical structure. That is, (1) introducing a skip path, (2) preparing dedicated latent variables for both the learning device and the estimating device, and (3) introducing a self-attention mechanism.
  • the neural network that estimates the latent variables in the upper layers only the latent variables in the lower layers are input.
  • the latent variables of all layers are added as inputs to the prediction neural network.
  • Existing NPs take a single latent variable z and a desired input x t as inputs.
  • the embodiment also adds information from all latent variables as inputs. As a result, the learning results of all latent variables are also taken into consideration during prediction.
  • Self-attention outputs an appropriately weighted latent representation r i when different input and output data (x i , y i ) are similar.
  • similar (x i , y i ) are thinned out by reducing the weight, and if the learning data exists in a specific area, it is possible to prevent biased learning in that area. .
  • FIG. 2 is a block diagram showing an example of a missing value imputation device according to one embodiment of the present invention.
  • the missing value complementing device 1 shown in FIG. 2 functions as a data complementing device that complements missing data.
  • the missing value imputation device 1 includes a learning device 2 , an estimation device 3 , a recording device 4 and an evaluation device 5 .
  • the learning device 2 learns a probabilistic expression of the target function as a latent variable from the input/output data.
  • the estimating device 3 estimates an unknown output based on the learning result.
  • the learning device 2 includes a preprocessing unit 21 , a latent expression estimating unit 22 , a latent expression aggregating unit 23 , a first layer latent variable estimating unit 24 , and an upper layer latent variable estimating unit 25 .
  • the preprocessing unit 21 acquires learning input/output data and generates a data set suitable for learning according to a predetermined preprocessing method.
  • the latent expression estimator 22 hierarchizes the latent variables, applies the given input/output data to a neural network, and estimates latent expressions expressing the latent variables.
  • the latent expression aggregation unit 23 aggregates the estimated latent expressions for each input/output data.
  • the first layer latent variable estimator 24 uses the aggregated latent expressions to obtain latent variable distribution parameters for the first layer of the layered latent variables.
  • the upper layer latent variable estimator 25 calculates latent variable distribution parameters for layers higher than the first layer. The obtained latent variable distribution parameters are recorded in the latent variable parameter recording unit 42 of the recording device 4 .
  • the estimation device 3 includes a preprocessing unit 31 , a latent variable generation unit 32 and an output estimation unit 33 .
  • the preprocessing unit 31 acquires a function to be estimated and generates a data set suitable for estimation according to a predetermined preprocessing method.
  • the latent variable generation unit 32 acquires the latent variable distribution parameters generated by learning from the latent variable parameter recording unit 42 of the recording device 4 .
  • the output estimator 33 estimates the output related to the function to be estimated based on the acquired distribution parameter of the latent variable.
  • the estimated output is provided to evaluation device 5 .
  • the evaluation device 5 evaluates the output estimation result and records the result in the evaluation value recording section 41 of the recording device 4 .
  • FIG. 3 is a schematic diagram for explaining the first layer latent variable estimation unit 24 and the upper layer latent variable estimation unit 25.
  • the first layer latent variable estimator 24 and the upper layer latent variable estimator 25 learn about the latent representation of the function (FIG. 3(A)). At that time, the introduction of skip paths and bi-directional latent variables enables effective learning of latent variables in the higher layers (FIG. 3(B))).
  • FIG. 3 shows an example in the case of three layers.
  • FIG. 4 is a schematic diagram for explaining the estimation device 3.
  • FIG. The estimating device 3 predicts an unknown output using the hierarchized NPs and the input to be complemented.
  • FIG. 5 is a flow chart of the learning phase in one embodiment of the invention.
  • the input of the function f of unknown shape to be estimated is defined as x
  • the output is defined as y.
  • Step S12 the learning device 2 obtains latent expressions (step S2). That is, the latent expression estimator 22 obtains the latent expression r i using an arbitrary neural network for each input/output data x i and y i .
  • step S13 the learning device 2 obtains parameters of latent variables. That is, the first layer latent variable estimator 24 uses an arbitrary neural network to obtain the distribution parameters ⁇ z1 and ⁇ z1 of the latent variable z 1 with the aggregate r * as an input. That is, the output of the first layer latent variable estimator 24 is the distribution parameters ⁇ z1 and ⁇ z1 of the latent variable z 1 in the first layer. The result is passed to the upper layer latent variable estimator 25 .
  • the upper layer latent variable estimation unit 25 estimates the distribution parameters ⁇ z2 , ⁇ z2 , . . . , ⁇ zL , ⁇ zL of the latent variables z 2 , .
  • L is the highest hierarchy.
  • To estimate z i for each stratum i (i ⁇ 2) we take as input the distribution parameters ⁇ zi ⁇ 1 , ⁇ zi ⁇ 1 of the latent variable z i ⁇ 1 and the deterministic Two of the outputs d i are used.
  • the skip path is another path from the aggregation r * to each latent variable constructed using a neural network
  • the deterministic output d i is the It is a feature vector created by preparing a neural network of paths of .
  • ⁇ zi and ⁇ zi of each layer are estimated using a neural network that receives the deterministic output d i ⁇ 1 and ⁇ zi ⁇ 1 and ⁇ zi ⁇ 1 as inputs.
  • the latent variable distribution parameters obtained in the above process are stored in the latent variable parameter recording unit 42 of the recording device 4 and updated as necessary.
  • step S14 the estimation device 3 obtains a predicted value (step S4). That is, the latent variable distribution parameter stored in the latent variable parameter recording unit 42 and the input xc used for prediction for learning are input to the estimating device 3, and the estimating device 3 predicts yc .
  • step S15 the evaluation device 5 performs evaluation based on the obtained estimation result. That is, the evaluation device 5 evaluates the evaluation prediction results based on the latent variable parameters and the prediction results of the estimation device 3 .
  • the evaluation value can be obtained by calculating the variational lower limit expression of Equation (2).
  • equation (2) z is written in bold to indicate that it is a vector.
  • the vector representation of z is written as z( ⁇ ). The obtained evaluation value is recorded in the evaluation value recording unit 41 .
  • Step S16 the missing value imputation device 1 updates each parameter. That is, the weight parameters of the neural network of the learning device 2 and the weight parameters of the neural network of the estimation device 3 are updated.
  • Equation (2) is used as an objective function, and the objective function is maximized using an algorithm such as the stochastic gradient descent method.
  • step S12 a self-attention mechanism may be used in the process of the latent expression estimator 22.
  • FIG. 1 the distance on the function space between the input and output x i and y i is introduced, and it is possible to increase the prediction accuracy by weighting so as to thin out duplicate x i and y i .
  • FIG. 6 is a flowchart of the estimation phase in one embodiment of the invention.
  • the estimation device 3 obtains latent variables for each layer. That is, the latent variable generator 32 generates the latent variable zi and the output d i from the skip path for each layer i (i ⁇ 1). As inputs, the latent variable parameters ⁇ z1 , ⁇ z1 , . Z i in each layer is generated from ⁇ zi and ⁇ zi using distribution parameters of corresponding latent variables.
  • the output d i following the skip path is prepared using a neural network. Each of d i is estimated using a neural network with d i+1 and z i+1 as inputs.
  • step S22 the estimating device 3 obtains parameters of predicted values. That is, the estimating device 3 uses a neural network with input x t to be predicted, decisive output d 1 , and first layer latent variable z 1 as inputs, and outputs distribution parameter ⁇ yt , ⁇ yt .
  • step S23 the estimation device 3 outputs the predicted value. That is, the estimation device 3 uses the output distribution parameters ⁇ yt and ⁇ yt estimated by the output estimation unit 33 to calculate and output y.
  • the functions of the learning device 2 can be realized by installing a program on the computer.
  • a computer can be made to function as the learning device 2 by executing a program provided as package software or online software.
  • FIG. 7 is a block diagram showing an example of a computer that executes the program of the learning device 2.
  • the computer 10 includes a CPU (Central Processing Unit) 101, a bridge circuit 102, a memory 103, a GPU (Graphics Processing Unit) 105, a BIOS-ROM 107, a storage 109, a USB connector 110, and a communication interface 112. Prepare.
  • the storage 109 is a non-volatile storage medium (block device) such as an HDD (Hard Disk Drive) or SSD (Solid State Drive).
  • the storage 109 stores basic programs such as an OS (Operating System) 201 and device drivers, and a learning program 202 for realizing the functions of the learning device 2 .
  • OS Operating System
  • the memory 103 includes ROM (Read Only Memory) and RAM (Random Access Memory).
  • a CPU 101 is a processor that controls the computer 10 .
  • the CPU 101 executes a BIOS (basic input/output system) stored in the BIOS-ROM 107 .
  • the CPU 101 also loads the learning program 202 from the storage 109 to the memory 103 and executes it.
  • the bridge circuit 102 connects the CPU 101 and each device on the PCI (Peripheral Component Interconnect) bus.
  • the bridge circuit 102 also includes a controller and the like that controls the storage 109 . Furthermore, it also has a function of executing communication with the GPU 105 via the PCIe bus.
  • the GPU 105 sends video signals to the monitor 104 .
  • a USB connector 110 connects a USB device or the like.
  • the learning program 202 may be installed on the computer 10 via a USB device.
  • the communication interface 112 connects the computer 10 to a communication network such as LAN (Local Area Network) or WAN (Wide Area Network).
  • the learning program 202 and various data may be stored not only in the storage 109, but also in a removable storage medium, for example, and read by the CPU 101 from a disk drive or the like. Alternatively, it may be stored in another computer connected via a communication network and read by CPU 101 via communication interface 112 .
  • the latent expression r is obtained directly from the entire data, without performing aggregation r * of latent expressions or learning a probabilistic latent variable z based thereon.
  • the latent expressions of the higher layers are learned. is not implemented.
  • ⁇ About regression> Regression which is the first example, will be described by taking reconstruction of a partial image of a face image as an example.
  • the input x corresponds to the pixel position of the image, that is, the two-dimensional discrete values
  • the output y is the red, blue, and yellow pixel values at the pixel position specified by x, that is, Corresponds to a three-dimensional vector of discrete values.
  • a probabilistic latent expression z( ⁇ ) corresponding to the function of the face is obtained by learning from a large number of face images by the learning device 2 of the present invention.
  • x and y may be discrete values, continuous values, vectors or matrices of any dimension.
  • ⁇ Regarding parameter optimization> As a second example, we present the parameter optimization of a black-box function.
  • yt is predicted in the form of a Gaussian distribution N( ⁇ , ⁇ 2 ) represented by the mean ⁇ yt and the variance ⁇ yt , so the prediction uncertainty at the desired position x t is in the form of the variance ⁇ yt is obtained by By using this uncertainty, it is also possible to efficiently apply search using any parameter optimization algorithm such as Bayesian optimization.
  • This parameter optimization is not limited to the example of drug activity, but can be applied to any task corresponding to the search for x that maximizes the objective function f: x ⁇ y, such as robot learning.
  • x and y may be discrete values, continuous values, vectors or matrices of any dimension, and it is possible to flexibly deal with data formats.
  • Applications of function approximation considering uncertainty include parameter optimization in robot learning and reward function approximation in reinforcement learning.
  • the computer 10 in FIG. 7 is not limited to a so-called server computer, and may be a desktop or notebook personal computer.
  • Portable terminals such as smartphones and tablets are also included in the category of computers.
  • the present invention is not limited to the above-described embodiments as they are, and can be embodied by modifying the constituent elements without departing from the gist of the invention at the implementation stage.
  • various inventions can be formed by appropriate combinations of the plurality of constituent elements disclosed in the above embodiments. For example, some components may be omitted from all components shown in the embodiments.
  • constituent elements of different embodiments may be combined as appropriate.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

Un dispositif d'apprentissage selon un mode de réalisation de la présente invention apprend, en tant que variables latentes, une représentation probabiliste d'une fonction cible à partir de données d'entrée/de sortie. Ce dispositif d'apprentissage est pourvu d'une unité d'estimation de représentation latente, d'une unité d'agrégation de représentation latente, d'une unité d'estimation de variable latente de premier niveau et d'une unité d'estimation de variable latente de niveau supérieur. L'unité d'estimation de représentation latente organise des variables latentes en une hiérarchie et applique des données d'entrée/de sortie à un réseau de neurones artificiels pour estimer des représentations latentes représentant les variables latentes. L'unité d'agrégation de représentation latente agrège les représentations latentes estimées pour chaque ensemble de données d'entrée/de sortie. L'unité d'estimation de variable latente de premier niveau utilise les représentations latentes agrégées pour obtenir des paramètres de distribution de variables latentes de premier niveau parmi les variables latentes dans la hiérarchie. L'unité d'estimation de variable latente de niveau supérieur obtient des paramètres de distribution de variables latentes à un niveau supérieur au premier niveau.
PCT/JP2021/006627 2021-02-22 2021-02-22 Dispositif et procédé d'apprentissage, et programme WO2022176196A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/JP2021/006627 WO2022176196A1 (fr) 2021-02-22 2021-02-22 Dispositif et procédé d'apprentissage, et programme

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2021/006627 WO2022176196A1 (fr) 2021-02-22 2021-02-22 Dispositif et procédé d'apprentissage, et programme

Publications (1)

Publication Number Publication Date
WO2022176196A1 true WO2022176196A1 (fr) 2022-08-25

Family

ID=82930514

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2021/006627 WO2022176196A1 (fr) 2021-02-22 2021-02-22 Dispositif et procédé d'apprentissage, et programme

Country Status (1)

Country Link
WO (1) WO2022176196A1 (fr)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017094267A1 (fr) * 2015-12-01 2017-06-08 株式会社Ｐｒｅｆｅｒｒｅｄ Ｎｅｔｗｏｒｋｓ Système de détection d'anomalie, procédé de détection d'anomalie, programme de détection d'anomalie et procédé de génération de modèle appris
JP2019075108A (ja) * 2017-10-18 2019-05-16 富士通株式会社 情報処理方法及び装置、並びに情報検出方法及び装置

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2017094267A1 (fr) * 2015-12-01 2017-06-08 株式会社Ｐｒｅｆｅｒｒｅｄ Ｎｅｔｗｏｒｋｓ Système de détection d'anomalie, procédé de détection d'anomalie, programme de détection d'anomalie et procédé de génération de modèle appris
JP2019075108A (ja) * 2017-10-18 2019-05-16 富士通株式会社 情報処理方法及び装置、並びに情報検出方法及び装置

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MAALøE LARS, FRACCARO MARCO, LIéVIN VALENTIN, WINTHER OLE: "BIVA: A Very Deep Hierarchy of Latent Variables for Generative Modeling", 6 November 2019 (2019-11-06), XP055846375, Retrieved from the Internet <URL:https://arxiv.org/pdf/1902.02102.pdf> [retrieved on 20210930] *
MARTA GARNELO; DAN ROSENBAUM; CHRIS J. MADDISON; TIAGO RAMALHO; DAVID SAXTON; MURRAY SHANAHAN; YEE WHYE TEH; DANILO J. REZENDE; S.: "Conditional Neural Processes", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 4 July 2018 (2018-07-04), 201 Olin Library Cornell University Ithaca, NY 14853 , XP081110939 *

Similar Documents

Publication Publication Date Title
NL2030226B1 (en) Methods, systems, articles of manufacture and apparatus to optimize resources in edge networks
JP6771645B2 (ja) ドメイン分離ニューラルネットワーク
KR101869438B1 (ko) 딥 러닝을 이용하여 환자의 진단 이력으로부터 질병 예후를 예측하는 방법 및 시스템
CN111126574A (zh) 基于内镜图像对机器学习模型进行训练的方法、装置和存储介质
US11681914B2 (en) Determining multivariate time series data dependencies
US20190122097A1 (en) Data analysis apparatus, data analysis method, and data analysis program
US10860895B2 (en) Imagination-based agent neural networks
US20220405649A1 (en) Quantum machine learning model feature space generation
CN111985631B (zh) 信息处理设备、信息处理方法及计算机可读记录介质
Wang et al. A proactive decision support method based on deep reinforcement learning and state partition
CN116070557A (zh) 使用强化学习的数据路径电路设计
Wu et al. Emulating facial biomechanics using multivariate partial least squares surrogate models
Gurung et al. Decentralized quantum federated learning for metaverse: Analysis, design and implementation
Song et al. Personalized federated learning with server-side information
WO2022176196A1 (fr) Dispositif et procédé d&#39;apprentissage, et programme
Bedi et al. STOWP: A light-weight deep residual network integrated windowing strategy for storage workload prediction in cloud systems
KR102561799B1 (ko) 디바이스에서 딥러닝 모델의 레이턴시를 예측하는 방법 및 시스템
Lupo Pasini et al. Fast and accurate predictions of total energy for solid solution alloys with graph convolutional neural networks
CN110689117A (zh) 基于神经网络的信息处理方法和装置
US11681726B2 (en) System for generating specialized phenotypical embedding
CN116070257A (zh) 使用去中心化数据上的堆叠集成来训练异构模型的系统和方法
Sun An influence diagram based cloud service selection approach in dynamic cloud marketplaces
WO2020054402A1 (fr) Dispositif de traitement de réseau neuronal, programme informatique, procédé de fabrication de réseau neuronal, procédé de fabrication de données de réseau neuronal, dispositif d&#39;utilisation de réseau neuronal, et procédé de réduction d&#39;échelle de réseau neuronal
CN115511070A (zh) 模型训练方法、装置及可读存储介质
Kalina et al. Robust training of radial basis function neural networks

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21926626

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21926626

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: JP