KR20220154902A - Electronic device and method for controlling electronic device - Google Patents

Abstract

Disclosed are an electronic device and a control method thereof. According to the present invention present disclosure, the electronic device includes: a memory which stores a scaling factor of a neural network model; and a processor which acquires quantization data by quantizing weight data used for calculation of the neural network model with a combination of code data and scaling factor data. The processor acquires code data corresponding to the weight data by generating a random number with a pre-set seed based on a random number generator, and learns the neural network model by updating the scaling factor data based on the acquired code data and the scaling factor data pre-stored in the memory. Accordingly, the electronic device may acquire a neural network model with compressibility.

Description

Electronic device and its control method {ELECTRONIC DEVICE AND METHOD FOR CONTROLLING ELECTRONIC DEVICE}

The present disclosure relates to an electronic device and a control method thereof, and more particularly, to an electronic device for quantizing weights of a neural network model and a control method thereof.

Recently, development and research on artificial intelligence systems that implement human-level intelligence have been conducted. An artificial intelligence system refers to a system that performs learning and reasoning based on a neural network model, unlike conventional rule-based systems, and is used in various fields such as voice recognition, image recognition, and future prediction.

In particular, recently, an artificial intelligence system that solves a given problem through a deep neural network based on deep learning has been developed.

A deep neural network is a neural network that includes multiple hidden layers between an input layer and an output layer. A model that implements a technology. In general, deep neural networks include multiple weight values in order to derive accurate result values.

On the other hand, since a deep neural network includes a huge amount of weight values, a problem arises in that resources required for calculation gradually increase. In addition, when computation is compressed or simplified on a deep neural network, accuracy of computation may be reduced.

The present disclosure has been made to solve the above-mentioned problems, and an object of the present disclosure relates to an electronic device and a control method for lightening a neural network model by quantizing weights of the neural network model.

According to an embodiment of the present disclosure, an electronic device includes a memory configured to store scaling factor data corresponding to a neural network model; and a processor for obtaining quantization data by quantizing weight data used in the operation of the neural network model into a combination of code data and scaling factor data, wherein the processor includes a random number generator. generator), a random number is generated with a preset seed to obtain code data corresponding to the weight data, and based on the obtained code data and scaling factor data previously stored in the memory The neural network model is learned by updating the scaling factor data.

Further, the processor obtains output data by performing a forward-pass on the neural network model based on the code data and scaling factor data stored in the memory, and the output data, the code data, and the Based on the scaling factor data, back-pass is performed on the neural network model to obtain a gradient value corresponding to the scaling factor data, and based on the gradient, the scaling factor Data may be updated, and the updated scaling factor data may be stored in the memory.

The processor generates first code data corresponding to the first weight data according to a preset seed based on the random number generator, and generates the first code data and at least one code data corresponding to the first code data. performs a multiplication operation on the first scaling factor data of , generates second code data corresponding to second weight data according to a preset seed based on the random number generator, and A multiplication operation may be performed on at least one second scaling factor data corresponding to the second code data.

The neural network model is composed of a plurality of layers, and the first code data is data corresponding to a weight between a plurality of nodes included in the first layer and a plurality of nodes included in the second layer of the neural network model. , The second code data may be data corresponding to weights between a plurality of nodes included in the second layer and a plurality of nodes included in the third layer of the neural network model.

The obtained code data includes a random number corresponding to each weight included in the weight data, and the random number may be one of -1 and +1.

The processor may normalize the obtained code data and perform an operation on the normalized code data and the scaling factor data to learn the neural network model.

And, the obtained code data may not be stored in the memory.

When the neural network model is learned, the processor generates the code data using the random number generator based on the preset seed, and the updated scaling factor data stored in the memory and the generated Output data may be generated by performing forward propagation on the learned neural network model using code data.

Meanwhile, a control method of an electronic device including a memory for storing scaling factor data corresponding to a neural network model according to another embodiment of the present disclosure is based on a random number generator, generating a random number to obtain code data corresponding to weight data used for calculation of the neural network model; and learning the neural network model by updating the scaling factor data based on the obtained code data and the scaling factor data stored in the memory.

The learning may include obtaining output data by performing a forward-pass on the neural network model based on the code data and scaling factor data stored in the memory; obtaining a gradient value corresponding to the scaling factor data by performing backward-pass on the neural network model based on the output data, the sign data, and the scaling factor data; updating the scaling factor data based on the gradient value; and storing the updated scaling factor data in the memory.

The obtaining of the output data may include generating first code data corresponding to the first weight data according to a preset seed based on the random number generator, and generating the first code data and the first code data. performing a multiplication operation on at least one first scaling factor data corresponding to the data; and generating second code data corresponding to the second weight data according to a preset seed based on the random number generator, and performing at least one second scaling corresponding to the second code data and the second code data. Performing a multiplication operation on factor data; may include.

The neural network model is composed of a plurality of layers, and the first code data is data corresponding to a weight between a plurality of nodes included in the first layer and a plurality of nodes included in the second layer of the neural network model. , The second code data may be data corresponding to weights between a plurality of nodes included in the second layer and a plurality of nodes included in the third layer of the neural network model.

The obtained code data may include a random number corresponding to each weight included in the weight data, and the random number may be one of -1 and +1.

The learning may include normalizing the generated code data; and learning the neural network model by performing an operation on the normalized code data and the scaling factor data.

[10] The method of claim 9, wherein the obtained code data is not stored in the memory.

Then, when the neural network model is learned, generating the code data using the random number generator based on the preset seed; and generating output data by performing forward propagation on the learned neural network model using the updated scaling factor data stored in the memory and the generated code data.

As described above, according to various embodiments of the present disclosure, an electronic device may obtain a neural network model with an improved compression rate.

1 is a block diagram briefly illustrating the configuration of an electronic device according to an embodiment of the present disclosure;
2 is a diagram illustrating quantization of weight data into code data and scaling factor data according to the present disclosure.
3 is a diagram for explaining a method of generating code data through a random number generator according to the present disclosure.
4 is a diagram for explaining code data generated by a random number generator according to a seed according to the present disclosure.
5 is a diagram for explaining a process of updating scaling factor data through forward propagation and back propagation according to the present disclosure.
6A is a diagram for explaining a Uniform Quantization method.
6B is a diagram for explaining a Non-Uniform Quantization method.
7 is a diagram for explaining an embodiment in which a plurality of code data is generated by a random number generator according to an embodiment of the present disclosure.
8 is a diagram for explaining a control method of an electronic device according to the present disclosure.
9 is a block diagram illustrating in detail the configuration of an electronic device according to the present disclosure.

The present disclosure relates to an electronic device for obtaining a lightweight neural network model by obtaining quantization data by quantizing weight data of a neural network model and a control method thereof.

The electronic device of the present disclosure can reduce a floating-point multiplication operation process required to perform an operation between weight data and input data by quantizing weight data into code data scaling factor data.

In addition, the electronic device may obtain a neural network model with an improved compression ratio by configuring code data with a combination of random numbers generated by the random number generator according to the present disclosure without storing code data.

Hereinafter, the present disclosure will be described in detail with reference to the drawings.

1 is a block diagram briefly illustrating the configuration of an electronic device 100 according to an embodiment of the present disclosure. As shown in FIG. 1 , the electronic device 100 may include a memory 110 and a processor 120 . However, the configuration shown in FIG. 1 is an exemplary diagram for implementing the embodiments of the present disclosure, and the electronic device 100 may additionally include appropriate hardware and software configurations that are obvious to those skilled in the art.

Meanwhile, in describing the present disclosure, the electronic device 100 is a device that obtains output data for input data by learning, compressing, or using a neural network model of a neural network model (or artificial intelligence model). , For example, the electronic device 100 may be implemented as a desktop PC, a laptop computer, a smart phone, a tablet PC, a server, and the like.

In addition, various operations performed by the electronic device 100 may be performed by a system in which a cloud computing environment is established. For example, a system in which a cloud computing environment is built may perform an operation between quantization data and input data using quantization data of a neural network model.

The memory 110 may store commands or data related to at least one other component of the electronic device 100 . Also, the memory 110 is accessed by the processor 120, and reading/writing/modifying/deleting/updating of data by the processor 120 may be performed.

In the present disclosure, the term memory refers to the memory 110, a ROM (not shown) in the processor 120, a RAM (not shown), or a memory card (not shown) mounted in the electronic device 100 (eg, micro SD). card, memory stick). In addition, the memory 110 may store programs and data for composing various screens to be displayed on the display area of the display.

In addition, the memory 110 may include a non-volatile memory capable of maintaining stored information even if power supply is interrupted, and a volatile memory requiring continuous power supply to maintain stored information. For example, the non-volatile memory may be implemented with at least one of OTPROM (one time programmable ROM), PROM (programmable ROM), EPROM (erasable and programmable ROM), EEPROM (electrically erasable and programmable ROM), mask ROM, and flash ROM, The volatile memory may be implemented as at least one of dynamic RAM (DRAM), static RAM (SRAM), and synchronous dynamic RAM (SDRAM).

The memory 110 may store weight data used for calculation of the neural network model. However, according to the present disclosure, the memory 110 may store only scaling factor data among quantization data obtained by quantizing weight data.

The weight data may include a plurality of weight values included in the weight data. In this case, the weight value may be implemented in the form of n (n is a natural number greater than or equal to 1) bit floating-point. For example, the weight data may be implemented as a 32-bit floating point. Weight data may be represented by at least one of a vector, matrix, or tensor.

The memory 110 may store quantization data in which weight data is quantized as a combination of sign data and scaling factor data. Quantization data may be expressed as at least one of a vector, matrix, or tensor according to a format of weight data.

The sign data may include 1 or -1, which is a sign value capable of determining only a sign without changing the size of weight data and a scaling factor on which an operation is performed. Scaling factor data may be expressed in a floating point format (eg, 32-bit floating point format) similar to the format of weight data. A method of quantizing the weight data will be described in a later section.

The memory 110 may store various types of input data. For example, the memory 110 may store voice data input through a microphone, image data or text data input through an input unit (eg, a camera, a keyboard, etc.). Input data stored in the memory 110 may include data received through an external device.

The processor 120 may be electrically connected to the memory 110 to control overall operations and functions of the electronic device 100 . The processor 120 may include one or a plurality of processors to control the operation of the electronic device 100 .

The processor 120 may load data necessary for performing various operations from non-volatile memory to volatile memory. Loading refers to an operation of loading and storing data stored in a non-volatile memory into a volatile memory so that the processor 120 can access the data.

In addition, the volatile memory may be implemented as a component included in the processor 120 as a component of the processor 120, but this is only one embodiment, and the volatile memory is implemented as a component separate from the processor 120 It can be.

The processor 120 may obtain quantization data by quantizing weight data. Quantizing the weight means simplifying the unit of the weight or expressing it in a different way in order to efficiently use the weight.

For example, the processor 120 may obtain quantization data by performing quantization using a binary coding method on weight values included in the weight data. And, the processor 120 may store the acquired quantization data in the memory 110 . Performing quantization of the binary code method on the weight values means that the weight values are quantized using a combination of code data and scaling factor data.

For example, performing binary code quantization on weight values based on k (k is a natural number greater than or equal to 1) bits means expressing weights in a way of summing the products of k sign values and scaling factors. do.

When k is 3, weight data may be quantized as shown in Equation 1 below. In Equation 1, W means weight data, A means scaling factor data, and B means code data.

[Equation 1]

When quantization is performed on weights using a binary code method, the processor 120 may determine a k value based on an accuracy level required when performing an operation of a neural network model. Since the weight can be expressed more accurately as the value of k increases, the value of k may be determined as a larger value to increase the accuracy of output data obtained through the neural network model.

Accordingly, when the level of accuracy required for performing the computation of the neural network model is high, the processor 120 may determine the value of k as a high value. An accuracy level required when performing an operation of a neural network model may be determined according to a type of input data or may be determined when a user designs a neural network model.

For example, when the input data is language data or voice data requiring high arithmetic accuracy, the processor 120 determines the value of k as 5, and the input data is image data requiring relatively low arithmetic accuracy. In this case, the processor 120 may determine the value of k as 3. However, this is only an example, and a k value corresponding to each type of input data may be allocated and may be freely changed by the user.

2 is a diagram illustrating quantization of weight data (W) into code data (b) and scaling factor data (α) according to the present disclosure.

Specifically, referring to FIG. 2 , the weight data W may be expressed in a 4x4 floating point format. As an embodiment of the present disclosure, the weight data (W) of FIG. 2 may mean weight data (W) connecting adjacent nodes in a neural network model. For example, the weight data W of FIG. 2 may mean a set of weights connecting nodes included in the first layer and nodes included in the second layer of the neural network model. That is, one floating point value of the weight data W of FIG. 2 may mean a weight value connecting any one node included in the first layer and any one node included in the second layer.

Further, quantization data may be composed of sign data (b) in which floating point values in the weight data (W) are expressed as -1 or +1, respectively, and scaling factor data (α) in the form of floating points.

As an example according to the present disclosure, one floating point data included in the weight data (W) of FIG. 2 may be implemented with 32 bits, and either +1 or -1 included in the code data (b) may be implemented with 1 bit. and the scaling factor data α may be implemented with 32 bits. However, the present disclosure is not limited thereto and may be implemented in various ways according to various data precisions.

According to the present disclosure, the processor 120 may generate the code data b of FIG. 2 using a random number generator. That is, code data (b) corresponding to the weight data (W) may be generated by generating a random number with a preset seed.

3 is a diagram for explaining a method of generating code data through a random number generator according to the present disclosure.

A random number generator is a component for generating a random number, which is a random number within a given range. The random number generator according to the present disclosure may randomly generate one of -1 and +1, and configure code data with a combination of random numbers generated through the random number generator.

Referring to FIG. 3 , the random number generator may obtain a plurality of random number combinations by randomly generating one of -1 and +1. In addition, code data may be configured by arranging the plurality of obtained random numbers.

The random number generator according to the present disclosure can generate code data by generating random numbers corresponding to each floating point of weight data, and a combination of generated random numbers is determined according to a seed value, which is an initial setting value. If fixed, the random number generator can generate random numbers with the same combination.

4 is a diagram for explaining code data generated by a random number generator according to a seed according to the present disclosure.

Referring to FIG. 4 , the random number generator may determine a combination of generated random numbers according to a seed value. That is, the random number generator may generate first code data 31 by generating a random number with a first seed value, and generate second code data 32 by generating a random number with a second seed value.

The processor 120 according to the present disclosure may improve the compression ratio of the neural network model by using code data generated by a random number generator without separately storing code data among quantization data in a memory. That is, according to the present disclosure, since only scaling factor data among quantization data is stored in the memory 110 and code data is not stored in the memory 110 but generated through a random number generator, there is no need to store the code data in the memory. Meaning that the code data is not stored in the memory 110 may mean, for example, that the code data is not stored in a non-volatile memory such as DRAM. Accordingly, the processor 120 may improve the compression rate of the neural network model through quantization data.

For example, one floating point data included in the weight data (W) is 32 bits, +1 or -1 included in the code data is 1 bit, scaling factor data (α) is 32 bits, and 128 codes When the data share one scaling factor data, the conventional compression rate becomes 1-(128*1bit+1*32bit)/(128*32bit)=96.09%. However, when using a random number generator according to the present disclosure, since there is no need to store code data in memory, a compression rate of 1-(128*0bit+1*32bit)/(128*32bit)=99.21% may be possible. .

5 is a diagram for explaining a process of updating scaling factor data through forward propagation and back propagation according to the present disclosure.

According to the present disclosure, the processor 120 may generate code data with an initially set seed value through a random number generator, and perform learning of a neural network model through the generated code data. In this case, the processor 120 may update scaling factor data based on sign data generated by the random number generator and scaling factor data stored in a memory. That is, in the learning process of the quantization neural network model, conventionally, both sign data and scaling factor data are updated and learning is performed. However, according to the present disclosure, only scaling factor data is updated, so resources and time required for updating can be saved. have.

Specifically, the processor 120 may obtain output data by performing forward-pass on the neural network model based on code data generated by the random number generator and scaling factor data stored in a memory. Forward-pass is a process of obtaining output data while performing an operation between input data and weight data. That is, when input data is input to the neural network model, the processor 120 may obtain output data by performing forward propagation on the neural network model, performing calculations on the input data, code data, and scaling factor data.

Then, the processor 120 performs backward-pass using the obtained output data, the scaling factor data stored in the memory, and the code data generated by the random number generator to obtain a gradient corresponding to the scaling factor data. value can be obtained. Backward-pass is a process for minimizing an error in output data, and is a process of updating weight data in a direction in which an error in output data is minimized. A gradient value is a value indicating the degree of error in output data, and means a slope indicating a point where the degree of error in output data is minimum. A smaller degree of error in the output data may indicate better performance of the neural network model, and the gradient value may be a learning index representing performance of the neural network model.

Also, the processor 120 may update scaling factor data based on a gradient value corresponding to the scaling factor data. And, the processor 120 may store the updated scaling factor data in the memory 110 .

Conventionally, sign data and scaling factor data of quantization data are updated through backpropagation. However, since code data is generated through a random number generator, the processor 120 according to the present disclosure may update only scaling factor data. That is, according to the present disclosure, since only scaling factor data is updated, time and resources required for updating existing code data can be saved.

And, the processor 120 may generate output data through the learned neural network model. Specifically, the processor 120 generates code data based on the seed value used when the learned neural network model is learned, and performs forward propagation using the learned scaling factor data and the generated code data to obtain output data. can create That is, as an example, the processor 120 may store a seed value used when the learned neural network model is learned in the memory 110 .

As an example according to the present disclosure, the processor 120 may normalize code data generated by the random number generator. The processor 120 may generate code data so that the random number included in the code data is uniform through a random number generator. That is, the random number generator may generate a random number such that the number of -1 and +1 included in the code data is the same. However, it is not limited thereto, and the random number generator may generate random numbers such that the ratio of the number of -1 and the number of +1 is similar within a preset error range (eg, 5%).

Also, the processor 120 may normalize the code data generated using the distribution convertor 500 so that the code data has a Gaussian distribution.

There are two ways to normalize code data. First, there is a uniform quantization method that expresses quantized code data values at equal intervals. 6A is a diagram for explaining a Uniform Quantization method, and FIG. 6B is a diagram for explaining a Non-Uniform Quantization method.

The uniform quantization method sets the minimum and maximum values of code data as shown in FIG. 6a and divides them at equal intervals to express them as integer values. In this case, the difference between the maximum and minimum values of the code data may be the scaling factor data value.

Second, there is a non-uniform quantization method that expresses quantized code data values as non-uniform values using a plurality of scaling factor data. For example, when expressing the non-uniform quantization method with only 2 bits as shown in FIG. 6B, 1 bit represents the value of the first scaling factor data (α1) among the 2 bits, and the remaining 1 bit represents the value of the second scaling factor data. (α2) can be represented. In this case, the code data values quantized by the number of four cases may form a distribution of non-uniform intervals.

The processor 120 may normalize the code data through the above two methods using the distribution converter 500, and perform forward propagation and back propagation of the neural network model using the normalized code data.

Meanwhile, functions related to artificial intelligence according to the present disclosure are operated through the processor 120 and the memory 110. Processor 120 may be composed of one or a plurality of processors. At this time, one or more processors may include general-purpose processors such as CPU (Central Processing Unit), AP (Application Processor), DSP (Digital Signal Processor), graphics-only processors such as GPU (Graphic Processing Unit) and VPU (Vision Processing Unit). Or it could be a processor dedicated to artificial intelligence, such as a Neural Processing Unit (NPU).

One or more processors 120 control input data to be processed according to predefined operation rules or artificial intelligence models stored in the memory 110 . Alternatively, when one or more processors are processors dedicated to artificial intelligence, the processors dedicated to artificial intelligence may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

A predefined action rule or an artificial intelligence model is characterized in that it is created through learning. Here, being made through learning means that a basic artificial intelligence model is learned using a plurality of learning data by a learning algorithm, so that a predefined action rule or artificial intelligence model set to perform a desired characteristic (or purpose) is created. means burden. Such learning may be performed in the device itself in which artificial intelligence according to the present disclosure is performed, or through a separate server and/or system.

Examples of learning algorithms include supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, but are not limited to the above examples.

The artificial intelligence model includes a plurality of artificial neural networks, and the artificial neural networks may be composed of a plurality of layers. Each of the plurality of neural network layers has a plurality of weight values, and a neural network operation is performed through an operation between an operation result of a previous layer and a plurality of weight values. A plurality of weights possessed by a plurality of neural network layers may be optimized by a learning result of an artificial intelligence model. For example, a plurality of weights may be updated so that a loss value or a cost value obtained from an artificial intelligence model is reduced or minimized during a learning process.

Examples of artificial neural networks include Convolutional Neural Network (CNN), Deep Neural Network (DNN), Recurrent Neural Network (RNN), Restricted Boltzmann Machine (RBM), Deep Belief Network (DBN), Bidirectional Recurrent Deep Neural Network (BRDNN), and There are deep Q-networks and the like, and the artificial neural network in the present disclosure is not limited to the above-described examples except for the cases specified.

7 is a diagram for explaining an embodiment in which a plurality of code data is generated by a random number generator according to an embodiment of the present disclosure.

Referring to FIG. 7 , the random number generator 300 according to the present disclosure may generate a plurality of code data corresponding to the number of layers of a neural network model. For example, the random number generator 300 generates random numbers according to the first seed and sequentially converts the first code data 40-1, the second code data 40-2 to the N-th code data 40-N. can create Here, the first code data 40 - 1 may correspond to a weight value for connecting a plurality of nodes included in the first layer and a plurality of nodes included in the second layer of the neural network model. And, the second code data 40 - 2 may correspond to a weight value for connecting a plurality of nodes included in the second layer and a plurality of nodes included in the third layer of the neural network model. In addition, the Nth code data 40 -N may correspond to a weight value for connecting a plurality of nodes included in the Nth layer and a plurality of nodes included in the N+1th layer of the neural network model.

And, as an embodiment, scaling factor data may be allocated to each of a plurality of code data. That is, the first scaling factor data 45-1 corresponding to the first code data 40-1 may perform an operation with the first code data 40-1, and the Nth code data 40-N The N-th scaling factor data 45-N corresponding to ) may perform an operation with the N-th code data 40-N. However, this is only an example, and a plurality of code data may share one scaling factor data. Alternatively, a plurality of scaling factor data may be allocated to one code data. That is, the first scaling factor data 45-1, the second scaling factor data 45-2, and the Nth scaling factor data 45-N may be implemented in plural numbers.

Although it has been described in FIG. 7 that a plurality of code data is generated through the first seed, the present disclosure is not limited thereto. That is, as an example, the first code data 40-1 is generated through the first seed, the second code data 40-2 is generated through the second seed, and the N-th code data is generated through the N-th seed. (40-N).

In addition, in the above-described embodiments, weight data and sign data have been described as two-dimensional data, but the present disclosure is not limited thereto and may be implemented in various ways such as a four-dimensional tensor.

8 is a diagram for explaining a control method of an electronic device according to the present disclosure.

The electronic device 100 may store scaling factor data (S810). The electronic device 100 according to the present disclosure may acquire quantization data by quantizing weight data used for computation of a neural network model as a combination of code data and scaling factor data. And, the electronic device 100 may store scaling factor data in memory. As an example, scaling factor data may be expressed in a floating point form, and in an initial stage of learning a neural network model, scaling factor data may be stored in a memory as an arbitrary floating point number. In addition, scaling factor data stored in the memory may be updated according to learning of the neural network model.

The electronic device 100 may obtain code data corresponding to the weight data by generating a random number with a preset seed based on the random number generator (S820). The sign data may include a random number corresponding to each weight included in the weight data. And, for example, the random number may be one of -1 and +1.

As an embodiment, the electronic device 100 generates first code data corresponding to the first weight data according to a preset seed based on a random number generator, and generates at least one code data corresponding to the first code data and the first code data. A multiplication operation may be performed on the first scaling factor data. Further, the electronic device 100 generates second code data corresponding to the second weight data according to a preset seed based on the random number generator, and generates the second code data and at least one second code data corresponding to the second code data. A multiplication operation may be performed on scaling factor data. Here, the neural network model is composed of a plurality of layers, and the first code data is data corresponding to weights between a plurality of nodes included in the first layer and a plurality of nodes included in the second layer of the neural network model, and the second The sign data may be data corresponding to weights between a plurality of nodes included in the second layer and a plurality of nodes included in the third layer of the neural network model.

Then, the electronic device 100 may learn the neural network model by updating scaling factor data based on the acquired code data and pre-stored scaling factor data (S830).

Specifically, the electronic device 100 may obtain output data by performing forward propagation on the neural network model based on the code data and the scaling factor data stored in the memory. When forward propagation is performed by inputting input data to the neural network model to obtain output data, the electronic device 100 performs back propagation on the neural network model based on the output data, sign data, and scaling factor data to obtain scaling factor data. A corresponding gradient value may be obtained. Also, the electronic device 100 may update scaling factor data based on the gradient value and store the updated scaling factor data in a memory.

That is, the electronic device 100 stores the scaling factor data updated through the learning process of the neural network model in the memory, but does not store the code data generated by the random number generator in the memory. Only information about can be stored in memory. Then, when the neural network model is learned, the electronic device 100 generates code data using a random number generator based on a preset seed, and obtains output data using the generated code data and scaling factor data stored in a memory. can

9 is a block diagram illustrating in detail the configuration of an electronic device according to the present disclosure. As shown in FIG. 9 , the electronic device 100 includes a memory 910, a processor 920, a communication unit 930, a display 940, a speaker 950, a microphone 960, and an input unit 970. can do. Since the memory 910 and the processor 920 have been specifically described with reference to the memory 110 and the processor 120 of FIG. 1 , overlapping descriptions will be omitted.

The communication unit 930 includes a circuit and can communicate with an external device. In this case, communication of the communication unit 930 with an external device may include communication through a third device (eg, a repeater, a hub, an access point, a server, or a gateway).

The communication unit 930 may include various communication modules to communicate with external devices. As an example, the communication unit 930 may include a wireless communication module, and for example, 5G (5TH Generation), LTE, LTE-A (LTE Advance), CDMA (code division multiple access), WCDMA (wideband CDMA) , and the like may include a cellular communication module using at least one.

As another example, the wireless communication module may include, for example, at least one of WiFi (Wireless Fidelity), Bluetooth, Bluetooth Low Energy (BLE), Zigbee, Radio Frequency (RF), or Body Area Network (BAN). can However, this is only one embodiment and the communication unit 930 may include a wired communication module.

The communication unit 930 may transmit scaling factor data constituting the neural network model trained according to the present disclosure to an external server. That is, the processor 920 according to the present disclosure may learn a neural network model, update scaling factor data corresponding to the neural network model, and store the updated scaling factor data in the memory 910 . Also, the communication unit 930 may transmit scaling factor data stored in the memory 910 to an external server.

In this case, information on the seed used to learn the neural network model may also be transmitted to an external server. That is, the processor 920 may update scaling factor data using sign data generated using a preset seed, and store information on the preset seed used for updating the scaling factor data in the memory 910. . Also, the communication unit 930 may transmit scaling factor data stored in the memory 910 to an external server together with information on a preset seed.

The external server may generate code data by generating a random number according to the received seed using a random number generator, perform forward propagation through the received scaling factor data and the generated code data, and obtain output data.

The communication unit 930 may receive input data input to the neural network model from an external device communicatively connected to the electronic device 100 . For example, the communication unit 930 receives various types of input data from an input device (eg, a camera, microphone, keyboard, etc.) connected to the electronic device 100 through wireless communication or an external server capable of providing various contents. can do.

The display 940 may display various information according to the control of the processor 120 . In particular, the display 940 may display input data or display output data obtained by performing an operation between weight data and input data. Here, displaying the output data may include an operation of displaying a screen including text or images generated based on the output data.

The display 940 may be a Liquid Crystal Display (LCD), Organic Light Emitting Diodes (OLED), Active-Matrix Organic Light-Emitting Diode (AM-OLED), Liquid Crystal on Silicon (LcoS) or Digital Light Processing (DLP). It can be implemented with various display technologies. Also, the display 940 may be coupled to at least one of a front area, a side area, and a rear area of the electronic device 100 in the form of a flexible display.

Also, the display 940 may be implemented as a touch screen having a touch sensor.

The speaker 950 is a component that outputs various audio data on which various processing tasks such as decoding, amplification, and noise filtering have been performed by an audio processing unit (not shown). In addition, the speaker 950 may output various notification sounds or voice messages.

For example, when an operation result between weight data and input data, that is, output data is output by a neural network model, the speaker 950 may output a notification sound indicating that the output data has been obtained.

The microphone 960 is a component capable of receiving voice input from a user. The microphone 960 may be provided inside the electronic device 100, but may be provided outside and electrically connected to the electronic device 100. In addition, when the microphone 960 is provided externally, the microphone 960 may transmit the generated user voice signal to the processor 120 through a wired/wireless interface (eg, Wi-Fi, Bluetooth).

The microphone 960 may receive a user voice including a wake-up word (or trigger word) capable of activating an artificial intelligence model composed of various artificial neural networks. When the user's voice including the wake-up word is received through the microphone 960, the processor 120 activates an artificial intelligence model and may perform an operation between weight data using the user's voice as input data.

The input unit 970 includes a circuit and can receive a user input for controlling the electronic device 100 . In particular, the input unit 970 may include a touch panel for receiving a user's touch using a user's hand or a stylus pen, and a button for receiving a user's manipulation. As another example, the input unit 970 may be implemented as another input device (eg, a keyboard, mouse, motion input unit, etc.). Meanwhile, the input unit 970 may receive first input data input from a user or various user commands.

The present embodiments can apply various transformations and have various embodiments, and specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the scope to the specific embodiments, and should be understood to include various modifications, equivalents, and/or alternatives of the embodiments of the present disclosure. In connection with the description of the drawings, like reference numerals may be used for like elements.

In describing the present disclosure, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the gist of the present disclosure, a detailed description thereof will be omitted.

In addition, the above-described embodiments may be modified in various forms, and the scope of the technical spirit of the present disclosure is not limited to the following embodiments. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the spirit of the disclosure to those skilled in the art.

Terms used in this disclosure are only used to describe specific embodiments, and are not intended to limit the scope of rights. Singular expressions include plural expressions unless the context clearly dictates otherwise.

In the present disclosure, expressions such as “has,” “can have,” “includes,” or “can include” indicate the presence of a corresponding feature (eg, numerical value, function, operation, or component such as a part). , which does not preclude the existence of additional features.

In this disclosure, expressions such as “A or B,” “at least one of A and/and B,” or “one or more of A or/and B” may include all possible combinations of the items listed together. . For example, “A or B,” “at least one of A and B,” or “at least one of A or B” (1) includes at least one A, (2) includes at least one B, Or (3) may refer to all cases including at least one A and at least one B.

Expressions such as "first," "second," "first," or "second," used in the present disclosure may modify various elements regardless of order and/or importance, and may refer to one element as It is used only to distinguish it from other components and does not limit the corresponding components.

A component (e.g., a first component) is "(operatively or communicatively) coupled with/to" another component (e.g., a second component); When referred to as "connected to", it should be understood that the certain component may be directly connected to the other component or connected through another component (eg, a third component).

On the other hand, when an element (eg, a first element) is referred to as being “directly connected” or “directly connected” to another element (eg, a second element), the element and the above It may be understood that other components (eg, a third component) do not exist between the other components.

The expression “configured to (or configured to)” as used in this disclosure means, depending on the situation, for example, “suitable for,” “having the capacity to.” ," "designed to," "adapted to," "made to," or "capable of." The term "configured (or set) to" may not necessarily mean only "specifically designed to" hardware.

Instead, in some contexts, the phrase "device configured to" may mean that the device is "capable of" in conjunction with other devices or components. For example, the phrase "a processor configured (or configured) to perform A, B, and C" may include a dedicated processor (eg, embedded processor) to perform the operation, or by executing one or more software programs stored in a memory device. , may mean a general-purpose processor (eg, CPU or application processor) capable of performing corresponding operations.

In an embodiment, ' ' or 'unit' performs at least one function or operation, and may be implemented as hardware or software, or a combination of hardware and software. In addition, a plurality of '' or a plurality of 'units' may be integrated into at least one and implemented by at least one processor, except for '' or 'units' that need to be implemented with specific hardware.

Meanwhile, various elements and regions in the drawings are schematically drawn. Therefore, the technical spirit of the present invention is not limited by the relative size or spacing drawn in the accompanying drawings.

Meanwhile, various embodiments described above may be implemented in a recording medium readable by a computer or a similar device using software, hardware, or a combination thereof. According to the hardware implementation, the embodiments described in this disclosure are application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), and field programmable gate arrays (FPGAs). ), processors, controllers, micro-controllers, microprocessors, and electrical units for performing other functions. In some cases, the embodiments described herein may be implemented by a processor itself. According to software implementation, embodiments such as procedures and functions described in this specification may be implemented as separate software. Each of these pieces of software may perform one or more functions and operations described herein.

Meanwhile, the method according to various embodiments of the present disclosure described above may be stored in a non-transitory readable medium. Such non-transitory readable media may be loaded and used in various devices.

A non-transitory readable medium is not a medium that stores data for a short moment, such as a register, cache, or memory, but a medium that stores data semi-permanently and can be read by a device. Specifically, programs for performing the various methods described above may be stored and provided in a non-transitory readable medium such as a CD, DVD, hard disk, Blu-ray disk, USB, memory card, or ROM.

According to one embodiment, the method according to various embodiments disclosed in this document may be included and provided in a computer program product. Computer program products may be traded between sellers and buyers as commodities. A computer program product may be distributed in the form of a device-readable storage medium (eg, compact disc read only memory (CD-ROM)) or online through an application store (eg, Play Store ^TM ). In the case of online distribution, at least part of the computer program product may be temporarily stored or temporarily created in a storage medium such as a manufacturer's server, an application store server, or a relay server's memory.

In addition, although the preferred embodiments of the present disclosure have been shown and described above, the present disclosure is not limited to the specific embodiments described above, and the technical field to which the disclosure belongs without departing from the subject matter of the present disclosure claimed in the claims. Of course, various modifications are possible by those skilled in the art, and these modifications should not be individually understood from the technical spirit or perspective of the present disclosure.

100: electronic device
110: memory
120: processor

Claims (16)

In electronic devices,
a memory for storing scaling factor data corresponding to the neural network model; and
A processor for obtaining quantization data by quantizing weight data used in the operation of the neural network model as a combination of code data and scaling factor data;
the processor,
Based on a random number generator, generating a random number with a preset seed to obtain code data corresponding to the weight data,
The electronic device learning the neural network model by updating the scaling factor data based on the obtained code data and the scaling factor data pre-stored in the memory.
According to claim 1,
the processor,
Obtaining output data by performing a forward-pass on the neural network model based on the code data and scaling factor data stored in the memory;
Obtaining a gradient value corresponding to the scaling factor data by performing backward-pass on the neural network model based on the output data, the sign data, and the scaling factor data;
Based on the gradient, updating the scaling factor data;
An electronic device that stores the updated scaling factor data in the memory.
According to claim 2,
The processor
First code data corresponding to first weight data is generated according to a preset seed based on the random number generator, and at least one first scaling factor corresponding to the first code data and the first code data is generated. perform a multiplication operation on the data;
Second code data corresponding to second weight data is generated according to a preset seed based on the random number generator, and the second code data and at least one second scaling factor corresponding to the second code data are generated. An electronic device that performs multiplication operations on data.
According to claim 3,
The neural network model is composed of a plurality of layers,
The first code data is data corresponding to weights between a plurality of nodes included in the first layer and a plurality of nodes included in the second layer of the neural network model,
The second code data is data corresponding to weights between a plurality of nodes included in the second layer and a plurality of nodes included in the third layer of the neural network model.
According to claim 1,
The electronic device, characterized in that the obtained code data includes a random number corresponding to each weight included in the weight data, and the random number is one of -1 and +1.
According to claim 1,
the processor,
Normalize the generated code data,
An electronic device learning the neural network model by performing an operation on the normalized code data and the scaling factor data.
According to claim 1,
The electronic device, characterized in that the obtained code data is not stored in the memory.
According to claim 1,
the processor,
When the neural network model is learned, the code data is generated using the random number generator based on the preset seed;
The electronic device generating output data by performing forward propagation on the learned neural network model using the updated scaling factor data stored in the memory and the generated code data.
A control method of an electronic device including a memory for storing scaling factor data corresponding to a neural network model,
obtaining code data corresponding to weight data used for calculation of the neural network model by generating a random number with a preset seed based on a random number generator; and
and learning the neural network model by updating the scaling factor data based on the obtained code data and the scaling factor data stored in the memory.
According to claim 9,
The learning step is
obtaining output data by performing a forward-pass on the neural network model based on the code data and scaling factor data stored in the memory;
obtaining a gradient value corresponding to the scaling factor data by performing backward-pass on the neural network model based on the output data, the sign data, and the scaling factor data;
updating the scaling factor data based on the gradient value; and
and storing the updated scaling factor data in the memory.
According to claim 10,
Obtaining the output data,
First code data corresponding to first weight data is generated according to a preset seed based on the random number generator, and at least one first scaling factor corresponding to the first code data and the first code data is generated. performing a multiplication operation on the data; and
Second code data corresponding to second weight data is generated according to a preset seed based on the random number generator, and the second code data and at least one second scaling factor corresponding to the second code data are generated. A control method comprising: performing a multiplication operation on data.
According to claim 11,
The neural network model is composed of a plurality of layers,
The first code data is data corresponding to weights between a plurality of nodes included in the first layer and a plurality of nodes included in the second layer of the neural network model,
The second code data is data corresponding to weights between a plurality of nodes included in the second layer and a plurality of nodes included in the third layer of the neural network model.
According to claim 9,
The control method characterized in that the obtained code data includes a random number corresponding to each weight included in the weight data, and the random number is one of -1 and +1.
According to claim 9,
The learning step is
normalizing the generated code data; and
and learning the neural network model by performing an operation on the normalized code data and the scaling factor data.
According to claim 9,
The control method, characterized in that the obtained code data is not stored in the memory.
According to claim 9,
generating the code data using the random number generator based on the preset seed when the neural network model is learned; and
and generating output data by performing forward propagation on the learned neural network model using the updated scaling factor data stored in the memory and the generated code data.

Images