WO2021095361A1 - 演算装置および学習済みモデル - Google Patents

演算装置および学習済みモデル Download PDF

Info

Publication number
WO2021095361A1
WO2021095361A1 PCT/JP2020/035254 JP2020035254W WO2021095361A1 WO 2021095361 A1 WO2021095361 A1 WO 2021095361A1 JP 2020035254 W JP2020035254 W JP 2020035254W WO 2021095361 A1 WO2021095361 A1 WO 2021095361A1
Authority
WO
WIPO (PCT)
Prior art keywords
unit
input
layer
neural network
data
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.)
Ceased
Application number
PCT/JP2020/035254
Other languages
English (en)
French (fr)
Japanese (ja)
Inventor
佑樹 前
敬文 金森
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Denso Corp
Tokyo Institute of Technology NUC
Original Assignee
Denso Corp
Tokyo Institute of Technology NUC
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 Denso Corp, Tokyo Institute of Technology NUC filed Critical Denso Corp
Priority to JP2021555926A priority Critical patent/JP7386462B2/ja
Publication of WO2021095361A1 publication Critical patent/WO2021095361A1/ja
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]

Definitions

  • the present disclosure relates to an arithmetic unit having a trained model of a neural network.
  • a neural network is a mathematical model created to express the mechanism of the brain on a computer. It is composed of an input layer, an intermediate layer, and an output layer, and each layer shows the strength of the connection between neurons. The weighting factor is specified. This weighting factor is determined by learning.
  • the weighting coefficient of the connection between neurons is deterministically determined by learning. For this reason, the inference result by the neural network is point estimation, and even if the result is incorrect, it may be output deterministically.
  • the Bayesian neural network is known as a neural network in which the inference result has probabilistic information.
  • the Bayesian neural network can obtain the inference result stochastically by having the weighting coefficient of the connection between neurons in the probability distribution.
  • the Bayesian neural network can obtain the prediction distribution by using Monte Carlo Dropout, which is one of the approximation methods of Bayesian deep learning.
  • Patent Document 1 a method called a model ensemble that integrates the inference results of a plurality of neural network models is also known (Patent Document 1). Combining the results of a plurality of neural network models is almost the same as obtaining the inference result stochastically, and this method can also improve the reliability of the inference result.
  • the Bayesian neural network described above has a much larger computational load than a normal neural network.
  • the Monte Carlo dropout requires tens to hundreds of samplings by the Monte Carlo method, which is extremely computationally intensive.
  • the method described in Patent Document 1 also calculates a plurality of neural networks, calculation time and memory capacity are required for the number of neural networks. Although such calculations can be performed for research purposes, they are difficult to apply to edge equipment that requires high-speed judgment, such as self-driving vehicles.
  • Non-Patent Document 1 As a method of obtaining the predicted distribution of the Bayesian neural network with one feed forward without performing multiple samplings, the function of each layer is approximated by the first-order Taylor expansion, and the expected value and variance of the function are analyzed. A method of propagation has been proposed (Non-Patent Document 1). If the function of each layer is approximated by the first-order Taylor expansion and the expected value of the function and the analytical solution of the variance are propagated, the predicted distribution regarded as the Gaussian distribution can be obtained at high speed with one feed forward.
  • the present disclosure provides an arithmetic unit provided with a neural network model capable of performing probabilistic inference with low memory and high speed.
  • the arithmetic unit of the present disclosure stores data and a storage unit that stores a trained neural network model, including a probability layer in which the weight coefficient is defined by a probability distribution and propagates the mean and variance of the probability distribution of output values to the subsequent stage. It includes an input unit for input, an inference unit that analyzes the data input in the input unit by inference by the neural network model, and an output unit that outputs the analysis result by the inference unit.
  • the trained model of the present disclosure is a trained model of a neural network for making a computer function to analyze input data, in which a weighting coefficient is defined by a probability distribution and the average of the probability distributions of output values. It is composed of a neural network model including a probability layer that propagates the variance to the subsequent stage, and is based on the average and variance of the probability distribution of the output values propagated to the output layer when the input data is applied to the neural network model. It has a structure to analyze the input data.
  • probabilistic inference can be performed at high speed with low memory.
  • FIG. 1 is a diagram showing a configuration of an arithmetic unit according to an embodiment.
  • FIG. 2 is a diagram showing an example of a neural network model.
  • FIG. 3 is a neural network model before conversion of the neural network model shown in FIG.
  • FIG. 4 is a diagram showing a calculation formula for propagating the expected value and variance in the processing of full coupling, dropout, ReLU, and sigmoid.
  • FIG. 5 is a diagram showing a calculation formula of the expected value and the variance of the dispersion mode and the upper bound mode when the learner is fully connected.
  • FIG. 6A is a diagram showing a configuration of a neural network model of the present embodiment.
  • FIG. 6B is a calculation formula of the output y of the neural network model.
  • FIG. 7A is a calculation formula for obtaining the expected value ⁇ and the variance ⁇ 2.
  • FIG. 7B is a diagram showing a calculation formula of the expected value and the variance when the two-class classification is performed using the expected value ⁇ and the variance ⁇ 2 which are the results of the regression analysis.
  • FIG. 8A is a diagram showing the configuration of a neural network model whose output is multidimensional.
  • FIG. 8B is a calculation formula of the output y of the neural network model whose output is multidimensional.
  • FIG. 9A is a calculation formula for obtaining the expected value ⁇ and the variance ⁇ 2.
  • FIG. 9B is a diagram showing a calculation formula of the expected value and the variance when multiclass classification is performed using the expected value ⁇ and the variance ⁇ 2 which are the results of regression analysis.
  • FIG. 10 is a diagram showing an application example of the arithmetic unit of the present embodiment.
  • FIG. 1 is a diagram showing a configuration of an arithmetic unit 1 according to the present embodiment.
  • the arithmetic unit 1 includes an input unit 10, an inference unit 11, a storage unit 12, and an output unit 13.
  • the input unit 10 has a function of accepting input of data to be analyzed.
  • the data to be analyzed includes, for example, image data, voice data, text data, and the like, but is not limited thereto.
  • the hardware of the input unit 10 is, for example, a communication interface that accepts data input from other sensors such as a camera and a microphone.
  • the storage unit 12 stores the trained neural network model.
  • the inference unit 11 reads the trained model stored in the storage unit 12, applies the data input from the input unit 10 to the trained model, and performs inference.
  • the hardware of the inference unit 11 is a device such as a CPU and RAM.
  • the arithmetic unit 1 of the present embodiment can be applied to both regression analysis and classification.
  • the output unit 13 has a function of outputting the analysis result by the inference unit 11.
  • the hardware of the output unit 13 is, for example, a communication interface for transmitting data related to the inference result to another application.
  • FIG. 2 is a diagram showing an example of a neural network model stored in the storage unit 12.
  • the upper part is the input layer L1, and from the upper part, the folding layer L2, the folding layer L3, the max pooling layer L4, the dropout layer L5, the Flatten layer L6, the fully bonded layer L7, the dropout layer L8, and the fully bonded layer L9.
  • the expected value and variance are output (L10).
  • dispersion is generated in the dropout layer L5, and then the expected value and dispersion are propagated.
  • the layer that propagates the variance is implemented by an interface of 2 inputs and 2 outputs. Therefore, the layer L v0 that generates the dispersion 0 is included so that the dropout layer L5 that generates the dispersion can receive the dispersion as an input.
  • the expected value ⁇ and the variance ⁇ 2 output from the previous layer are received as inputs, and the expected value ⁇ and the variance ⁇ are used by the calculation formula shown in the BNN column of FIG. Output 2
  • the neural network model shown in FIG. 2 is a transformation of the neural network model shown in FIG. 3 so that the expected value and the variance can be propagated.
  • FIG. 4 is a diagram showing a calculation formula that outputs a hat of ⁇ as an expected value and a square of a hat of ⁇ as a variance in the processing of full coupling, dropout, ReLU, and sigmoid.
  • x Layer input (normal NN layer input with 1 input and 1 output)
  • y Layer output (normal NN layer output with 1 input and 1 output)
  • ⁇ Expected value of x squared: variance of x ⁇ hat: expected value of y squared of ⁇ hat: variance of y
  • FIG. 4 the calculation formulas of a normal neural network (NN) with 1 input and 1 output are shown side by side.
  • an affine layer such as a fully connected layer
  • the input is regarded as a random variable
  • the analytical solution of the expected value and variance is calculated.
  • a non-linear layer such as dropout, ReLU, sigmoid, etc.
  • approximate Bayesian inference such as Gaussian approximation is performed.
  • the expected value and variance can be obtained in the same manner as above by performing approximate Bayesian inference.
  • the arithmetic unit of the present embodiment has two modes for propagating the variance.
  • the first mode is an independent variance propagation mode (hereinafter referred to as "independent mode") that calculates the output variance assuming that the input values are independent, and the second mode is input.
  • This is a mode of upper bound variance propagation (hereinafter referred to as "upper bound mode”) in which the variance to be output is calculated assuming that the values are non-independent.
  • FIG. 5 is a diagram showing the calculation formulas of the expected value and the variance of the independent mode and the upper bound mode when the learner is fully connected. Basically, the calculation is performed in the independent mode assuming that the input values are independent. Therefore, the formulas for calculating the expected value and variance in the independent mode are the same as the formulas shown in FIG. In the upper bound mode, the sum of the products of the variance and the weight is squared, so that the variance is larger than in the independent mode.
  • the independent variance propagation is performed when the independence assumption is incorrect, the variance will be underestimated and overconfident, so when the independence assumption is incorrect, the calculation in the upper bound mode will result in overconfidence. You can prevent it from becoming.
  • upper bound variance propagation is performed when the assumption of independence is correct, the variance will be overestimated and the confidence will be insufficient, but if the independence dispersion propagation is performed, the confidence will be prevented.
  • the independent mode and the upper bound mode have different characteristics.
  • the arithmetic unit may switch the mode of distributed propagation depending on the application.
  • the basic configuration of the arithmetic unit of the second embodiment is the same as that of the arithmetic unit 1 of the first embodiment (see FIG. 1).
  • the arithmetic unit of the second embodiment has a different configuration of the neural network model stored in the storage unit 12 from that of the first embodiment.
  • the calculation load is reduced by limiting the probability layer included in the neural network to only one layer.
  • FIG. 6A is a diagram showing the configuration of the neural network model stored in the storage unit 12.
  • the neural network model has an input layer, an intermediate layer, and an output layer.
  • three layers are provided as intermediate layers, but the number of intermediate layers is not limited to three, and may be four or more layers, or two or one layer.
  • the layer closer to the output layer among the three intermediate layers constitutes a stochastic layer having a weighting coefficient stochastically.
  • the probability layer defines the weighting coefficient stochastically by combining the Bernoulli distribution and the weighting coefficient.
  • Other layers of the neural network model define the weighting factors deterministically.
  • the inference unit 11 inputs the data to be analyzed input by the input unit 10 to the input layer, and performs the calculation according to the weighting coefficient of the neural network model.
  • the inference unit 11 of the present embodiment analyzes the input data by using the value appearing in the probability layer instead of the value appearing in the output layer.
  • FIG. 6B is a calculation formula of the output value y of the neural network model shown in FIG. 6A.
  • the value obtained by adding the weight w 0 to the sum of the values obtained by multiplying the value z k appearing in the probability layer by the d k determined from the Bernoulli distribution and the weight coefficient w k is used as the output of the neural network model.
  • the weight w 0 is a bias term meaning an intercept.
  • the Bernoulli distribution is given as a given probability distribution, but the given probability distribution is not limited to the Bernoulli distribution, and another probability distribution such as the Gaussian distribution may be used.
  • FIG. 7A is a diagram showing a calculation formula showing the expected value and the variance obtained by the inference unit 11.
  • the expected value ⁇ is a value obtained by adding a weight w 0 to the sum of the values obtained by multiplying the value z k appearing in the probability layer by the d k determined from the Bernoulli distribution and the weighting coefficient w k.
  • the inference unit 11 of the present embodiment obtains the variance ⁇ 2 obtained by using the probability distribution p as an index showing uncertainty.
  • FIG. 7A shows the result of regression analysis of the data to be analyzed
  • FIG. 7B shows the expectation when two-class classification is performed using the expected value ⁇ and the variance ⁇ 2 which are the results of regression analysis. It is a figure which shows the calculation formula of a value and a variance. The smaller the variance ⁇ 2, the higher the reliability of the expected value ⁇ . By performing the calculation using these calculation formulas, the inference unit 11 can perform the classification.
  • FIGS. 6A, 6B, 7A, and 7B described above the case where the output y is one-dimensional is described, but when the output y is multidimensional, FIGS. 8A, 8B, and 9A are shown. It looks like 9B. That is, FIG. 8A is a diagram showing the configuration of a neural network model whose output is multidimensional, and FIG. 8B is a calculation formula of the output y of the neural network model whose output is multidimensional.
  • FIG. 9A shows a calculation formula for obtaining the expected value ⁇ and the variance ⁇ 2 when the output is multidimensional, and FIG. 9B shows the multi-class classification using the expected value ⁇ and the variance ⁇ 2 which are the results of regression analysis. It is a figure which shows the calculation formula of an expected value and a variance.
  • the training of the neural network model is performed using the teacher data.
  • the error between the inference result obtained by inputting data to the input unit 10 and the correct label is obtained, and the weighting coefficient of the neural network model is updated by the error backpropagation method so that the error is minimized.
  • the probability distribution constituting the probability layer is given, it is not updated, but the probability layer is updated by updating the weighting coefficient. In this way, when learning is performed, learning is performed using the values that appear in the output layer.
  • the trained model described above is also included in the scope of the present disclosure.
  • the example in which the probability layer is arranged at the position closest to the output layer is given, but the arrangement of the probability layer does not have to be the position closest to the output layer.
  • the probability layer is one layer in the neural network model. By setting the probability layer to one layer in this way, the calculation shown in FIG. 7A becomes possible.
  • the probability layer is composed of a combination of a given probability distribution and a weighting coefficient has been given, but the probability distribution itself may be updated by learning.
  • FIG. 10 is a diagram showing an application example of the arithmetic unit of the first embodiment and the second embodiment.
  • the ECU 2 Engine Control Unit
  • the ECU 2 has the same configuration as the arithmetic unit 1 shown in FIG.
  • the ECU 2 receives the input of the image taken by the camera 20 of the autonomous driving vehicle, classifies the objects reflected in the image into classes (for example, trucks, passenger cars, motorcycles, people, etc.), and determines the reliability of the classified classes. Infer. Then, the ECU 2 transmits the inferred result to the vehicle control unit 21 of the autonomous driving vehicle.
  • the vehicle control unit 21 controls the vehicle based on the data transmitted from the ECU 2.
  • the arithmetic device is a device that detects an object based on the image taken by the camera mounted on the vehicle, the weight coefficient is defined by the probability distribution, and the average and dispersion of the probability distribution of the output vector are propagated to the subsequent stage.
  • a storage unit that stores a learned neural network model, including a probability layer to be used, an input unit that inputs an image taken by a camera of an autonomous driving vehicle, and an object reflected in the image in a class by inference by the neural network model. It includes an inference unit that classifies and infers its reliability, and an output unit that outputs the object class and reliability data inferred by the inference unit to the vehicle control unit.
  • the vehicle control unit controls the self-driving vehicle based on the data received from the arithmetic unit. Since the vehicle control unit also receives reliability data, for example, when the reliability of the inference result is low, it is possible to perform control such as slowing down the vehicle speed while performing control using the inference result.
  • the arithmetic device is a device that detects an object based on an image taken by a camera mounted on a vehicle.
  • the weight coefficient is defined by a probability distribution, and the average and dispersion of the probability distribution of the output vector are propagated to the subsequent stage.
  • the vehicle control unit controls the self-driving vehicle based on the data received from the arithmetic unit.
  • autonomous vehicles are safety-critical autonomous systems, it is useful to provide information on the reliability of data analysis results. Further, in the self-driving vehicle, in order to avoid an accident, it is necessary to instantly recognize the situation, so that it is not possible to perform a time-consuming process. Since the ECU 2 of the present embodiment can perform probabilistic inference with low memory and high speed, it can be suitably used for controlling an autonomous driving vehicle.
  • the arithmetic unit 1 of the present embodiment can be applied to other than the control of the autonomous driving vehicle.
  • the substance may be classified into two classes, benign / malignant, and the index of its reliability may be obtained, or the type of tumor may be classified into multiple classes and the index of its reliability may be obtained.
  • the arithmetic unit of the present disclosure can be applied to various situations.
  • the present disclosure is particularly useful when performing inference on memory-saving edge devices.

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)
  • Image Analysis (AREA)
PCT/JP2020/035254 2019-11-11 2020-09-17 演算装置および学習済みモデル Ceased WO2021095361A1 (ja)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP2021555926A JP7386462B2 (ja) 2019-11-11 2020-09-17 演算装置および学習済みモデル

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2019203988 2019-11-11
JP2019-203988 2019-11-11

Publications (1)

Publication Number Publication Date
WO2021095361A1 true WO2021095361A1 (ja) 2021-05-20

Family

ID=75912167

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2020/035254 Ceased WO2021095361A1 (ja) 2019-11-11 2020-09-17 演算装置および学習済みモデル

Country Status (2)

Country Link
JP (1) JP7386462B2 (https=)
WO (1) WO2021095361A1 (https=)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114648652A (zh) * 2022-03-14 2022-06-21 北京计算机技术及应用研究所 一种利用卷积神经网络和正交变换计算图像哈希值的方法
KR20230172428A (ko) 2022-06-15 2023-12-22 가부시키가이샤 야스카와덴키 데이터 분석 시스템, 데이터 분석 방법 및 프로그램

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018104563A2 (en) * 2016-12-09 2018-06-14 Tomtom Global Content B.V. Method and system for video-based positioning and mapping

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018104563A2 (en) * 2016-12-09 2018-06-14 Tomtom Global Content B.V. Method and system for video-based positioning and mapping

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
ANCONA MARCO, ÖZTIRELI CENGIZ, GROSS MARKUS: "Explaining deep neural networks with a polynomial time algorithm for Shapley values approximation, arXiv:1903.10992v4", ARXIV (CORNELL UNIVERSITY), 21 June 2019 (2019-06-21), pages 1 - 10, XP081162867, Retrieved from the Internet <URL:https://arxiv.org/abs/1903.10992v4> [retrieved on 20201020] *
INOUE, MASAAKI ET AL: "The neural network with q-normal distribution", PROCEEDINGS OF THE 2003 IEICE GENERAL CONFERENCE, 3 March 2003 (2003-03-03) *
WEN, H. L. ET AL.: "Probabilistic wind power forecasting via Bayesian deep learning based prediction intervals", PROCEEDINGS OF THE 2019 IEEE 17TH INTERNATIONAL CONFERENCE ON INDUSTRIAL INFORMATICS (INDIN, 25 July 2019 (2019-07-25), pages 1091 - 1096, XP033700784, ISBN: 978-1-7281-2927-3, DOI: 10.1109/INDIN41052.2019.8972125 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114648652A (zh) * 2022-03-14 2022-06-21 北京计算机技术及应用研究所 一种利用卷积神经网络和正交变换计算图像哈希值的方法
KR20230172428A (ko) 2022-06-15 2023-12-22 가부시키가이샤 야스카와덴키 데이터 분석 시스템, 데이터 분석 방법 및 프로그램

Also Published As

Publication number Publication date
JP7386462B2 (ja) 2023-11-27
JPWO2021095361A1 (https=) 2021-05-20

Similar Documents

Publication Publication Date Title
Arnelid et al. Recurrent conditional generative adversarial networks for autonomous driving sensor modelling
Jain et al. An automated hyperparameter tuned deep learning model enabled facial emotion recognition for autonomous vehicle drivers
Kutyniok The mathematics of artificial intelligence
US20180121802A1 (en) Method of converting neural network and recognition apparatus using the same
KR102306658B1 (ko) 이종 도메인 데이터 간의 변환을 수행하는 gan의 학습 방법 및 장치
CN112424788A (zh) 使用三维眼睛注视矢量的车辆乘员参与
Hu et al. A framework for probabilistic generic traffic scene prediction
US11494613B2 (en) Fusing output of artificial intelligence networks
US11961275B2 (en) Device and method for training a normalizing flow
US6601052B1 (en) Selective attention method using neural network
Asemi et al. Adaptive neuro-fuzzy inference system for evaluating dysarthric automatic speech recognition (ASR) systems: a case study on MVML-based ASR
US20230394304A1 (en) Method and Apparatus for Neural Network Based on Energy-Based Latent Variable Models
JP7295338B2 (ja) 証明可能ロバストな深層生成モデルを用いたシステム及び方法
EP4396730B1 (en) Automated variational inference using stochastic models with irregular beliefs
JP7386462B2 (ja) 演算装置および学習済みモデル
Zhong et al. Large language model based system with causal inference and Chain-of-Thoughts reasoning for traffic scene risk assessment
CN115661560B (zh) 驾驶舱中人脸检测方法、目标检测模型训练方法及装置
DE102024127224A1 (de) Föderiertes lernen mit basismodell-destillation
EP4607408A1 (en) Using diffusion model to generate graph data
Zhang et al. Network security prediction and situational assessment using neural network-based method
JP2008085472A (ja) 音源定位・同定装置
Zhu et al. Recent developments on statistical transfer learning
Zec et al. Recurrent conditional gans for time series sensor modelling
Rajeshwar Hybrid particle Swarm optimization-gravitational search algorithm based deep belief network: Speech emotion recognition: Hybrid PSO-GSA based DBN
Tarabin et al. Detecting distracted drivers using convolutional 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: 20887952

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2021555926

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20887952

Country of ref document: EP

Kind code of ref document: A1