WO2022249327A1 - 学習装置、学習方法及び学習プログラム - Google Patents
学習装置、学習方法及び学習プログラム Download PDFInfo
- Publication number
- WO2022249327A1 WO2022249327A1 PCT/JP2021/019982 JP2021019982W WO2022249327A1 WO 2022249327 A1 WO2022249327 A1 WO 2022249327A1 JP 2021019982 W JP2021019982 W JP 2021019982W WO 2022249327 A1 WO2022249327 A1 WO 2022249327A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- learning
- learning device
- matrix
- parameters
- deep learning
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims description 37
- 230000006870 function Effects 0.000 claims abstract description 46
- 238000013136 deep learning model Methods 0.000 claims abstract description 33
- 239000011159 matrix material Substances 0.000 claims description 61
- 238000004364 calculation method Methods 0.000 claims description 12
- 238000009826 distribution Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 description 29
- 238000013135 deep learning Methods 0.000 description 15
- 238000010586 diagram Methods 0.000 description 12
- 238000003860 storage Methods 0.000 description 9
- 239000013598 vector Substances 0.000 description 8
- 238000011156 evaluation Methods 0.000 description 5
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000009499 grossing Methods 0.000 description 4
- 230000010365 information processing Effects 0.000 description 4
- 238000005457 optimization Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 239000007787 solid Substances 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 238000012946 outsourcing Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 239000010454 slate Substances 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
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
- the present invention relates to a learning device, a learning method, and a learning program.
- deep learning and deep neural networks have achieved great success in areas such as image recognition and speech recognition.
- image recognition using deep learning when an image is input to a model that includes many nonlinear functions of deep learning, the result of identifying what the image shows is output.
- convolutional networks and ReLU are commonly used in image recognition.
- a deep neural network trained by deep learning may simply be referred to as a deep learning model or model.
- Non-Patent Documents 1 and 2 As a method of making deep learning robust against hostile attacks, adversarial learning that is added as data when learning adversarial attacks in advance has been proposed (see, for example, Non-Patent Documents 1 and 2).
- Entropy-SGD described in Non-Patent Document 4 smoothes the objective function, but the learning efficiency may not be sufficiently high.
- a learning device includes a calculation unit that calculates an objective function by Entropy-SGD when data created as a hostile attack is input to a deep learning model; an updating unit that updates the parameters of the deep learning model so that the function is optimized.
- deep learning models can be made robust against noise.
- FIG. 1 is a diagram illustrating the structure of the entire deep learning model.
- FIG. 2 is a diagram illustrating a configuration example of the learning device according to the first embodiment.
- FIG. 3 is a diagram for explaining the Entropy-SGD algorithm.
- FIG. 4 is a diagram for explaining the algorithm of the embodiment.
- FIG. 5 is a diagram explaining the algorithm of the embodiment.
- FIG. 6 is a flowchart showing the flow of deep learning.
- FIG. 7 is a flow chart showing the flow of learning using Entropy-SGD.
- FIG. 8 is a flowchart showing the flow of update processing in learning.
- FIG. 9 is a flowchart showing the flow of update processing according to the embodiment.
- FIG. 10 is a flowchart showing the flow of update processing according to the embodiment.
- FIG. 10 is a flowchart showing the flow of update processing according to the embodiment.
- FIG. 11 is a flow chart showing the flow of update processing in adversarial learning.
- FIG. 12 is a flow chart showing the flow of update processing according to an embodiment in adversarial learning.
- FIG. 13 is a flow chart showing the flow of update processing according to an embodiment in adversarial learning.
- FIG. 14 is a diagram of an example of a computer that executes a program;
- FIG. 1 is a diagram illustrating the structure of the entire deep learning model. In the following description, it is assumed that deep learning is performed by the learning device 10a.
- a deep learning model has an input layer that receives signals, one or more hidden layers that transform the signals from the input layers, and a final layer that transforms the signals from the hidden layers into outputs such as probabilities. .
- FIG. 6 is a flowchart showing the flow of deep learning. As shown in FIG. 6, first, the learning device 10a applies an input randomly selected from a prepared data set to the discriminator (step S101).
- the learning device 10a calculates the output of the classifier and uses it and the label of the dataset to calculate the loss function (step S102). Then, the learning device 10a updates the parameters of the discriminator using the gradient of the loss function (step S103). Note that the loss function is an example of the objective function.
- step S104 If the evaluation criteria are not satisfied (step S104, No), the learning device 10a returns to step S101 and repeats the process. On the other hand, if the evaluation criteria are satisfied (step S104, Yes), the learning device 10a ends the process.
- the learning device 10a updates the parameters so that the loss function becomes smaller. Since the loss function is usually set to a function that becomes smaller as the discriminator output and label match, the learning process enables the discriminator to discriminate the input label.
- the evaluation criteria in step S104 are, for example, whether or not a separately prepared data set can be correctly identified.
- image recognition by deep learning will be described here as an example, the embodiments can be applied to various identification tasks other than image recognition.
- C is the image channel (three channels in the case of RGB format)
- H is the vertical dimension
- W is the horizontal dimension.
- the deep learning model repeats the non-linear function and the linear operation and outputs an output through a function called softmax function in the final layer.
- softmax function a function called softmax function in the final layer.
- the output of formula (1) represents the score for each label in class classification, and the element of the output with the largest score for i obtained by formula (2) is the recognition result of deep learning.
- Image recognition is one of class classification, and the model f s (z ⁇ ( ⁇ )) that performs classification is called a discriminator.
- ⁇ is obtained by optimizing the average of the data as shown in formula (3).
- ⁇ is a parameter called the learning rate.
- Newton's method is a method of more efficiently performing optimization using gradients, which optimizes equation (5).
- B(x i ) is a region of distance ⁇ centered at x i
- x ′ i obtained by maximization is called a hostile attack.
- Entropy-SGD is a method for improving smoothness in deep learning. For the original loss function L(x, y, ⁇ ), Entropy-SGD minimizes the loss function of equation (7).
- Equation (7) is the local entropy of the probability density function p ⁇ ( ⁇ ′) shown in Equation (8).
- Ep ⁇ ( ⁇ ') [ ⁇ '] is the expected value of the probability density function p ⁇ ( ⁇ ').
- FIG. 3 is a diagram for explaining the Entropy-SGD algorithm.
- FIG. 7 is a flowchart showing the flow of learning using Entropy-SGD.
- the learning device 10a first initializes parameters (step S201). Next, the learning device 10a updates parameters using Entropy-SGD (step S202).
- step S203 If the evaluation criteria are not satisfied (step S203, No), the learning device 10a returns to step S202 and repeats the process. On the other hand, if the evaluation criteria are satisfied (step S203, Yes), the learning device 10a ends the process.
- E p ⁇ ( ⁇ ′)[ ⁇ ′] is approximately obtained by a method called Stochastic Gradient Langevin Dynamics (SGLD).
- SGLD Stochastic Gradient Langevin Dynamics
- FIG. 8 is a flowchart showing the flow of update processing in learning.
- the part surrounded by the dashed line in FIG. 8 corresponds to the third to eighth lines in FIG. 3, that is, the process of calculating the expected value of ⁇ ' by SGLD.
- the learning device 10a first increases l by 1 (step S301). Then, the learning device 10a applies an input randomly selected from the data set to the discriminator (step S302).
- the learning device 10a calculates the gradient, samples ⁇ ' according to p ⁇ ( ⁇ ') , and updates the average of ⁇ ' using this (step S303).
- step S304 if l is equal to or less than L (step S304, Yes), the learning device 10a returns to step S301 and repeats the process. On the other hand, if l is not equal to or less than L (step S304, No), the learning device 10a updates the model parameters (step S305).
- FIG. 2 is a diagram illustrating a configuration example of the learning device according to the first embodiment.
- the learning device 10 receives an input of a learning data set, performs model learning, and outputs a trained model.
- the learning device 10 has an interface section 11 , a storage section 12 and a control section 13 .
- the interface unit 11 is an interface for inputting and outputting data.
- the interface unit 11 includes a NIC (Network Interface Card).
- the interface unit 11 may include an input device such as a mouse and a keyboard, and an output device such as a display.
- the storage unit 12 is a storage device such as an HDD (Hard Disk Drive), an SSD (Solid State Drive), an optical disc, or the like. Note that the storage unit 12 may be a rewritable semiconductor memory such as RAM (Random Access Memory), flash memory, NVSRAM (Non Volatile Static Random Access Memory).
- the storage unit 12 stores an OS (Operating System) and various programs executed by the learning device 10 .
- the storage unit 12 also stores model information 121 .
- the model information 121 is information such as parameters for constructing a deep learning model (classifier).
- the model information 121 includes weights and biases of each layer of the deep neural network.
- the deep learning model constructed by the model information 121 may be a learned one or a pre-learned one.
- the control unit 13 controls the learning device 10 as a whole.
- the control unit 13 is, for example, an electronic circuit such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit), or an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).
- the control unit 13 also has an internal memory for storing programs defining various processing procedures and control data, and executes each processing using the internal memory. Further, the control unit 13 functions as various processing units by running various programs.
- the controller 13 has a calculator 131 and an updater 132 .
- the deep learning loss function L(x, y, ⁇ ) becomes a non-smooth function when adversarial learning or the like is performed. Gradient-based optimization is not efficient in that case.
- the learning device 10 has the following configuration. That is, the calculation unit 131 calculates an objective function by Entropy-SGD when data created as a hostile attack is input to the deep learning model. Also, the updating unit 132 updates the parameters of the deep learning model so that the objective function is optimized.
- the learning device 10 can learn according to the following Example 1 or Example 2, which is a more efficient version of Example 1.
- the learning device 10 uses the fact that the Hessian matrix becomes a variance-covariance matrix in Entropy-SGD, estimates the variance-covariance matrix by SGLD, and multiplies the inverse matrix of the Hessian matrix like the Newton method to improve efficiency. do.
- the learning device 10 calculates the Hessian matrix of Entropy-SGD.
- the (i, j) component of the Hessian matrix is given by equation (10).
- the matrix containing the components of formula (10) is formula (11).
- ⁇ ⁇ ' is the variance-covariance matrix of the probability density function p ⁇ ( ⁇ ').
- FIG. 4 is a diagram for explaining the algorithm of the embodiment.
- the learning device 10 approximates E p ⁇ [ ⁇ i ' ⁇ j ’] from the 13th to 17th lines in FIG. Ep ⁇ [ ⁇ i '] and Ep ⁇ [ ⁇ j '] are approximated, and in line 24, the inverse matrix of the Hessian matrix is calculated and multiplied by the gradient. As a result, speedup can be expected as with the Newton method.
- FIG. 9 shows the flow of update processing in this case.
- FIG. 9 is a flowchart showing the flow of update processing according to the embodiment. The processing of FIG. 9 corresponds to the algorithm of FIG.
- the learning device 10 first increases l by 1 (step S401). Then, the learning device 10 applies an input randomly selected from the data set to the discriminator (step S402).
- the learning device 10 calculates the gradient, samples ⁇ ' according to p ⁇ ( ⁇ ') , and updates the average of ⁇ ' using this (step S403).
- the learning device 10 updates the variance-covariance matrix using ⁇ ' (step S404).
- step S405, Yes if l is equal to or less than L (step S405, Yes), the learning device 10 returns to step S401 and repeats the process.
- step S405 the learning device 10 calculates the inverse matrix of the unit matrix and the estimated (updated) variance-covariance matrix (step S406). Then, the learning device 10 updates the model parameters using the calculated inverse matrix (step S407).
- the calculation unit 131 calculates a first matrix, which is a variance-covariance matrix of parameters according to the probability distribution used in Entropy-SGD, by SGLD (Stochastic Gradient Langevin Dynamics).
- the updating unit 132 updates the parameters of the deep learning model using the first matrix.
- the updating unit 132 also updates the parameters of the deep learning model by multiplying the gradient by the inverse matrix of the first matrix, which is a Hessian matrix.
- the learning device 10 only needs to multiply each parameter by the reciprocal of the variance.
- the learning device 10 approximates a variance-covariance matrix (covariance matrix) with a covariance of 0 by the algorithm indicated by the pseudo code in FIG.
- FIG. 5 is a diagram explaining the algorithm of the embodiment.
- the learning device 10 approximates E p ⁇ [ ⁇ i ' ⁇ j ’] from the 11th to 13th lines in FIG. E p ⁇ [ ⁇ i '] and E p ⁇ [ ⁇ j '] are approximated, and the inverse matrix of the Hessian matrix is calculated on the 18th line and multiplied by the gradient.
- FIG. 10 shows the flow of update processing in this case.
- FIG. 10 is a flowchart showing the flow of update processing according to the embodiment. The processing of FIG. 10 corresponds to the algorithm of FIG.
- the learning device 10 first increases l by 1 (step S501). Then, the learning device 10 applies an input randomly selected from the data set to the discriminator (step S502).
- the learning apparatus 10 calculates the gradient, samples ⁇ ' according to p ⁇ ( ⁇ ') , and updates the average of ⁇ ' using this (step S503).
- the learning device 10 updates the variance using ⁇ ' (step S504).
- step S505 if l is equal to or less than L (step S505, Yes), the learning device 10 returns to step S501 and repeats the process.
- step S505 if l is not equal to or less than L (step S505, No), the learning device 10 calculates a vector consisting of the identity matrix and the estimated (updated) variance (step S506). Then, the learning device 10 updates the model parameters using the calculated vector (step S507).
- the calculation unit 131 is the first Compute the matrix of .
- the updating unit 132 updates the parameters of the deep learning model using the first matrix.
- Entropy-SGD Examples 1 and 2 described so far are applicable to adversarial learning.
- an effect of improving learning efficiency while smoothing the objective function of adversarial learning is produced.
- FIG. 11 is a flowchart showing the flow of update processing in adversarial learning. As shown in FIG. 11, the learning device 10a first increases l by 1 (step S601). Then, learning device 10a randomly selects an input from the data set (step S602).
- the learning device 10a creates a hostile attack from the selected input (step S603). Then, the learning device 10a inputs (applies) the created hostile attack to the discriminator (step S604).
- the learning device 10a calculates the gradient, samples ⁇ ' according to p ⁇ ( ⁇ ') , and updates the average of ⁇ ' using this (step S605).
- step S606 if l is equal to or less than L (step S606, Yes), the learning device 10a returns to step S601 and repeats the process. On the other hand, if l is not equal to or less than L (step S606, No), the learning device 10a updates the model parameters (step S607).
- FIG. 12 is a flowchart showing the flow of update processing according to the embodiment in adversarial learning.
- the learning device 10 first increases l by 1 (step S701). Then, the learning device 10 randomly selects an input from the data set (step S702).
- the learning device 10 creates a hostile attack from the selected input (step S703). Then, the learning device 10 inputs (applies) the created hostile attack to the discriminator (step S704).
- the learning device 10 calculates the gradient, samples ⁇ ' according to p ⁇ ( ⁇ ') , and updates the average of ⁇ ' using this (step S705).
- the learning device 10 updates the variance-covariance matrix using ⁇ ' (step S706).
- step S707 Yes
- the learning device 10 returns to step S701 and repeats the process.
- step S707 the learning device 10 calculates the inverse matrix of the unit matrix and the estimated (updated) variance-covariance matrix (step S708). Then, the learning device 10 updates the model parameters using the calculated inverse matrix (step S709).
- FIG. 13 is a flowchart showing the flow of update processing according to the embodiment in adversarial learning. As shown in FIG. 13, the learning device 10 first increases l by 1 (step S801). Learning device 10 then randomly selects an input from the data set (step S802).
- the learning device 10 creates a hostile attack from the selected input (step S803). Then, the learning device 10 inputs (applies) the created hostile attack to the discriminator (step S804).
- the learning device 10 calculates the gradient, samples ⁇ ' according to p ⁇ ( ⁇ ') , and updates the average of ⁇ ' using this (step S805).
- the learning device 10 updates the variance using ⁇ ' (step S806).
- step S807 if l is equal to or less than L (step S807, Yes), the learning device 10 returns to step S801 and repeats the process.
- step S807 the learning device 10 calculates a vector consisting of the identity matrix and the estimated (updated) variance (step S808). Then, the learning device 10 updates the model parameters using the calculated vector (step S809).
- the calculation unit 131 calculates the objective function by Entropy-SGD when data created as a hostile attack is input to the deep learning model.
- the updating unit 132 updates the parameters of the deep learning model so that the objective function is optimized.
- the learning device 10 can improve learning efficiency while smoothing the objective function of adversarial learning.
- the calculation unit 131 also calculates a first matrix, which is a variance-covariance matrix of parameters according to the probability distribution used in Entropy-SGD, by SGLD (Stochastic Gradient Langevin Dynamics).
- the updating unit 132 updates the parameters of the deep learning model using the first matrix. As a result, the learning device 10 can improve the learning efficiency of adversarial learning.
- the calculation unit 131 calculates the first matrix assuming that the covariance of the variance-covariance matrix calculated by SGLD (Stochastic Gradient Langevin Dynamics) of parameters according to the probability distribution used in Entropy-SGD is 0. .
- the updating unit 132 updates the parameters of the deep learning model using the first matrix. Thereby, the learning device 10 can further improve the learning efficiency of adversarial learning.
- the updating unit 132 updates the parameters of the deep learning model by multiplying the gradient by the inverse matrix of the first matrix, which is the Hessian matrix. This allows the learning device 10 to smooth the gradient.
- each component of each device illustrated is functionally conceptual, and does not necessarily need to be physically configured as illustrated.
- the specific form of distribution and integration of each device is not limited to the illustrated one, and all or part of them can be functionally or physically distributed or Can be integrated and configured.
- all or any part of each processing function performed by each device is realized by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or hardware by wired logic can be realized as
- the learning device 10 can be implemented by installing a program for executing the above-described learning process as package software or online software on a desired computer.
- the information processing device can function as the learning device 10 by causing the information processing device to execute the above program.
- the information processing apparatus referred to here includes a desktop or notebook personal computer.
- information processing devices include mobile communication terminals such as smartphones, mobile phones and PHS (Personal Handyphone Systems), and slate terminals such as PDAs (Personal Digital Assistants).
- the learning device 10 can also be implemented as a server device that uses a terminal device used by a user as a client and provides the client with services related to the above processing.
- the server device is implemented as a server device that provides a service in which a data set is input and a trained deep learning model is output.
- the server device may be implemented as a web server, or may be implemented as a cloud that provides services related to the above processing by outsourcing.
- FIG. 14 is a diagram showing an example of a computer that executes programs.
- the computer 1000 has a memory 1010 and a CPU 1020, for example.
- Computer 1000 also has hard disk drive interface 1030 , disk drive interface 1040 , serial port interface 1050 , video adapter 1060 and network interface 1070 . These units are connected by a bus 1080 .
- the memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM (Random Access Memory) 1012 .
- the ROM 1011 stores a boot program such as BIOS (Basic Input Output System).
- BIOS Basic Input Output System
- Hard disk drive interface 1030 is connected to hard disk drive 1090 .
- a disk drive interface 1040 is connected to the disk drive 1100 .
- a removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100 .
- Serial port interface 1050 is connected to mouse 1110 and keyboard 1120, for example.
- Video adapter 1060 is connected to display 1130, for example.
- the hard disk drive 1090 stores, for example, an OS 1091, application programs 1092, program modules 1093, and program data 1094. That is, a program that defines each process of the learning device 10 is implemented as a program module 1093 in which computer-executable code is described. Program modules 1093 are stored, for example, on hard disk drive 1090 .
- the hard disk drive 1090 stores a program module 1093 for executing processing similar to the functional configuration of the learning device 10 .
- the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).
- the setting data used in the processing of the above-described embodiment is stored as program data 1094 in the memory 1010 or the hard disk drive 1090, for example. Then, the CPU 1020 reads the program modules 1093 and program data 1094 stored in the memory 1010 and the hard disk drive 1090 to the RAM 1012 as necessary, and executes the processes of the above-described embodiments.
- the program modules 1093 and program data 1094 are not limited to being stored in the hard disk drive 1090, but may be stored in a removable storage medium, for example, and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program modules 1093 and program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). Program modules 1093 and program data 1094 may then be read by CPU 1020 through network interface 1070 from other computers.
- LAN Local Area Network
- WAN Wide Area Network
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)
Abstract
Description
まず、図1を用いて深層学習モデルについて説明する。図1は、深層学習モデル全体の構造を例示する図である。なお、以降の説明では深層学習は学習装置10aによって実行されるものとする。
深層学習において滑らかさを改善する方法としてEntropy-SGDがある。元々の損失関数L(x,y,θ)に対して、Entropy-SGDは(7)式の損失関数を最小化する。
図2を用いて、第1の実施形態に係る学習装置の構成について説明する。図2は、第1の実施形態の学習装置の構成例を示す図である。学習装置10は、学習用データセットの入力を受け付け、モデルの学習を行い、学習済みモデルを出力する。
学習装置10は、Entropy-SGDにおいてヘッセ行列が分散共分散行列になることを用いて、SGLDによって当該分散共分散行列を推定し、ニュートン法のようにヘッセ行列の逆行列を掛けることで効率化する。
逆行列を計算するにはO(d3)の計算コストがかかるため、より効率的な方法として共分散を0と仮定してΣが各パラメータの分散からなる対角行列であると仮定する。すると、ヘッセ行列の逆行列は対角行列でその(i,i)成分は(12)式となる。
これまで説明してきたように、計算部131は、深層学習モデルに敵対的攻撃として作成されたデータを入力したときの目的関数をEntropy-SGDにより計算する。更新部132は、目的関数が最適化されるように深層学習モデルのパラメータを更新する。これにより、学習装置10は、敵対的学習の目的関数を滑らかにしつつ学習効率を向上させることができる。
また、図示した各装置の各構成要素は機能概念的なものであり、必ずしも物理的に図示のように構成されていることを要しない。すなわち、各装置の分散及び統合の具体的形態は図示のものに限られず、その全部又は一部を、各種の負荷や使用状況等に応じて、任意の単位で機能的又は物理的に分散又は統合して構成することができる。さらに、各装置にて行われる各処理機能は、その全部又は任意の一部が、CPU(Central Processing Unit)及び当該CPUにて解析実行されるプログラムにて実現され、あるいは、ワイヤードロジックによるハードウェアとして実現され得る。
一実施形態として、学習装置10は、パッケージソフトウェアやオンラインソフトウェアとして上記の学習処理を実行するプログラムを所望のコンピュータにインストールさせることによって実装できる。例えば、上記のプログラムを情報処理装置に実行させることにより、情報処理装置を学習装置10として機能させることができる。ここで言う情報処理装置には、デスクトップ型又はノート型のパーソナルコンピュータが含まれる。また、その他にも、情報処理装置にはスマートフォン、携帯電話機やPHS(Personal Handyphone System)等の移動体通信端末、さらには、PDA(Personal Digital Assistant)等のスレート端末等がその範疇に含まれる。
11 インタフェース部
12 記憶部
13 制御部
121 モデル情報
131 計算部
132 更新部
Claims (6)
- 深層学習モデルに敵対的攻撃として作成されたデータを入力したときの目的関数をEntropy-SGDにより計算する計算部と、
前記目的関数が最適化されるように前記深層学習モデルのパラメータを更新する更新部と、
を有することを特徴とする学習装置。 - 前記計算部は、Entropy-SGDの中で用いられる確率分布にしたがったパラメータの分散共分散行列である第1の行列をSGLD(Stochastic Gradient Langevin Dynamics)により計算し、
前記更新部は、前記第1の行列を用いて前記深層学習モデルのパラメータを更新することを特徴とする請求項1に記載の学習装置。 - 前記計算部は、Entropy-SGDの中で用いられる確率分布にしたがったパラメータのSGLD(Stochastic Gradient Langevin Dynamics)により計算される分散共分散行列の共分散を0と仮定した第1の行列を計算し、
前記更新部は、前記第1の行列を用いて前記深層学習モデルのパラメータを更新することを特徴とする請求項1に記載の学習装置。 - 前記更新部は、ヘッセ行列である前記第1の行列の逆行列を勾配に掛けて前記深層学習モデルのパラメータを更新することを特徴とする請求項2又は3に記載の学習装置。
- 学習装置によって実行される学習方法であって、
深層学習モデルに敵対的攻撃として作成されたデータを入力したときの目的関数をEntropy-SGDにより計算する計算工程と、
前記目的関数が最適化されるように前記深層学習モデルのパラメータを更新する更新工程と、
を含むことを特徴とする学習方法。 - コンピュータを、請求項1から4のいずれか一項に記載の学習装置として機能させるための学習プログラム。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2023523800A JPWO2022249327A1 (ja) | 2021-05-26 | 2021-05-26 | |
PCT/JP2021/019982 WO2022249327A1 (ja) | 2021-05-26 | 2021-05-26 | 学習装置、学習方法及び学習プログラム |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/019982 WO2022249327A1 (ja) | 2021-05-26 | 2021-05-26 | 学習装置、学習方法及び学習プログラム |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022249327A1 true WO2022249327A1 (ja) | 2022-12-01 |
Family
ID=84228647
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2021/019982 WO2022249327A1 (ja) | 2021-05-26 | 2021-05-26 | 学習装置、学習方法及び学習プログラム |
Country Status (2)
Country | Link |
---|---|
JP (1) | JPWO2022249327A1 (ja) |
WO (1) | WO2022249327A1 (ja) |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020235011A1 (ja) * | 2019-05-21 | 2020-11-26 | 日本電信電話株式会社 | 学習装置、学習方法及び学習プログラム |
JP2021022316A (ja) * | 2019-07-30 | 2021-02-18 | Kddi株式会社 | 学習装置、学習方法及び学習プログラム |
-
2021
- 2021-05-26 WO PCT/JP2021/019982 patent/WO2022249327A1/ja active Application Filing
- 2021-05-26 JP JP2023523800A patent/JPWO2022249327A1/ja active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020235011A1 (ja) * | 2019-05-21 | 2020-11-26 | 日本電信電話株式会社 | 学習装置、学習方法及び学習プログラム |
JP2021022316A (ja) * | 2019-07-30 | 2021-02-18 | Kddi株式会社 | 学習装置、学習方法及び学習プログラム |
Non-Patent Citations (1)
Title |
---|
GAURI JAGATAP; AMEYA JOSHI; ANIMESH BASAK CHOWDHURY; SIDDHARTH GARG; CHINMAY HEGDE: "Adversarially Robust Learning via Entropic Regularization", ARXIV.ORG, 19 February 2021 (2021-02-19), XP081882045 * |
Also Published As
Publication number | Publication date |
---|---|
JPWO2022249327A1 (ja) | 2022-12-01 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110689139A (zh) | 用于机器学习的方法和计算机系统 | |
JP7010371B2 (ja) | 学習済みモデル更新装置、学習済みモデル更新方法、プログラム | |
WO2022110640A1 (zh) | 一种模型优化方法、装置、计算机设备及存储介质 | |
US11030246B2 (en) | Fast and accurate graphlet estimation | |
US20180357566A1 (en) | Unsupervised learning utilizing sequential output statistics | |
US20210034976A1 (en) | Framework for Learning to Transfer Learn | |
WO2023051140A1 (zh) | 用于图像特征表示生成的方法、设备、装置和介质 | |
US11636384B1 (en) | Spherical random features for polynomial kernels | |
Sherzer et al. | Can machines solve general queueing systems? | |
JP2013037471A (ja) | 確率モデル更新システム、確率モデル更新装置、確率モデル更新方法およびプログラム | |
Chamoso et al. | Social computing for image matching | |
US20210326757A1 (en) | Federated Learning with Only Positive Labels | |
US20230297674A1 (en) | Detection device, detection method, and detection program | |
WO2022249327A1 (ja) | 学習装置、学習方法及び学習プログラム | |
CN111630530A (zh) | 数据处理系统和数据处理方法 | |
He et al. | An incremental kernel density estimator for data stream computation | |
WO2019208523A1 (ja) | 学習装置、学習方法及び学習プログラム | |
JP6928346B2 (ja) | 予測装置、予測方法および予測プログラム | |
US20240232625A1 (en) | Training device, training method, and training program | |
WO2024013911A1 (ja) | 学習装置、学習方法、学習プログラム、推論装置、推論方法、及び推論プログラム | |
US20230267316A1 (en) | Inference method, training method, inference device, training device, and program | |
JP2020087148A (ja) | 情報処理装置、情報処理方法およびプログラム | |
JP7396467B2 (ja) | 学習装置、分類装置、学習方法、分類方法、および、プログラム | |
WO2023195138A1 (ja) | 学習方法、学習装置及び学習プログラム | |
JP7327631B2 (ja) | 分類方法、分類装置及び分類プログラム |
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: 21942981 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2023523800 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18563397 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21942981 Country of ref document: EP Kind code of ref document: A1 |