KR20230138372A - Apparatus and method of neural networm operation - Google Patents

Abstract

Disclosed are an apparatus for a neural network arithmetic operation and a method thereof. To this end, the apparatus for a neural network arithmetic operation according to one embodiment of the present invention comprises: a receiver which receives data to perform a neural network arithmetic operation; and a processor which extracts calibration data based on learning data for allowing the neural network to learn from the data, generates a look-up table (LUT) for performing a nonlinear arithmetic operation included in the neural network by means of an auxiliary network corresponding to the layer of the neural network based on the calibration data, and updates the parameter of the LUT based on the output of the nonlinear arithmetic operation and the output of the auxiliary network.

Description

Neural network operation device and method {APPARATUS AND METHOD OF NEURAL NETWORM OPERATION}

The embodiments below relate to neural network computing devices and methods.

In the conventional neural network calculation, the method of using LUT (Look Up Table) generates errors when approximating a non-linear function in which the range of output values changes more rapidly than the amount of change in input values. In the conventional neural network calculation method, the accuracy of prediction of the final neural network model was reduced due to errors generated.

In addition, since a large amount of Look Up Table (LUT) indexes are required to approximate sections with high fluctuation rates, which increases hardware costs, a neural network calculation method that preserves accuracy while maintaining the number of LUT indexes is required. do.

However, when approximating a nonlinear function of a neural network using a single LUT, the performance of neural network inference may deteriorate because multiple layers of the neural network have different statistical characteristics of input data.

In the neural network calculation device, the neural network calculation device according to one embodiment includes a receiver that receives data for performing neural network calculation, and a calibration device based on training data for training the neural network among the data. ) Extract data, generate a LUT (Look Up Table) for performing a non-linear operation included in the neural network through an auxiliary network corresponding to a layer of the neural network based on the calibration data, and perform the non-linear operation of the non-linear operation. and a processor that updates parameters of the LUT based on the output and the output of the auxiliary network.

The processor may extract the calibration data by extracting a predetermined ratio of data from the learning data.

The processor generates a first LUT based on a first auxiliary network corresponding to the first layer of the neural network, and generates a second LUT based on a second auxiliary network corresponding to the second layer of the neural network. can do.

The processor generates a first nonlinear operation output corresponding to the first layer of the neural network based on the calibration data, and inputs the first nonlinear operation output to the second layer of the neural network to perform forward propagation. propagation) can be performed.

The processor may generate the LUT by determining the scale and bias of the LUT for approximating the non-linear operation.

The processor may fine-tune the parameters of the auxiliary network by performing back propagation based on the output of the auxiliary network and the output of the non-linear operation.

The processor may fine-tune the parameters of the auxiliary network based on the mean absolute error between the output of the auxiliary network and the output of the non-linear operation.

The processor may simultaneously learn the first layer and the second layer based on the output of the nonlinear operation.

The non-linear operation may include a GELU (Gaussian Error Linear Unit) operation, a softmax operation, or a layer normalization operation.

In a neural network calculation method, the neural network calculation method according to an embodiment includes receiving data for performing a neural network calculation, and performing calibration based on training data for training the neural network among the data. ) Extracting data, generating a LUT (Look Up Table) for performing a non-linear operation included in the neural network through an auxiliary network corresponding to a layer of the neural network based on the calibration data, and updating parameters of the LUT based on the output of the non-linear operation and the output of the auxiliary network.

The step of extracting the calibration data may include extracting the calibration data by extracting a predetermined ratio of data from the learning data.

Generating the LUT may include generating a first LUT based on a first auxiliary network corresponding to a first layer of the neural network, and based on a second auxiliary network corresponding to a second layer of the neural network. This may include generating a second LUT.

The step of updating the parameters includes generating a first non-linear operation output corresponding to the first layer of the neural network based on the calibration data, and transmitting the first non-linear operation output to the second layer of the neural network. It may include performing forward propagation by inputting.

Generating the LUT may include generating the LUT by determining a scale and bias of the LUT for approximating the non-linear operation.

The step of updating the parameters may include fine-tuning the parameters of the auxiliary network by performing backpropagation based on the output of the auxiliary network and the output of the non-linear operation.

Fine-tuning the parameters of the auxiliary network may include fine-tuning the parameters of the auxiliary network based on a mean absolute error between the output of the auxiliary network and the output of the non-linear operation. .

The step of updating the parameter may include simultaneously learning the first layer and the second layer based on the output of the non-linear operation.

The non-linear operation may include a GELU (Gaussian Error Linear Unit) operation, a softmax operation, or a layer normalization operation.

Figure 1 shows a schematic block diagram of a neural network computing device.
FIG. 2 shows the operation of the neural network computing device shown in FIG. 1.
Figure 3a is a diagram for explaining the learning process of the auxiliary network.
Figure 3b shows an example of an auxiliary network.
Figure 4 shows the performance of the neural network computing device.
Figure 5 shows a flowchart of the operation for measuring the performance of Figure 4.
FIG. 6 shows a flowchart of the operation of the neural network computing device shown in FIG. 1.

Specific structural or functional descriptions of the embodiments are disclosed for illustrative purposes only and may be changed and implemented in various forms. Accordingly, the actual implementation form is not limited to the specific disclosed embodiments, and the scope of the present specification includes changes, equivalents, or substitutes included in the technical idea described in the embodiments.

Terms such as first or second may be used to describe various components, but these terms should be interpreted only for the purpose of distinguishing one component from another component. For example, a first component may be named a second component, and similarly, the second component may also be named a first component.

When a component is referred to as being “connected” to another component, it should be understood that it may be directly connected or connected to the other component, but that other components may exist in between.

Singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, terms such as “comprise” or “have” are intended to designate the presence of the described features, numbers, steps, operations, components, parts, or combinations thereof, and are intended to indicate the presence of one or more other features or numbers, It should be understood that this does not exclude in advance the possibility of the presence or addition of steps, operations, components, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person of ordinary skill in the art. Terms as defined in commonly used dictionaries should be interpreted as having meanings consistent with the meanings they have in the context of the related technology, and unless clearly defined in this specification, should not be interpreted in an idealized or overly formal sense. No.

Hereinafter, embodiments will be described in detail with reference to the attached drawings. In the description with reference to the accompanying drawings, identical components will be assigned the same reference numerals regardless of the reference numerals, and overlapping descriptions thereof will be omitted.

Figure 1 shows a schematic block diagram of a neural network computing device.

Referring to FIG. 1, the neural network calculation device 10 can perform neural network calculation. The neural network calculation device 10 can perform calculations on non-linear functions included in neural network calculations. The neural network calculation device 10 can improve calculation speed by generating a Look Up Table (LUT) that approximates the non-linear function of the neural network and replacing the non-linear function with the LUT.

A non-linear function may mean a function other than a linear function. Nonlinear functions can be expressed in a form where the graph between the variables of the function is not a straight line. For example, non-linear functions include sigmoid (or logistic function, hyperbolic tangent, ReLU (Rectified Linear Unit) function, leaky ReLU function, parametric) It may include a ReLU function, an Exponential Linear Units (ELUs) function, a softmax function, a swish function, a Gaussian Error Linear Unit (GeLU) function, and/or a Scaled Exponential Linear Unit (SELU) function.

The neural network calculation device 10 can perform calculations on nonlinear functions using a look-up table (LUT). The neural network calculation device 10 may generate an LUT using an auxiliary network corresponding to each layer included in the neural network that performs the main calculation. The auxiliary network may be implemented in the form of a neural network.

A neural network can refer to an overall model in which artificial neurons (nodes), which form a network through the combination of synapses, change the strength of the synapse connection through learning and have problem-solving capabilities.

Neurons in a neural network can contain combinations of weights or biases. A neural network may include one or more layers consisting of one or more neurons or nodes. Neural networks can infer the results they want to predict from arbitrary inputs by changing the weights of neurons through learning.

Neural networks may include deep neural networks. Neural networks include CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), perceptron, multilayer perceptron, FF (Feed Forward), RBF (Radial Basis Network), DFF (Deep Feed Forward), and LSTM. (Long Short Term Memory), GRU (Gated Recurrent Unit), AE (Auto Encoder), VAE (Variational Auto Encoder), DAE (Denoising Auto Encoder), SAE (Sparse Auto Encoder), MC (Markov Chain), HN (Hopfield) Network), BM (Boltzmann Machine), RBM (Restricted Boltzmann Machine), DBN (Depp Belief Network), DCN (Deep Convolutional Network), DN (Deconvolutional Network), DCIGN (Deep Convolutional Inverse Graphics Network), GAN (Generative Adversarial Network) ), Liquid State Machine (LSM), Extreme Learning Machine (ELM), Echo State Network (ESN), Deep Residual Network (DRN), Differential Neural Computer (DNC), Neural Turning Machine (NTM), Capsule Network (CN), It may include Kohonen Network (KN) and Attention Network (AN).

The neural network computing device 10 may be implemented within a personal computer (PC), a data server, or a portable device.

Portable devices include laptop computers, mobile phones, smart phones, tablet PCs, mobile internet devices (MIDs), personal digital assistants (PDAs), and enterprise digital assistants (EDAs). , digital still camera, digital video camera, portable multimedia player (PMP), personal navigation device or portable navigation device (PND), handheld game console, e-book ( It can be implemented as an e-book) or a smart device. A smart device may be implemented as a smart watch, smart band, or smart ring.

The neural network calculation device 10 can perform nonlinear calculation of the neural network by generating a LUT corresponding to each layer, considering that the layers included in the neural network have different statistical characteristics of input data. . The neural network computation device 10 can improve the performance of neural network computation by processing non-linear computation using the LUT corresponding to each layer.

The neural network computing device 10 includes a receiver 100 and a processor 200. The neural network computing device 10 may further include a memory 300.

The receiver 100 may receive data for performing a neural network operation. Receiver 100 may include a receiving interface. The receiver 100 may output the received data to the processor 200. Data related to the neural network include model parameters (or weights) of the neural network, input data for performing neural network operations, data output from the neural network, training data for training the neural network, and/or related to neural network operations. May contain information.

The processor 200 may process data stored in the memory 300. The processor 200 may execute computer-readable code (eg, software) stored in the memory 300 and instructions triggered by the processor 200 .

The “processor 200” may be a data processing device implemented in hardware that has a circuit with a physical structure for executing desired operations. For example, the intended operations may include code or instructions included in the program.

For example, data processing devices implemented in hardware include microprocessors, central processing units, processor cores, multi-core processors, and multiprocessors. , ASIC (Application-Specific Integrated Circuit), and FPGA (Field Programmable Gate Array).

The processor 200 may extract calibration data from the data based on training data for training a neural network. The processor 200 may extract calibration data by extracting a predetermined ratio of data from the training data. For example, the predetermined ratio may be 1/10.

The processor 200 may generate a look up table (LUT) for performing a nonlinear operation included in the neural network through an auxiliary network corresponding to a layer of the neural network based on the calibration data. Nonlinear operations may include operations on nonlinear functions. For example, non-linear operations may include GELU (Gaussian Error Linear Unit) operations, softmax operations, or layer normalization operations.

The processor 200 may generate a LUT by determining the scale and bias of the LUT for approximating a non-linear operation. The processor 200 may generate the first LUT based on the first auxiliary network corresponding to the first layer of the neural network. The processor 200 may generate a second LUT based on a second auxiliary network corresponding to the second layer of the neural network.

The processor 200 may update the parameters of the LUT based on the output of the nonlinear operation and the output of the auxiliary network. The processor 200 may generate a first non-linear operation output corresponding to the first layer of the neural network based on the calibration data. The processor 200 may perform forward propagation by inputting the first nonlinear operation output to the second layer of the neural network.

The processor 200 may fine tune the parameters of the auxiliary network by performing backpropagation based on the output of the auxiliary network and the output of the nonlinear operation. The processor 200 may fine-tune the parameters of the auxiliary network based on the mean absolute error between the output of the auxiliary network and the output of the non-linear operation.

The processor 200 may simultaneously learn the first layer and the second layer based on the output of the nonlinear operation.

The memory 300 may store data or calculation results for calculations (eg, neural network calculations). The memory 300 may store instructions (or programs) executable by the processor 200. For example, the instructions may include instructions for executing the operation of the processor and/or the operation of each component of the processor.

The memory 300 may be implemented as a volatile memory device or a non-volatile memory device.

Volatile memory devices may be implemented as dynamic random access memory (DRAM), static random access memory (SRAM), thyristor RAM (T-RAM), zero capacitor RAM (Z-RAM), or twin transistor RAM (TTRAM).

Non-volatile memory devices include EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, MRAM (Magnetic RAM), Spin-Transfer Torque (STT)-MRAM (MRAM), and Conductive Bridging RAM (CBRAM). , FeRAM (Ferroelectric RAM), PRAM (Phase change RAM), Resistive RAM (RRAM), Nanotube RRAM (Nanotube RRAM), Polymer RAM (PoRAM), Nano Floating Gate Memory (NFGM), holographic memory, molecular electronic memory device, or insulation resistance change memory.

FIG. 2 shows the operation of the neural network computing device shown in FIG. 1.

Referring to FIG. 2, a processor (eg, processor 200 of FIG. 1) can train a neural network and an auxiliary network. The processor 200 may generate an LUT by determining the characteristics of each layer constituting the neural network. The processor 200 may learn the auxiliary network by fine-tuning the parameters of the auxiliary network based on backpropagation. The processor 200 can improve the computational performance of the neural network by generating an optimal LUT based on the finely tuned auxiliary network.

The processor 200 may generate an auxiliary network corresponding to each of a plurality of layers included in the neural network by using the weight of the auxiliary network that generates the LUT corresponding to the entire neural network as an initial value.

The processor 200 may perform fine-tuning learning based on the difference between the output of the LUT generated by the auxiliary network and the output of the non-linear operation of the neural network. The processor 200 may perform fine-tuning learning by using the average absolute error between the output of the LUT and the output of the nonlinear operation of the neural network as a loss function.

Through this, the processor 200 can naturally learn the neural network and/or the auxiliary neural network by performing backpropagation to reduce the loss function without analyzing statistical characteristics of the input data for each layer.

The neural network may include a first layer (210-1), a second layer (210-2), ..., an N-th layer N (210-N). Here, N may mean a natural number. Each layer may include non-linear operations. The first auxiliary network 250-1, the second auxiliary network 250-2, ..., the Nth auxiliary network N (250-N) may correspond to each layer. The processor 200 may generate and learn the first LUT 230-1, the second LUT 2 230-2, ..., the N-th LUT 230-N through a plurality of auxiliary networks. .

The processor 200 may generate a plurality of LUTs 230-1, 230-2, ..., 230-N by determining the scale and bias of the LUT for approximating a non-linear operation. . The processor 200 may generate the first LUT 230-1 based on the first auxiliary network 250-1 corresponding to the first layer 210-1 of the neural network. The processor 200 may generate the second LUT 230-2 based on the second auxiliary network 250-2 corresponding to the second layer 210-1 of the neural network.

The processor 200 may update parameters of the plurality of LUTs 230-1, 230-2, ..., 230-N based on the output of the non-linear operation and the output of the auxiliary network. The processor 200 may simultaneously learn the first layer 210-1 and the second layer 210-2 based on the output of the non-linear operation.

The processor 200 may generate a first non-linear operation output corresponding to the first layer 210-1 of the neural network based on the calibration data. The processor 200 may perform forward propagation by inputting the first nonlinear operation output to the second layer 210-2 of the neural network.

The processor 200 may fine-tune the parameters of the auxiliary network by performing backpropagation based on the output of the auxiliary network and the output of the nonlinear operation. The processor 200 may fine-tune the parameters of the auxiliary network based on the mean absolute error between the output of the auxiliary network and the output of the non-linear operation.

FIG. 3A is a diagram for explaining the learning process of an auxiliary network, and FIG. 3B shows an example of an auxiliary network.

Referring to FIGS. 3A and 3B, a processor (eg, processor 200 of FIG. 1) may generate an LUT corresponding to a nonlinear operation of a neural network. The processor 200 may generate a LUT by configuring an auxiliary network.

The example in FIG. 3A is a diagram for explaining a learning process using the auxiliary network 310 and nonlinear operation 330 corresponding to one layer. Learning can also be performed for other auxiliary networks and layers in the same manner as described later.

The processor 200 may use the weight of one auxiliary network 310 for which training has been completed in advance as an initial value of the weight of the auxiliary network 310 corresponding to all layers.

The processor 200 may extract calibration data from data based on training data for training a neural network. The processor 200 may extract calibration data by extracting a predetermined ratio of data from the training data. Since the calibration data is extracted from the training data of the neural network, it may have statistically similar characteristics to the training data.

The processor 200 may perform fine tuning of the auxiliary network 310 based on the extracted calibration data.

The processor 200 may finely adjust the parameters of the auxiliary network 310 by performing backpropagation based on the output of the auxiliary network 310 and the output of the nonlinear operation 330. The processor 200 may finely adjust the parameters of the auxiliary network 310 based on the mean absolute error (MAE) between the output of the auxiliary network 310 and the output of the nonlinear operation 330.

As the layers of the neural network become deeper, errors resulting from the approximation of the nonlinear operation 330 may accumulate. The processor 200 can prevent error accumulation caused by learning multiple layers simultaneously by using the output of the nonlinear operation 330 as the output transmitted to the next layer. Through this, the processor 200 can improve learning efficiency because it can learn all layers of the neural network at once without accumulating errors.

The auxiliary network 310 may be in the form of a neural network including one hidden layer. When generating a LUT of N-entries, the auxiliary network 310 may be composed of N-1 neurons. At this time, ReLU may be used as the nonlinear function.

The first layer of the auxiliary network 310 may include both weight and bias parameters, and the second layer may include only the weight parameters.

Figure 4 shows the performance of the neural network computing device.

Referring to FIG. 4, a processor (e.g., processor 200 of FIG. 1) can improve the performance of neural network operations by generating a LUT through fine tuning of the auxiliary network. The example of FIG. 4 may indicate that the LUT that has performed fine-tuning shows superior performance compared to the baseline and the LUT that has not performed fine-tuning.

The example of FIG. 4 may represent the score for each task when the layer normalization operation, which is a non-linear operation of the RoBERTa model, is approximated with a LUT.

The baseline column represents the performance when the nonlinear operation of the RoBERTa model is used as is, the LUT column represents the performance when the nonlinear operation is replaced with one LUT, and the fine-tuned LUT column corresponds to each layer. Performance can be shown when using a LUT that has been fine-tuned using an auxiliary network.

The processor 200 performs fine tuning to generate an optimized LUT for each layer, thereby recovering performance degradation that occurs when replacing nonlinear operations using only one LUT in most tasks.

The processor 200 can perform learning more efficiently by simultaneously performing fine-tuning on a plurality of layers in parallel, compared to a method of sequentially performing fine-tuning for each layer.

Figure 5 shows a flowchart of the operation for measuring the performance of Figure 4.

Referring to FIG. 5, a processor (e.g., processor 200 in FIG. 1) places an auxiliary network for generating a LUT in each layer constituting the neural network to generate a LUT corresponding to the nonlinear operation of the neural network. Can (510). At this time, the weight of one auxiliary network learned in advance can be used as the initial value of the weight of the auxiliary network for all layers.

The processor 200 may perform forward propagation learning using the calibration data (530). The processor 200 may use part of the learning data of the neural network as calibration data. For example, the processor 200 may use 1/10 of the training data of the neural network as calibration data. The processor 200 may transfer the output of the nonlinear operation to the next layer.

The processor 200 may generate an LUT by performing backpropagation learning based on the output of the auxiliary network and the output of the nonlinear operation (550). The processor 200 may generate the LUT by fine-tuning the weights of the auxiliary network (570).

The processor 200 may measure the accuracy of the task using the finely tuned auxiliary network (590). The processor 200 may measure task accuracy through a finely tuned auxiliary network by repeatedly performing operations 530 to 570. Processor 200 may update parameters of the neural network and/or auxiliary network through fine-tuning.

Through the above-described fine-tuning method, the processor 200 can perform fine-tuning on the neural network and/or auxiliary network without requiring a label using regression.

The processor 200 can prevent the accumulation of errors that occur during fine tuning by transferring the output of the nonlinear function to the input of the next layer.

FIG. 6 shows a flowchart of the operation of the neural network computing device shown in FIG. 1.

Referring to FIG. 6, a receiver (e.g., receiver 100 of FIG. 1) may receive data for performing a neural network operation (610).

A processor (e.g., processor 200 in FIG. 1) may extract calibration data based on training data for training a neural network from data (630). The processor 200 may extract calibration data by extracting a predetermined ratio of data from the training data. For example, the predetermined ratio may be 1/10.

The processor 200 may generate an LUT for performing a nonlinear operation included in the neural network through an auxiliary network corresponding to a layer of the neural network based on the calibration data (650). For example, non-linear operations may include GELU operations, softmax operations, or layer normalization operations.

The processor 200 may generate a LUT by determining the scale and bias of the LUT for approximating a non-linear operation. The processor 200 may generate the first LUT based on the first auxiliary network corresponding to the first layer of the neural network. The processor 200 may generate a second LUT based on a second auxiliary network corresponding to the second layer of the neural network.

The processor 200 may update the parameters of the LUT based on the output of the nonlinear operation and the output of the auxiliary network (670). The processor 200 may generate a first non-linear operation output corresponding to the first layer of the neural network based on the calibration data. The processor 200 may perform forward propagation by inputting the first nonlinear operation output to the second layer of the neural network.

The processor 200 may fine-tune the parameters of the auxiliary network by performing backpropagation based on the output of the auxiliary network and the output of the nonlinear operation. The processor 200 may fine-tune the parameters of the auxiliary network based on the average absolute error between the output of the auxiliary network and the output of the non-linear operation.

The processor 200 may simultaneously learn the first layer and the second layer based on the output of the nonlinear operation.

The embodiments described above may be implemented with hardware components, software components, and/or a combination of hardware components and software components. For example, the devices, methods, and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, and a field programmable gate (FPGA). It may be implemented using a general-purpose computer or a special-purpose computer, such as an array, programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and software applications running on the operating system. Additionally, a processing device may access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, a single processing device may be described as being used; however, those skilled in the art will understand that a processing device includes multiple processing elements and/or multiple types of processing elements. It can be seen that it may include. For example, a processing device may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are possible.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. Software and/or data may be used on any type of machine, component, physical device, virtual equipment, computer storage medium or device to be interpreted by or to provide instructions or data to a processing device. , or may be permanently or temporarily embodied in a transmitted signal wave. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on a computer-readable recording medium.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. A computer-readable medium may include program instructions, data files, data structures, etc., singly or in combination, and the program instructions recorded on the medium may be specially designed and constructed for the embodiment or may be known and available to those skilled in the art of computer software. It may be possible. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -Includes optical media (magneto-optical media) and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

The hardware devices described above may be configured to operate as one or multiple software modules to perform the operations of the embodiments, and vice versa.

As described above, although the embodiments have been described with limited drawings, those skilled in the art can apply various technical modifications and variations based on this. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

Therefore, other implementations, other embodiments, and equivalents of the claims also fall within the scope of the claims described below.

Claims (19)

In the neural network computing device,
A receiver that receives data for performing neural network operations;
Extracting calibration data from the data based on training data for training the neural network,
Based on the calibration data, generate a LUT (Look Up Table) for performing a non-linear operation included in the neural network through an auxiliary network corresponding to a layer of the neural network,
A processor that updates parameters of the LUT based on the output of the non-linear operation and the output of the auxiliary network.
A neural network computing device including a.
According to paragraph 1,
The processor,
Extracting the calibration data by extracting a predetermined ratio of data from the learning data,
Neural network computing device.
According to paragraph 1,
The processor,
Generate a first LUT based on a first auxiliary network corresponding to the first layer of the neural network,
Generating a second LUT based on a second auxiliary network corresponding to the second layer of the neural network,
Neural network computing device.
According to paragraph 1,
The processor,
Generating a first non-linear operation output corresponding to a first layer of the neural network based on the calibration data,
Performing forward propagation by inputting the first nonlinear operation output to a second layer of the neural network,
Neural network computing device.
According to paragraph 1,
The processor,
Generating the LUT by determining a scale and bias of the LUT for approximating the non-linear operation,
Neural network computing device.
According to paragraph 1,
The processor,
Fine-tuning the parameters of the auxiliary network by performing back propagation based on the output of the auxiliary network and the output of the nonlinear operation,
Neural network computing device.
According to clause 6,
The processor,
Fine-tuning the parameters of the auxiliary network based on the mean absolute error between the output of the auxiliary network and the output of the non-linear operation,
Neural network computing device.
According to paragraph 3,
The processor,
Simultaneously learning the first layer and the second layer based on the output of the nonlinear operation,
Neural network computing device.
According to paragraph 1,
The nonlinear operation is,
Including GELU (Gaussian Error Linear Unit) operation, softmax operation, or layer normalization operation,
Neural network computing device.
In the neural network calculation method,
Receiving data for performing a neural network operation;
extracting calibration data from the data based on training data for training the neural network;
generating a Look Up Table (LUT) for performing a non-linear operation included in the neural network through an auxiliary network corresponding to a layer of the neural network based on the calibration data; and
Updating parameters of the LUT based on the output of the non-linear operation and the output of the auxiliary network.
Neural network calculation method including.
According to clause 10,
The step of extracting the calibration data is,
Extracting the calibration data by extracting a predetermined ratio of data from the learning data.
Neural network calculation method including.
According to clause 10,
The step of generating the LUT is,
generating a first LUT based on a first auxiliary network corresponding to a first layer of the neural network; and
generating a second LUT based on a second auxiliary network corresponding to the second layer of the neural network
Neural network calculation method including.

According to clause 10,
The step of updating the parameters is,
generating a first non-linear operation output corresponding to a first layer of the neural network based on the calibration data; and
Performing forward propagation by inputting the first nonlinear operation output to a second layer of the neural network
Neural network calculation method including. According to clause 10,
The step of generating the LUT is,
Generating the LUT by determining a scale and bias of the LUT for approximating the non-linear operation.
Neural network calculation method including.
According to clause 10,
The step of updating the parameters is,
Fine-tuning the parameters of the auxiliary network by performing backpropagation based on the output of the auxiliary network and the output of the non-linear operation.
Neural network calculation method including.
According to clause 15,
The step of fine-tuning the parameters of the auxiliary network is,
Fine-tuning the parameters of the auxiliary network based on the mean absolute error between the output of the auxiliary network and the output of the non-linear operation.
Neural network calculation method including.
According to clause 12,
The step of updating the parameters is,
Simultaneously learning the first layer and the second layer based on the output of the nonlinear operation
Neural network calculation method including.
According to clause 10,
The nonlinear operation is,
Including GELU (Gaussian Error Linear Unit) operation, softmax operation, or layer normalization operation,
Neural network calculation method.
A computer program stored in a computer-readable medium in combination with hardware to execute the method of any one of claims 10 to 18.

Images