WO2022114913A1 - 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화 시스템 및 그 방법 - Google Patents
인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화 시스템 및 그 방법 Download PDFInfo
- Publication number
- WO2022114913A1 WO2022114913A1 PCT/KR2021/017832 KR2021017832W WO2022114913A1 WO 2022114913 A1 WO2022114913 A1 WO 2022114913A1 KR 2021017832 W KR2021017832 W KR 2021017832W WO 2022114913 A1 WO2022114913 A1 WO 2022114913A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- neural network
- algorithm
- artificial intelligence
- intelligence technology
- resistance variable
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 62
- 238000005516 engineering process Methods 0.000 title claims abstract description 46
- 238000013473 artificial intelligence Methods 0.000 title claims abstract description 42
- 238000005457 optimization Methods 0.000 title claims abstract description 34
- 238000013528 artificial neural network Methods 0.000 claims abstract description 77
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 27
- 238000010606 normalization Methods 0.000 claims description 20
- 238000011176 pooling Methods 0.000 claims description 15
- 238000004590 computer program Methods 0.000 claims description 9
- 238000010586 diagram Methods 0.000 description 18
- 238000004891 communication Methods 0.000 description 6
- 238000012545 processing Methods 0.000 description 4
- 230000006870 function Effects 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000003247 decreasing effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 210000000653 nervous system Anatomy 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000005236 sound signal Effects 0.000 description 2
- 210000000225 synapse Anatomy 0.000 description 2
- 230000003044 adaptive effect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000001151 other effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000000946 synaptic effect Effects 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 230000000007 visual 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- 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
-
- 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
- G06N3/065—Analogue means
-
- 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
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- 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
- G06N3/09—Supervised learning
-
- 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
- G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
Definitions
- the present disclosure relates to a system and method for optimizing a layer of a stacked resistance variable memory device using artificial intelligence technology. More specifically, in a BNN (Binary Neural Networks) model, the optimal parameters are obtained by classifying the BNN model parameters into physical parameters and hyperparameters, and the minimum size channel with high accuracy and minimum deviation using the obtained parameters. To a system and method for optimizing a layer of a stacked resistance variable memory device using an artificial intelligence technology for calculating a value and optimizing a layer of a stacked resistance conversion memory (3D RRAM) using the calculated channel value.
- a BNN Binary Neural Networks
- Such a neuromorphic processor is a neural network device for driving various neural networks such as Binary Neural Networks (BNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Feedforward Neural Network (FNN), etc. and can be used in fields such as data classification or image recognition.
- BNN Binary Neural Networks
- CNN Convolutional Neural Network
- RNN Recurrent Neural Network
- FNN Feedforward Neural Network
- a neural network device requires a large amount of computation on complex input data.
- a technology capable of efficiently processing a neural network operation is required.
- a technology capable of minimizing the loss of accuracy while reducing the amount of computation required to process complex input data is required.
- An object of the present disclosure is to provide a system and method for optimizing a layer of a stacked resistance variable memory device using artificial intelligence technology.
- the layer optimization method of the stacked resistance variable memory device using artificial intelligence technology is a method using a neural network device for optimizing the layer of the stacked resistance variable memory device using artificial intelligence technology, the neural network device classifying the BNN parameters into physical parameters and hyperparameters in the BNN model; obtaining, by the neural network device, an optimal parameter using the physical parameter and the hyper parameter; and calculating, by the neural network device, a minimum channel size in the BNN model by using the optimal parameter.
- the physical parameters include the number of convolutional layers, the channel (filter) size, the kernel size, the presence or absence of batch normalization and the presence or absence of a pooling layer, and the hyper parameter is an optimization device (Optimizer), a learning rate (learning rate), and may include a momentum (Momentum).
- Optimizer optimization device
- Learning rate learning rate
- Momentum momentum
- the optimization device includes a batch gradient descent algorithm, a stochastic gradient descent (SGD) algorithm, a gradient descent algorithm, a mini-batch gradient descent algorithm, a momentum ( Momentum) algorithm, Adagrad algorithm, RMSprop algorithm, and may include at least one algorithm of Adam (Adam) algorithm.
- SGD stochastic gradient descent
- Grad gradient descent
- mini-batch gradient descent algorithm a momentum ( Momentum) algorithm
- Adagrad Adagrad algorithm
- RMSprop algorithm and may include at least one algorithm of Adam (Adam) algorithm.
- the acquiring of the optimal parameter may include: calculating an optimal algorithm using an algorithm of the optimization device included in the hyper parameter; calculating an optimal learning rate for adjusting a weight and intensity of a kernel update in BNN using the learning rate included in the hyper parameter; and calculating an optimal momentum in consideration of a momentum value in the BNN using the momentum included in the hyper parameter.
- the calculating of the optimal algorithm may include: calculating an optimal algorithm combination among at least one algorithm included in the optimization device; and calculating the ratio of the optimal algorithm combination.
- the ratio of the Adam algorithm and the SGD algorithm may be 3:7, and when the kernel size is 5 ⁇ 5, the ratio of the Adam algorithm and the SGD algorithm may be 6:4.
- the same optimal learning rate may be obtained, and the optimal learning rate may have a learning rate of 0.03.
- the optimal momentum may have a value of 0.5 momentum when the kernel size is 3x3, and 0.6 momentum when the kernel size is 5x5.
- the obtaining of the optimal parameter may include determining whether the batch normalization and the pooling layer exist.
- the batch normalization may be included in each of the convolutional layers.
- the batch normalization may be performed by [Equation 1].
- the pooling layer may be located in the last two layers among the convolutional layers.
- the calculating of the minimum channel size may include determining accuracy of the minimum channel size.
- the accuracy may be 96% or more when the kernel size is 3x3.
- the calculating of the minimum channel size may further include determining the accuracy of the minimum channel size by applying an error rate for each layer for each minimum channel size.
- the accuracy may be 94.069% at the 40nm node, the accuracy may be 93.777% at the 20nm node, and the accuracy may be 93.07% at the 10nm node.
- the system for optimizing the layer of the stacked resistance variable memory device using the artificial intelligence technology may perform a neural network method for optimizing the layer of the stacked resistance variable memory device using the artificial intelligence technology by the neural network device.
- the program according to an embodiment of the present disclosure is stored in a computer-readable recording medium so as to perform the layer optimization method of the stacked resistance variable memory device using the artificial intelligence technology in combination with a computer that is hardware.
- robot devices such as drones, advanced driver assistance systems (ADAS), etc., smart TVs, smartphones, medical devices, mobile devices, image display devices, measurement devices, IoT devices, etc.
- ADAS advanced driver assistance systems
- smart TVs smartphones
- medical devices mobile devices
- image display devices measurement devices
- IoT devices IoT devices
- FIG. 1 is a conceptual diagram illustrating a system for optimizing a layer of a stacked resistance variable memory device using an artificial intelligence technology according to an embodiment of the present disclosure.
- FIG. 2 is a view for explaining a method of optimizing a layer of a stacked resistance variable memory device using an artificial intelligence technology according to an embodiment of the present disclosure.
- FIG. 3 is a diagram for explaining a step of obtaining an optimal parameter shown in FIG. 2 .
- 4 and 5 are diagrams for explaining the step of calculating the optimal algorithm combination shown in FIG.
- 6 and 7 are diagrams for explaining the step of calculating the optimal learning rate shown in FIG.
- FIG. 8 and 9 are diagrams for explaining the step of calculating the optimum momentum shown in FIG. 3 .
- 10 to 12 are diagrams for explaining the step of determining the presence or absence of the batch normalization and pooling layer shown in FIG. 3 .
- 13 and 14 are diagrams for explaining the step of calculating the minimum channel size shown in FIG. 2 .
- FIG. 15 is a diagram for explaining a step of optimizing the layer of the RRAM having the minimum deviation shown in FIG. 2 .
- FIG. 16 is a hardware configuration diagram of a computing device capable of implementing the neural network device shown in FIG. 1 .
- first, second, etc. are used to describe various elements, these elements are not limited by these terms, of course. These terms are only used to distinguish one component from another. Accordingly, it goes without saying that the first component mentioned below may be the second component within the spirit of the present invention.
- FIG. 1 is a conceptual diagram illustrating a system for optimizing a layer of a stacked resistance variable memory device using an artificial intelligence technology according to an embodiment of the present disclosure.
- the hierarchical optimization system 1 of the stacked resistance variable memory device using the artificial intelligence technology which is an embodiment of the present disclosure, is a filter of the BNN (Binary Neural Networks) model by the neural network device 10 .
- BNN Binary Neural Networks
- FIG. 2 is a view for explaining a method of optimizing a layer of a stacked resistance variable memory device using an artificial intelligence technology, according to an embodiment of the present disclosure.
- FIG. 3 is a diagram for explaining a step of obtaining an optimal parameter shown in FIG. 2 .
- 4 and 5 are diagrams for explaining a method of calculating an optimal algorithm combination shown in FIG. 3 .
- 6 and 7 are diagrams for explaining a method of calculating the optimal learning rate shown in FIG. 3 .
- 8 and 9 are diagrams for explaining a method of calculating the optimum momentum shown in FIG. 3 .
- 10 to 12 are diagrams for explaining the step of determining the presence or absence of the batch normalization and pooling layer shown in FIG. 3 .
- FIG. 13 and 14 are diagrams for explaining a method of calculating the minimum channel size shown in FIG. 2 .
- FIG. 15 is a diagram for explaining a method of optimizing a layer of an RRAM having a minimum deviation shown in FIG. 2 .
- the layer optimization method of the stacked resistance variable memory device using artificial intelligence technology was performed in the coded condition, the present invention is not limited thereto.
- Kernel Size Convolutional Layer Channel Size 3 ⁇ 3 2 10 15 20 4 10 15 20 6 10 15 20 5 ⁇ 5 2 10 15 20 4 10 15 20 6 10 15 20
- the neural network device 10 may classify BNN parameters into physical parameters and hyper parameters in the BNN model ( S10 ).
- the physical parameter may include the number of convolutional layers, the channel (filter) size, the kernel size, the presence or absence of batch normalization and the presence or absence of a pooling layer, but is not limited thereto. .
- the hyperparameter may include, but is not limited to, an optimizer, a learning rate, and a momentum, the number of hidden units, the mini-batch size, the number of hidden layers, and the learning rate. decay) can be included.
- the neural network device 10 may obtain an optimal parameter by using the physical parameter and the hyper parameter ( S20 ).
- the neural network device 10 may calculate the optimal algorithm by using the algorithm of the optimization device included in the hyper parameter ( S100 ).
- the algorithm of the optimization device is a batch gradient descent algorithm that considers all data when calculating an error, and stochastic gradient descent that calculates only one randomly selected data rather than all data when adjusting parameter values.
- SGD Spochastic Gradient Descent
- Gradient Descent Algorithm that adjusts the values of parameters by calculating only a set amount
- Mini-Batch Gradient Descent Mini-Batch
- Momentum Algorithm that applies the law of physics called inertia
- Adagrad Algorithm that applies different learning rates to each parameter
- RMS Prop that improves the disadvantage of decreasing learning rate of Adagrad Algorithm (RMSprop)
- Adam Adaptive Moment Estimation
- the neural network device 10 has low optimal accuracy to calculate the optimal algorithm, but the SGD algorithm in which the speed is improved by adjusting the weights for randomly extracted data rather than adjusting the weights for all data
- the optimal combination of algorithms can be calculated using the Adam algorithm, which combines the strengths of the RMSprop algorithm and the momentum algorithm. That is, in the neural network device 10, the ratio of the Adam algorithm and the SGD algorithm is 3:7 when the kernel size is 3x3, and the ratio of the Adam algorithm and the SGD algorithm is 6:4 when the kernel size is 5x5. can be calculated.
- the neural network device 10 may calculate an optimal ratio between the Adam algorithm and the SGD algorithm.
- the optimal ratio with the highest accuracy can be calculated by comparing the case where the Adam algorithm is high and the case where the SGD algorithm is high.
- the ratio of Adam algorithm and SGD algorithm is 1:9, the ratio of Adam algorithm and SGD algorithm is 2:8, the ratio of Adam algorithm and SGD algorithm is 3:7, the ratio of Adam algorithm and SGD algorithm is 4: 6, the ratio of the Adam algorithm and the SGD algorithm is 5:5, the ratio of the Adam algorithm and the SGD algorithm is 6:4, the ratio of the Adam algorithm and the SGD algorithm is 7:3, the ratio of the Adam algorithm and the SGD algorithm is 8:1 and It is repeated for the condition that the ratio of Adam algorithm and SGD algorithm is 9:1.
- the optimal ratio of the Adam algorithm and the SGD algorithm is 3:7 when the kernel size is 3x3, and the optimal combination of the Adam algorithm and the SGD algorithm is 6:4 when the kernel size is 5x5. can be calculated.
- the accuracy according to a change in the channel size (refer to FIG. 5(a) ) and the number of convolutional layers is 4 Among the accuracy according to the change in the channel size in the case of (see Fig. 5 (b)) and the accuracy according to the change in the channel size when the number of convolution layers is 6 (see Fig. 5 (c)), the Adam algorithm and the SGD algorithm It can be seen that the highest accuracy is achieved when the ratio of is 3:7.
- the neural network device 10 may use the learning rate included in the hyper parameter to calculate the optimal learning rate for adjusting the weight and the intensity of the kernel update in the BNN ( S110 ). That is, when the kernel size is 3x3 and the kernel size is 5x5, the neural network device 10 may calculate the same optimal learning rate of 0.03.
- the neural network device 10 can calculate an optimal learning rate of 0.03 with an accuracy of 99% by comparing and analyzing the learning rate that has the meaning of how much to learn when learning from 0.01 to 0.1 once. .
- the accuracy according to the change in the channel size (refer to FIG. 6(a)) and the number of convolutional layers are 4
- the learning rate is 0.03 among the accuracy according to the change in the channel size (see Fig. 6 (b)) and the accuracy according to the change in the channel size when the number of convolution layers is 6 (see Fig. 6 (c))
- the neural network apparatus 10 may calculate the optimal momentum by considering the momentum value in the BNN by using the momentum included in the hyper parameter ( S120 ). That is, the neural network device 10 may calculate an optimal momentum having a momentum of 0.5 when the kernel size is 3 ⁇ 3 and a momentum of 0.6 when the kernel size is 5 ⁇ 5.
- the neural network device 10 considers different momentum values from 0.1 to 1.0 in order to calculate an optimal momentum value to reduce training time, and when the kernel size is 3 ⁇ 3, the momentum is 0.5, and the kernel When the size is 5 ⁇ 5, the optimum momentum with a momentum of 0.6 can be calculated.
- the neural network device 10 may determine the presence or absence of batch normalization (S130). That is, the neural network device 10 may determine whether batch normalization is required for the convolutional layer.
- batch normalization may be calculated by Equation 1 below.
- the neural network device 10 may determine whether a pooling layer exists (S140). That is, the neural network device 10 may determine whether a pooling layer exists in the convolutional layer.
- the pooling layer may be located in the last two layers among the convolutional layers.
- the pooling layer when the pooling layer is located in the last two layers among the convolutional layers, referring to FIG. 12 , it can be seen that the pooling layer has the highest accuracy.
- step S130 of determining the presence or absence of the batch normalization may be performed after the step S140 of determining the existence of the pooling layer.
- steps S130 and S140 may be simultaneously performed.
- the neural network device 10 may calculate the minimum channel size in the BNN model by using the optimal parameters (S30). That is, the neural network device 10 may determine the accuracy of the minimum channel size and calculate the corresponding minimum channel size.
- the neural network device 10 may calculate the channel size starting from 9 and decreasing one by one.
- the channel size may be independent of the number of convolutional layers.
- the accuracy is 96% or more when the kernel size is 3 ⁇ 3. That is, it is possible to minimize the layer of the stacked resistance variable memory device while maintaining or increasing the accuracy in the BNN model.
- the neural network device 10 may determine the accuracy of the minimum channel size by applying the error rate for each layer to each minimum channel size.
- the accuracy is 94.069% at the 40nm node, the accuracy is 93.777% at the 20nm node, and the accuracy is 93.07 at the 10nm node. It can be %.
- the neural network device 10 may optimize the RRAM layer having the minimum deviation by using the minimum channel size (S40).
- the neural network device 10 may optimize the channel size to 8 by reducing the channel size of 50 by 80% or more while maintaining accuracy.
- Such a neural network device 10 may include various portable electronic communication devices for performing a layer optimization method of a stacked resistance variable memory device using artificial intelligence technology.
- a smart phone for performing a layer optimization method of a stacked resistance variable memory device using artificial intelligence technology.
- PDA personal digital assistant
- a tablet for example, a wearable device, for example, a watch-type terminal (Smartwatch), a glass-type terminal (including Smart Glass), HMD (Head Mounted Display), etc.) and various Internet of Things (IoT) terminals, but are not limited thereto.
- IoT Internet of Things
- 16 is a hardware configuration diagram of an exemplary computing device capable of implementing the neural network device 10 .
- the computing device 800 loads one or more processors 810 , a storage 850 for storing a computer program 851 , and a computer program 851 executed by the processor 810 . and a memory 820 , a bus 830 , and a network interface 840 .
- processors 810 the computing device 800 loads one or more processors 810 , a storage 850 for storing a computer program 851 , and a computer program 851 executed by the processor 810 .
- a memory 820 for storing a computer program 851 , and a computer program 851 executed by the processor 810 .
- a bus 830 for storing a computer program 851
- a network interface 840 a network interface
- the processor 810 controls the overall operation of each component of the computing device 800 .
- the processor 810 includes a central processing unit (CPU), a micro processor unit (MPU), a micro controller unit (MCU), a graphic processing unit (GPU), or any type of processor well known in the art of the present disclosure. can be In addition, the processor 810 may perform an operation on at least one computer program for executing the method of optimizing the layer of the stacked resistance variable memory device using artificial intelligence technology according to an embodiment of the present disclosure.
- Computing device 800 may include one or more processors.
- the memory 820 stores data supporting various functions of the computing device 800 .
- the memory 820 stores a plurality of computer programs (app, application program, or application software) driven in the computing device 800 , data, instructions, and one or more information for the operation of the computing device 800 .
- At least some of the computer programs may be downloaded from an external device (not shown).
- at least a part of the computer program may exist on the computing device 800 from the time of shipment for basic functions (eg, receiving a message, sending a message) of the computing device 800 .
- the memory 820 may load one or more computer programs 851 from the storage 850 in order to execute the method for optimizing the layer of the stacked resistance variable memory device using artificial intelligence technology according to an embodiment of the present disclosure.
- RAM random access memory
- the bus 830 provides communication functions between components of the computing device 800 .
- the bus 830 may be implemented as various types of buses, such as an address bus, a data bus, and a control bus.
- the network interface 840 supports wired/wireless Internet communication of the computing device 800 . Also, the network interface 840 may support various communication methods other than Internet communication. To this end, the network interface 840 may be configured to include a communication module well-known in the technical field of the present disclosure.
- the storage 850 may non-temporarily store one or more computer programs 851 .
- the storage 850 is a non-volatile memory, such as a read only memory (ROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a flash memory, a hard disk, a removable disk, or well in the art to which the present disclosure pertains. It may be configured to include any known computer-readable recording medium.
- the computing device 800 may further include an input unit and an output unit in addition to the components illustrated in FIG. 16 .
- the input unit may include a camera for receiving an image signal, a microphone for receiving an audio signal, and a user input unit for receiving information from a user.
- the user input unit may include one or more of a touch key and a mechanical key.
- the image data collected through the camera or the audio signal collected through the microphone may be analyzed and processed by a user's control command.
- the output unit is for visual, auditory or tactile output of the command processing result, and may include a display unit, an optical output unit, a speaker, a haptic output unit, and an optical output unit.
- Steps of a method or algorithm described in relation to an embodiment of the present disclosure may be implemented directly in hardware, as a software module executed by hardware, or by a combination thereof.
- a software module may include random access memory (RAM), read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, hard disk, removable disk, CD-ROM, or this It may reside in any type of computer-readable recording medium well known in the art to which the disclosure pertains.
- the disclosed technology can be applied to a neural network using a stacked resistance conversion memory (3D RRAM), a neural network device, and a neural network system.
- 3D RRAM stacked resistance conversion memory
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Image Analysis (AREA)
- Complex Calculations (AREA)
- Semiconductor Memories (AREA)
Abstract
Description
Kernel Size | Convolutional Layer | Channel Size |
3×3 |
2 |
10 |
15 | ||
20 | ||
4 |
10 | |
15 | ||
20 | ||
6 |
10 | |
15 | ||
20 | ||
5×5 |
2 |
10 |
15 | ||
20 | ||
4 |
10 | |
15 | ||
20 | ||
6 |
10 | |
15 | ||
20 |
Claims (20)
- 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 장치를 이용한 방법에 있어서,상기 뉴럴 네트워크 장치가 BNN모델에서 BNN매개변수를 물리적 매개변수 및 하이퍼 매개변수로 분류하는 단계;상기 뉴럴 네트워크 장치가 상기 물리적 매개변수 및 상기 하이퍼 매개변수를 이용하여 최적의 파라미터를 획득하는 단계;상기 뉴럴 네트워크 장치가 상기 최적의 파라미터를 이용하여 상기 BNN모델에서 최소채널크기를 산출하는 단계;를 포함하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제1항에 있어서,상기 물리적 매개변수는 컨볼루션 레이어(Convolutional Layer)의 수, 채널(필터) 크기, 커널 크기, 배치 정규화(Batch Normalization) 유무 및 풀링 계층(Pooling Layer) 유무를 포함하고,상기 하이퍼 매개변수는 최적화 기기(Optimizer), 학습률(learning rate) 및 모멘텀(Momentum)을 포함하는,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제2항에 있어서,상기 최적화 기기는 배치 경사 하강(Batch Gradient Descent) 알고리즘, 확률적 경사 하강(Stochastic Gradient Descent, SGD) 알고리즘, 경사 하강(Gradient Descent) 알고리즘, 미니 배치 경사 하강(Mini-Batch Gradient Descent) 알고리즘, 모멘텀(Momentum) 알고리즘, 아다그라드(Adagrad) 알고리즘, 알엠에스프롭(RMSprop) 알고리즘 및 아담(Adam) 알고리즘 중 적어도 하나의 알고리즘을 포함하는,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제3항에 있어서,상기 최적의 파라미터를 획득하는 단계는,상기 하이퍼 매개변수에 포함된 상기 최적화 기기의 알고리즘을 이용하여 최적알고리즘을 산출하는 단계;상기 하이퍼 매개변수에 포함된 상기 학습률을 이용하여 BNN에서 가중치 및 커널 업데이트의 강도를 조절하는 최적학습률을 산출하는 단계; 및상기 하이퍼 매개변수에 포함된 상기 모멘텀을 이용하여 BNN에서 운동량값을 고려하여 최적모멘텀을 산출하는 단계;를 포함하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제4항에 있어서,상기 최적알고리즘을 산출하는 단계는,상기 최적화 기기에 포함된 적어도 하나의 알고리즘 사이의 최적의 알고리즘조합을 산출하는 단계; 및상기 최적의 알고리즘조합의 비율을 산출하는 단계;를 포함하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제5항에 있어서,상기 최적알고리즘은,커널 크기가 3×3인 경우 Adam 알고리즘과 SGD 알고리즘의 비율이 3:7 이고,커널 크기가 5×5인 경우 Adam 알고리즘과 SGD 알고리즘의 비율이 6:4 인,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제4항에 있어서,커널 크기가 3×3 및 커널 크기가 5×5인 경우 동일한 최적학습률을 갖고,상기 최적학습률은 0.03 학습률을 가지는,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제4항에 있어서,상기 최적모멘텀은,커널 크기가 3×3인 경우 0.5 모멘텀의 값을 가지고,커널 크기가 5×5인 경우 0.6 모멘텀의 값을 가지는,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제4항에 있어서,상기 최적의 파라미터를 획득하는 단계는,상기 배치 정규화 및 상기 풀링 계층의 존재유무를 판단하는 단계;를 포함하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제9항에 있어서,컨볼루션 레이어의 수가 4이고, 채널 크기가 9이고, 커널 크기가 3×3인 경우, 상기 배치 정규화는 상기 컨볼루션 레이어 각각에 포함되는,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제10항에 있어서,상기 배치 정규화는 상기 컨볼루션 레이어 각각에 포함되지 않는 경우, 8%의 정확도가 차이나는,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제9항에 있어서,컨볼루션 레이어의 수가 4이고, 채널 크기가 9이고, 커널 크기가 3×3인 경우, 상기 풀링 계층은 상기 컨볼루션 레이어 중 마지막 2개의 계층에 위치하는,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제1항에 있어서,상기 최소채널크기를 산출하는 단계는,상기 최소채널크기의 정확성을 판단하는 단계;를 포함하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제14항에 있어서,상기 최소채널크기가 8인 경우, 커널 크기가 3×3인 일 때 정확도가 96% 이상인,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제15항에 있어서,상기 최소채널크기를 산출하는 단계는,상기 최소채널크기별로 계층별 에러율을 적용하여 상기 최소채널크기의 정확성을 판단하는 단계;를 더 포함하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제16항에 있어서,상기 최소채널크기가 8이고, 커널 크기가 3×3인 인 경우,40nm 노드에서 정확도가 94.069%이고, 20nm 노드에서 정확도가 93.777%이며, 10nm 노드에서 정확도가 93.07%인,인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제1항에 있어서,상기 뉴럴 네트워크 장치가 상기 최소채널크기를 이용하여 최소편차를 갖는 RRAM의 계층을 최적화하는 단계;를 포함하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 방법.
- 제1항 내지 제18항 중 어느 한 항에 기재된 방법을 뉴럴 네트워크 장치에 의해 수행하는, 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화하는 뉴럴 네트워크 시스템.
- 하드웨어인 컴퓨터와 결합되어, 제1항 내지 제18항 중 어느 한 항에 기재된 방법을 수행할 수 있도록 컴퓨터에서 독출가능한 기록매체에 저장된 컴퓨터 프로그램.
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US18/039,402 US20240020511A1 (en) | 2020-11-30 | 2021-11-30 | System and method for performing layer optimization of stacked resistive random-access memory device by using artificial intelligence technology |
CN202180091192.XA CN116724317A (zh) | 2020-11-30 | 2021-11-30 | 利用人工智能技术的堆叠式阻变存储器元件的层优化系统及其方法 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
KR1020200163675A KR102539876B1 (ko) | 2020-11-30 | 2020-11-30 | 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화 시스템 및 그 방법 |
KR10-2020-0163675 | 2020-11-30 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022114913A1 true WO2022114913A1 (ko) | 2022-06-02 |
Family
ID=81756228
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/KR2021/017832 WO2022114913A1 (ko) | 2020-11-30 | 2021-11-30 | 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화 시스템 및 그 방법 |
Country Status (4)
Country | Link |
---|---|
US (1) | US20240020511A1 (ko) |
KR (1) | KR102539876B1 (ko) |
CN (1) | CN116724317A (ko) |
WO (1) | WO2022114913A1 (ko) |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110991621A (zh) * | 2019-12-13 | 2020-04-10 | 中国科学院计算技术研究所 | 一种基于通道数搜索卷积神经网络的方法 |
CN111160534A (zh) * | 2019-12-31 | 2020-05-15 | 中山大学 | 一种适用于移动终端的二值神经网络前向传播框架 |
KR20200091785A (ko) * | 2019-01-23 | 2020-07-31 | 주식회사 스트라드비젼 | 하드웨어 최적화를 통해 모바일 장치 또는 높은 정확도의 소형 네트워크에 사용될 cnn 파라미터 양자화를 최적화하기 위하여 cnn 레이어를 변환하는 방법 및 장치 |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR102526650B1 (ko) | 2017-05-25 | 2023-04-27 | 삼성전자주식회사 | 뉴럴 네트워크에서 데이터를 양자화하는 방법 및 장치 |
-
2020
- 2020-11-30 KR KR1020200163675A patent/KR102539876B1/ko active IP Right Grant
-
2021
- 2021-11-30 US US18/039,402 patent/US20240020511A1/en active Pending
- 2021-11-30 CN CN202180091192.XA patent/CN116724317A/zh active Pending
- 2021-11-30 WO PCT/KR2021/017832 patent/WO2022114913A1/ko active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20200091785A (ko) * | 2019-01-23 | 2020-07-31 | 주식회사 스트라드비젼 | 하드웨어 최적화를 통해 모바일 장치 또는 높은 정확도의 소형 네트워크에 사용될 cnn 파라미터 양자화를 최적화하기 위하여 cnn 레이어를 변환하는 방법 및 장치 |
CN110991621A (zh) * | 2019-12-13 | 2020-04-10 | 中国科学院计算技术研究所 | 一种基于通道数搜索卷积神经网络的方法 |
CN111160534A (zh) * | 2019-12-31 | 2020-05-15 | 中山大学 | 一种适用于移动终端的二值神经网络前向传播框架 |
Non-Patent Citations (2)
Title |
---|
BOGDAN PENKOVSKY; MARC BOCQUET; TIFENN HIRTZLIN; JACQUES-OLIVIER KLEIN; ETIENNE NOWAK; ELISA VIANELLO; JEAN-MICHEL PORTAL; DAMIEN : "In-Memory Resistive RAM Implementation of Binarized Neural Networks for Medical Applications", ARXIV.ORG, 20 June 2020 (2020-06-20), 201 Olin Library Cornell University Ithaca, NY 14853 , XP081699610 * |
XUANYI DONG; YI YANG: "Network Pruning via Transformable Architecture Search", ARXIV.ORG, 23 May 2019 (2019-05-23), 201 Olin Library Cornell University Ithaca, NY 14853 , XP081500056 * |
Also Published As
Publication number | Publication date |
---|---|
KR20220075521A (ko) | 2022-06-08 |
KR102539876B1 (ko) | 2023-06-02 |
US20240020511A1 (en) | 2024-01-18 |
CN116724317A (zh) | 2023-09-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2017213398A1 (en) | Learning model for salient facial region detection | |
WO2019156283A1 (en) | Dynamic memory mapping for neural networks | |
WO2020111647A1 (en) | Multi-task based lifelong learning | |
WO2020238353A1 (zh) | 数据处理方法和装置、存储介质及电子装置 | |
Santhalingam et al. | Sign language recognition analysis using multimodal data | |
WO2022059969A1 (ko) | 심전도 데이터 분류를 위한 심층 신경망 사전 학습 방법 | |
EP3620982A1 (en) | Sample processing method and device | |
WO2018212584A2 (ko) | 딥 뉴럴 네트워크를 이용하여 문장이 속하는 클래스를 분류하는 방법 및 장치 | |
WO2020222391A1 (en) | System and method for invertible wavelet layer for neural networks | |
WO2023282569A1 (en) | Method and electronic device for generating optimal neural network (nn) model | |
WO2020231005A1 (ko) | 영상 처리 장치 및 그 동작방법 | |
WO2020192523A1 (zh) | 译文质量检测方法、装置、机器翻译系统和存储介质 | |
WO2020222378A1 (en) | Convolution streaming engine for deep neural networks | |
CN114925320B (zh) | 一种数据处理方法及相关装置 | |
WO2019045147A1 (ko) | 딥러닝을 pc에 적용하기 위한 메모리 최적화 방법 | |
WO2022114913A1 (ko) | 인공지능 기술을 이용한 적층형 저항 변화 메모리 소자의 계층 최적화 시스템 및 그 방법 | |
WO2020096332A1 (en) | System and method for cached convolution calculation | |
WO2022071708A1 (ko) | 컨텐츠 추천 서비스를 제공하는 전자 장치 및 그 방법 | |
WO2022244997A1 (en) | Method and apparatus for processing data | |
WO2022139325A1 (ko) | 과적합 없는 단일 신경망 기반 다중 도메인 적응 학습을 위한 컴퓨터 시스템, 및 그의 방법 | |
WO2021242073A1 (ko) | 전자 장치 및 이의 제어 방법 | |
WO2022196955A1 (ko) | 인공지능 모델을 사전 학습시키는 방법 및 장치 | |
WO2021010540A1 (ko) | 딥러닝 내부의 데이터 추출 방법 및 장치 | |
WO2022045526A1 (ko) | 정규화된 합성곱 신경망 시스템 및 합성곱 신경망 정규화 방법 | |
CN112035649A (zh) | 问答模型处理方法、装置、计算机设备及存储介质 |
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: 21898738 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202180091192.X Country of ref document: CN |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18039402 Country of ref document: US |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21898738 Country of ref document: EP Kind code of ref document: A1 |