Deep learning machine and operation method thereof

Abstract

A deep learning machine includes a classification unit having a labeling criterion and configured to label input data according to the labeling criterion, a conversion unit configured to integerize first-class input data labeled as requiring integerization among the input data labeled by the classification unit, a first learning data unit configured to receive the first-class input data integerized by the conversion unit and infer output data, and a second learning data unit configured to receive second-class input data labeled as not requiring integerization and infer the output data.

Description

Technical Field

[0001] This invention relates to a deep learning machine and its operation method, which can reduce the amount of computation required to process complex input data, while preventing feature loss due to integer operations, and maximizes the performance of deep learning by using additional classification deep learning to distinguish input data. This does not lead to a decrease in the recognition performance based on floating-point operations in deep learning inference based on integer operations. Background Technology

[0002] Machine learning emerged to address the challenge of finding rules to distinguish objects, such as in image recognition. Machine learning is a technique that uses known features of the data to be learned as a foundation, primarily employing statistical methods to receive and represent the data, and then generalizes this representation to unknown data. To use machine learning rules obtained in this way, features of the input data need to be extracted. Deep learning techniques can be used in this process.

[0003] However, this deep learning technique requires a large amount of computation on complex and massive input data. Therefore, the recent trend is to convert floating-point operations to integers and embed integer operations to drive deep learning techniques in resource-constrained systems, such as smartphones.

[0004] However, unlike floating-point based algorithms, integer-based algorithms suffer from performance degradation during the quantization of unprocessed data. This performance degradation is particularly pronounced when data is concentrated in specific areas, i.e., extreme ends, during the quantization process. Therefore, technologies are needed to address this issue.

[0005] It should be understood that the matters described in the above-mentioned related technologies are only used to facilitate the understanding of the background of the present invention and should not be regarded as prior art known to those skilled in the art. Summary of the Invention

[0006] This overview introduces selected concepts in a simplified form, which will be further described in detail below. This overview is not intended to identify key or essential features of the claimed subject matter, nor is it intended to help determine the scope of protection of the claimed subject matter.

[0007] Therefore, in view of the above problems, the present invention is proposed. The purpose of the present invention is to provide a deep learning machine and its operation method, which can reduce the amount of computation required to process complex input data, prevent feature value loss due to integer operations, and maximize the performance of deep learning by adding classification deep learning to distinguish input data. In deep learning inference based on integer operations, it does not lead to a decrease in recognition performance based on floating-point operations.

[0008] In one general aspect, a deep learning machine includes: a classification unit having a labeling criterion and for labeling input data according to the labeling criterion; a transformation unit for integerizing first-class input data labeled by the classification unit as requiring integerization; a first learning data unit for receiving the first-class input data integerized by the transformation unit and inferring output data; and a second learning data unit for receiving second-class input data labeled as not requiring integerization and inferring output data.

[0009] The labeling standard can be provided based on an inference performance value derived by receiving integerized input data and inferring the output data.

[0010] The output data can be inferred based on the third quantization learning data, which is obtained by deep learning and quantization of the input data.

[0011] The labeling standard can be provided based on fourth quantization learning data obtained through deep learning and quantization rules, which are rules about the difference between the inference performance value of the input data and the inference performance value of the integerized input data.

[0012] A classification unit may have at least one labeling criterion corresponding to the category of the input data.

[0013] The first learning data unit can infer the output data based on the first learning data, which is derived by deep learning on the input data labeled as the first class.

[0014] The first learning data unit can infer the output data based on the first quantized learning data, which is obtained by quantizing the first learning data.

[0015] The second learning data unit can infer the output data based on the second quantized learning data, which is obtained by deep learning and quantization of the input data labeled as the second type.

[0016] In another general aspect, a deep learning machine operation method includes: a classification unit having a labeling standard receiving input data and labeling the input data according to the labeling standard; a transformation unit integerizing first type input data that is labeled as needing integerization in the input data labeled by the classification unit; and a first learning data unit receiving the integerized first type input data and inferring output data.

[0017] The labeling standard can be provided based on an inference performance value derived by receiving the integerized input data and inferring the output data based on a third quantization learning data obtained by deep learning and quantization of the input data.

[0018] The labeling standard can be provided through deep learning fourth quantization learning data, which is obtained through deep learning and quantization rules that are rules about the difference between the inference performance value of the input data and the inference performance value of the integerized input data.

[0019] When labeling in this classification unit, labeling can be performed according to at least one labeling standard corresponding to the category of the input data.

[0020] When inferring the output data in the first learning data unit, the output data can be inferred based on the first quantized learning data, which is obtained by deep learning and quantization of the input data labeled as the first class.

[0021] The deep learning machine operation method may include: a second learning data unit receiving input data that is labeled as not requiring integerization of a second class, and inferring output data after the first learning data unit infers the output data.

[0022] When inferring the output data in the second learning data unit, the output data can be inferred based on the second quantized learning data, which is obtained by deep learning and quantization of the input data labeled as the second type.

[0023] Other features and aspects will become apparent from the following detailed description, drawings and claims. Attached Figure Description

[0024] Figure 1 This is a structural diagram of a deep learning machine according to an embodiment of the present invention.

[0025] Figure 2 This is a block diagram of a deep learning machine according to an embodiment of the present invention.

[0026] Figure 3 This is a flowchart of a deep learning machine computation method according to an embodiment of the present invention. Detailed Implementation

[0027] The descriptions of the specific structures and functions of embodiments of the present invention disclosed in this specification or application are intended to describe embodiments of the present invention. Embodiments of the present invention can be implemented in various forms and should not be construed as limited to the embodiments described in this specification and application. Embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] The "~ unit" used in this embodiment can be implemented as software such as a task, class, subroutine, process, object, execution thread, program, program-based artificial intelligence (AI) accelerator, or a combination of software executing in a predetermined area of ​​memory, or hardware such as an application-specific integrated circuit (ASIC), digital signal processor (DSP), vision processing unit (VPU), neuromorphic IC, or hardware-based AI accelerator. Furthermore, the "~ unit" in this embodiment can be implemented through software or a combination of software and hardware. Additionally, the "~ unit" can be contained in a computer-readable storage medium. Moreover, the "~ unit" in this embodiment can be distributed across multiple software or hardware components or combinations thereof. In this case, the data processing of the "~ unit" in this embodiment can be distributed, centralized, or accelerated through cloud computing, edge computing, or AI edge computing.

[0029] Furthermore, the term "deep learning" used in this embodiment is a type of machine learning and can be used interchangeably with it. Machine learning is a technique involving systems and algorithms that learn and reason based on empirical data to improve their performance. In other words, machine learning is a method of building specific models to infer output data based on input data. Here, deep learning includes hidden data used to extract features from input data by learning parameters such as weights and biases, as well as activation functions in algorithms such as artificial neural networks, and transforming the extracted features into output data between input and output data, thereby serving as a method for building specific models.

[0030] Therefore, learning in this specification can refer to the process of using learning data to determine learning parameters for the purposes of a corresponding system, such as classification, regression analysis, and clustering of input data. Furthermore, the concept of learning in this specification includes learning itself (e.g., training or retraining) or reasoning (e.g., prediction or determination). Of course, deep learning in this specification includes learning using regressive artificial neural networks (such as convolutional neural networks (CNNs)) and machine learning using probability-based, geometry-based, and ensemble-based algorithms. In addition, deep learning in this specification includes supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning.

[0031] Figure 1This is a structural diagram of a deep learning machine according to an embodiment of the present invention. Figure 2 This is a block diagram of a deep learning machine according to an embodiment of the present invention. Figure 3 This is a flowchart of a deep learning machine computation method according to an embodiment of the present invention.

[0032] refer to Figure 1 This is a structural diagram of a deep learning machine according to an embodiment of the present invention. The deep learning machine according to an embodiment of the present invention includes: a classification unit D, which has a labeling standard and labels input data according to the labeling standard; a conversion unit C, which integerizes the first type of input data labeled by the classification unit as needing to be integerized; a first learning data unit T1, which receives the first type of input data integerized by the conversion unit and infers output data; and a second learning data unit T2, which receives the second type of input data labeled as not needing to be integerized and infers output data.

[0033] Figure 1 Input unit I can receive a deep learning model, training data used to train or learn the deep learning model, retraining data, input data used to obtain output using the learned model, and output history or user usage history after output. Input unit I can acquire unprocessed input data and preprocessed input data for learning the deep learning model. Here, preprocessing can refer to extracting input features from the input data. Furthermore, the input data of input unit I can be data sent / received to / from other devices or users via wired / wireless communication or interfaces. Data types can include image information, audio information, corresponding signals, data, and user input information.

[0034] Figure 1 The input data of input unit I is labeled in classification unit D as either a first class requiring integerization or a second class not requiring integerization. Classification unit D may have a labeling criterion. In conversion unit C, the input data labeled as requiring integerization in the first class is integerized. First learning data unit T1 infers output data from the integerized first-class input data using a deep learning model. Furthermore, second learning data unit T2 infers output data from the second-class input data labeled as not requiring integerization using a deep learning model without going through conversion unit C, or infers output data from the second-class input data labeled as not requiring integerization using conversion unit C (not shown), which converts the input data to floating-point data.

[0035] Here, integerization is a type of quantization, a model compression technique that reduces model size without losing information representation, thereby reducing computational load and improving computational speed and efficiency. Furthermore, in this specification, integerization means converting input data, including floating-point data, into integer data, including fixed-point or integer data. Specifically, integerization can refer to converting input data into data suitable for integer operators; in other words, the operators cannot handle floating-point data.

[0036] Here, input data requiring integerization may refer to input data whose information representation (e.g., the resolution of image information) has been reduced to or exceeds a predetermined reference value, or whose learning performance value or inference performance value has been reduced to or exceeds a predetermined reference value. Furthermore, the first category requiring integerization and the second category not requiring integerization are incompatible. However, the first category can include multiple types, such as a first category with a higher integerization requirement and a first category with a lower integerization requirement. Similarly, the second category can include multiple types, such as a second category with a greater degree of no integerization requirement and a second category with a lesser degree of no integerization requirement.

[0037] Figure 1 The control unit M controls the device based on output data inferred from either the first learning data unit T1 or the second learning data unit T2. Furthermore, Figure 1 The control unit M can control the learning amount, learning speed, inference amount, and inference speed of the first learning data unit T1 or the second learning data unit T2. For example, the control unit M can control the power generation device, power transmission device, braking device, or steering device of autonomous driving, the electrical device that facilitates autonomous driving, or its peripheral devices, based on output data obtained by extracting features from images measured in the vehicle. Here, the control unit M can control the device to prioritize learning or output speed or distance information of obstacles in the image, thereby performing high-priority functions regarding vehicle collision risk.

[0038] Figure 1 The control unit M can determine the optimal control point for performing a specific function by combining the data stored in the storage unit S and the output data during the control steps. Figure 1 The storage unit S can store the programs and / or data used by the components that make up the deep learning machine. Furthermore, Figure 1 The storage unit S can store learning parameters for learning, quantization parameters for quantization, input data, and output data.

[0039] Deep learning requires extensive computation on complex and massive input data. Therefore, introducing deep learning into resource-constrained systems such as smartphones can significantly slow down computation. This problem becomes particularly severe when the input data is unprocessed or floating-point data. Thus, to drive deep learning in such systems, it is necessary to integerize and embed floating-point operations. However, unlike algorithms based on floating-point input data, deep learning algorithms based on integer input data suffer a performance degradation during the integerization process of unprocessed data. Specifically, this performance degradation occurs significantly when the data is concentrated in a specific region, i.e., at extreme endpoints, during the integerization of data values ​​and the quantization of the learning data obtained by learning those values.

[0040] Therefore, to solve this problem, the deep learning machine according to embodiments of the present invention adds a classification network and combines it with the deep learning model. That is, in this embodiment, the input data is divided into a first type of input data and a second type of input data, namely, input data that needs to be integerized and input data that does not need to be integerized. The input data that needs to be integerized is integerized and inference is performed based on its features, while the input data that does not need to be integerized is inferred based on its features. Therefore, the deep learning machine according to embodiments of the present invention can ensure high computational speed for inference through integer-based input data, while maintaining and improving the performance and accuracy of inference through floating-point-based input data.

[0041] refer to Figure 1 This is a structural diagram of a deep learning machine according to an embodiment of the present invention. A labeling standard can be provided based on an inference performance value derived by receiving integerized input data and inference output data. Furthermore, the inference performance value can be derived by inferring output data based on third quantized learning data, which is obtained from deep learning and quantized input data. That is, in an embodiment of the present invention, the inference performance value can be determined using preprocessed learning data, and a labeling standard can be provided based on the inference performance value in a step prior to the integerized inference step.

[0042] Figure 1 The third learning data unit T3 receives unintegerized input data (not shown) from the storage unit S or the input unit I, and performs deep learning on it to generate third learning data. Furthermore, the third learning data unit T3 quantizes the third learning data to generate third quantized learning data, and receives both unintegerized and integerized input data through the third quantized learning data, and infers output data. Therefore, a labeling standard can be provided for the classification unit D based on the inference performance of the third learning data unit T3 and the inference performance based on inference on integerized input data (not shown).

[0043] While such inference performance values ​​can be supervised, this is undesirable for machine learning purposes. Specifically, in image detection and cell segmentation, these inference performance values ​​can be calculated using the mean intersection-over-union (mIOU). In this case, inference can be performed based on third-level learning data or third-level quantized learning data obtained through deep learning on unintegerized input data. This can lead to increased complexity, thus requiring the generation of a simple classification network. Inference performance is improved when input data requiring or not requiring integerization is labeled according to the inference performance value actively measured by a simple classification network. Therefore, the deep learning machine according to embodiments of the invention has a labeling criterion based on an inference performance value that is based on third-level quantized learning data, thereby enabling the definition of a criterion for input data requiring integerization and improving its inference performance based on that criterion.

[0044] refer to Figure 1 This is a structural diagram of a deep learning machine according to an embodiment of the present invention. It can provide a labeling standard based on fourth quantization learning data, which is obtained from deep learning and quantization rules. These rules relate to the difference between the inference performance value of the input data and the inference performance value of the integerized input data. That is, in this embodiment, the labeling standard is derived through deep learning based on the fourth quantization learning data.

[0045] Figure 1 The fourth learning data unit T4 receives inference performance values, which are obtained by receiving unintegerized input data (not shown) from the storage unit S or the input unit I and inferring output data. This inference can be performed in the third learning data unit T3 described above. The fourth learning data unit T4 receives integerized input data and the inference performance values ​​of the input data from the third learning data unit T3, and performs deep learning on rules that cause a decrease in inference performance due to integerization of the input data. The fourth learning data unit T4 obtains a labeling standard based on the fourth quantized learning data obtained by quantizing the rule, and updates the labeling standard. Therefore, the deep learning machine according to the embodiment of the present invention performs deep learning on input data that requires integerization, and updates the standard by using a labeling standard obtained based on the inference performance values ​​of the learning data including the third quantized learning data. Therefore, the classification performance of the machine is improved, and the corresponding integer-based inference performance is also improved.

[0046] Figure 2 This is a block diagram of a deep learning machine according to an embodiment of the present invention. The classification unit D may have at least one labeling criterion corresponding to the category of the input data. Input to... Figure 2The input data of input unit I can include image data. In this case, the image data can be the sum of still image data segmented by time and audio data. However, as... Figure 2 As shown, devices that receive such images in real time (short-range (SR) radar, long-range (LR) radar, lidar (light detection and ranging, LiDAR), SVM cameras, and cameras) receive different types of image data depending on their purpose and function. Therefore, Figure 2 The classification unit D can have different labeling standards applicable to the input data included in the image data of the device input to the input unit I. These labeling standards can be provided by setting the aforementioned fourth learning data unit T4 differently for each device. Therefore, the control unit M can control the relevant devices according to the labeling standards adapted to each fourth learning data unit T4, based on the reasoning of the first learning data unit T1 and the second learning data unit T2.

[0047] Therefore, the deep learning machine according to embodiments of the present invention can optimize the amount and type of information in the input data, the amount and type of information in the output data suitable for the amount and type of information in the input data, and improve it by providing a labeling standard corresponding to the category of the input data. Specifically, the output data of each of the above-described devices can be improved for classification and localization, image detection, semantic segmentation, and object segmentation that are important in image information. For example, each device may need a low-level stage for collision warning and emergency braking and a high-level stage for maintaining vehicle distance. This embodiment can optimize for the more important stage between the two. Therefore, if the labeling of the input data has been optimized according to the improvement, the deep learning machine according to embodiments of the present invention can control the reduction of the learning amount of input data that does not need improvement or the increase of the learning amount of input data that needs improvement. Therefore, in this embodiment, integer operations are optimized to suit the image information and the necessary output information, thereby enabling precise control and fast computation.

[0048] refer to Figure 1 This is a structural diagram of a deep learning machine according to an embodiment of the present invention. A first learning data unit T1 can infer output data based on first learning data, which is derived through deep learning of a first type of input data. Furthermore, the first learning data unit can infer output data based on first quantized learning data obtained by quantizing the first learning data. Additionally, a second learning data unit can infer output data based on second quantized learning data, which is obtained through deep learning and quantization of a second type of input data. In other words, the embodiments of the present invention perform learning and inference suitable for each type.

[0049] Learning data obtained through deep learning can be provided using either floating-point or integer-based input data. However, inference performance is high and computation speed is fast when performing integer-based inference based on first and second learning data, which are obtained through deep learning using a first class that needs to be integerized and a second class that does not need to be integerized as input data suitable for integer operations. Quantizing the first learning data is advantageous in this process. Therefore, the deep learning machine according to embodiments of the present invention can improve the performance of integer-based inference by performing inference on first / second learning data obtained through relearning (quantizing) input data features based on label criteria classification.

[0050] Figure 3 This is a flowchart of a deep learning machine operation method according to an embodiment of the present invention. The deep learning machine operation method according to the present invention includes: step S200, in which a classification unit with a labeling standard receives input data and labels the input data according to the labeling standard; step S300, in which a conversion unit integerizes the first type of input data in the labeled input data that is labeled as needing integerization; and step S500, in which a first learning data unit receives the integerized first type of input data and infers output data.

[0051] In the first step of the deep learning machine operation method according to an embodiment of the present invention, input data, integerized input data, learning data, quantized learning data, and a learning inference model are included, such as... Figure 3 (S100) shows the input. This input can be performed in the input unit I or the storage unit S. In the next step, the classification unit labels the input data into a first class requiring integerization and a second class not requiring integerization, respectively, according to a labeling criterion (S200). This labeling criterion can be provided by deriving third quantization learning data or by inference performance values ​​based on the third quantization learning data (S202). Alternatively, the labeling criterion can be provided by deep learning fourth learning data (not shown) or fourth quantization learning data, which is derived through deep learning rules regarding the differences between inference performance values ​​(S204).

[0052] In the next step, the conversion unit converts the first type of input data, which is marked as needing integerization, into integers (S300). Simultaneously, the conversion unit can convert the second type of input data, which is marked as not needing integerization, into floating-point data (S302). In the final step, the first learning data unit infers output data based on the integerized first type of input data (S500). Furthermore, the second learning data unit can infer output data based on the integerized second type of input data (S502).

[0053] Before the final step, the first learning data unit can generate first quantized learning data by performing deep learning and quantization on the input data labeled as the first class (S400). Therefore, inference in the final step can be performed based on the first quantized learning data. Similarly, before the final step, the second learning data unit can generate second quantized learning data by performing deep learning and quantization on the input data labeled as the second class (S402). Therefore, inference in the final step can be performed based on the second quantized learning data.

[0054] Therefore, the deep learning machine computation method according to embodiments of the present invention adds a classification network and combines it with a deep learning model. Thus, the deep learning machine computation method according to embodiments of the present invention can ensure high computational speed for inference using integer-based input data, while maintaining and improving the performance and accuracy of inference using floating-point-based input data.

[0055] exist Figure 3 In the flowchart of the deep learning machine operation method according to an embodiment of the present invention, labeling can be performed according to a labeling standard provided based on an inference performance value. This inference performance value is obtained by inferring output data from input data that has been integerized based on third quantized learning data, which is obtained by performing deep learning and quantization (S202) on the input data labeled in the labeling step (S200) performed in the classification unit. Inference performance is improved when input data requiring integerization or input data that does not require integerization is labeled according to the inference performance value actively measured through this simple classification network. Therefore, the deep learning machine operation method according to an embodiment of the present invention provides a labeling standard based on the inference performance value based on third quantized learning data, thereby clarifying the standard for input data requiring integerization and improving inference performance according to this standard.

[0056] exist Figure 3 In the flowchart of the deep learning machine operation method according to an embodiment of the present invention, labeling can be performed according to a labeling standard provided by deep learning fourth quantization learning data, which is obtained by deep learning and quantization rules. These rules relate to the difference between the inference performance value of the input data and the inference performance value of the integerized input data (S204) in the labeling step (S200) performed in the classification unit. In this step, the labeling standard is updated by providing a labeling standard based on the fourth quantization learning data. Therefore, the deep learning machine according to the embodiment of the present invention performs deep learning and updates the standard using a labeling standard provided by the inference performance value of learning data including third quantization learning data. Thus, the machine's classification performance is improved, and its corresponding integer-based inference performance is also enhanced.

[0057] exist Figure 3 In the flowchart of the deep learning machine operation method according to an embodiment of the present invention, the labeling step (S200) performed in the classification unit can label the data according to at least one labeling criterion corresponding to the category of the input data. Accordingly, the deep learning machine operation method according to the embodiment of the present invention can optimize the information content and type of the input data, the information content and type of the output data suitable for the information content and type of the input data, and improve it by providing a labeling criterion corresponding to the category of the input data. Therefore, the deep learning machine operation method according to the embodiment of the present invention optimizes integer operations to adapt to image information and necessary output information, thereby achieving precise control and fast computation.

[0058] exist Figure 3 In the flowchart of the deep learning machine operation method according to an embodiment of the present invention, in the step of inferring output data in the first learning data unit (S500), output data can be inferred based on first quantized learning data, which is obtained through deep learning and quantization of input data labeled as a first type (S400). Furthermore, after the step of inferring output data in the first learning data unit (S500), the second learning data unit can receive second type input data labeled as not requiring integerization and infer output data (S502). Additionally, in the step of inferring output data in the second learning data unit (S502), output data can be inferred based on second quantized learning data, which is obtained through deep learning and quantization of input data labeled as a second type (S402).

[0059] The deep learning machine computation method according to embodiments of the present invention is based on integer arithmetic in the inference step. When performing integer arithmetic-based inference based on first learning data and second learning data, the inference performance is high and the computation speed is fast. The first and second learning data are obtained through deep learning using a first class requiring integerization and a second class not requiring integerization as input data suitable for integer arithmetic. Quantizing the first learning data is advantageous in this process. Therefore, the deep learning machine computation method according to embodiments of the present invention can improve the performance of integer arithmetic-based inference by performing inference on the first / second learning data obtained through relearning (quantization) of input data features based on labeling criteria classification.

[0060] This invention relates to a deep learning machine and its operation method, and more specifically, to a deep learning machine and its operation method that additionally uses classification deep learning to distinguish input data, which does not cause a decrease in recognition performance based on floating-point operations in deep learning inference based on integer operations.

[0061] The deep learning technology of this invention requires extensive computation on complex and massive input data. Therefore, introducing deep learning into resource-constrained systems such as smartphones significantly reduces computational speed. Thus, to drive deep learning in such systems, it is necessary to integerize floating-point operations. However, unlike algorithms based on floating-point input data, deep learning algorithms based on integer input data suffer a decline in recognition performance during the integerization of unprocessed data. In particular, when data is concentrated in a specific region, i.e., at extreme endpoints, this performance degradation occurs significantly during the integerization of data values ​​and the quantization of the learning data obtained by learning those values.

[0062] Accordingly, to address this problem, the deep learning machine and its operation method according to embodiments of the present invention add a classification network and combine it with the deep learning model. That is, the deep learning machine and its operation method according to this embodiment divide the input data into a first type of input data and a second type of input data, namely, input data requiring integerization and input data not requiring integerization. The input data requiring integerization is integerized and inference is performed based on its features, while the input data not requiring integerization is inferred based on its features before inference using the deep learning model. Therefore, the deep learning machine and its operation method according to embodiments of the present invention can ensure high inference speed through integer-based input data, while maintaining and improving the performance and accuracy of inference through floating-point-based input data.

[0063] The deep learning machine and its operation method according to the present invention can reduce the amount of computation required to process complex input data, while preventing feature loss due to integer operations, and can maximize the performance of deep learning by using additional classification deep learning to distinguish input data. In deep learning inference based on integer operations, it does not lead to a decrease in recognition performance based on floating-point operations.

[0064] Although specific embodiments of the invention have been described and illustrated above, those skilled in the art will understand that various modifications and variations can be made without departing from the spirit and scope of the invention as disclosed in the appended claims.

Claims

1. A deep learning machine, comprising: A classification unit having a labeling criterion and for labeling input data into a first class requiring integerization and a second class not requiring integerization according to the labeling criterion, wherein the labeling criterion is provided based on inference performance values, and wherein the input data includes image data; A conversion unit is used to integerize the first type of input data that is marked as needing to be integerized in the input data labeled by the classification unit; The first learning data unit is used to receive the first type of input data that has been integerized by the conversion unit and to infer output data. as well as The second learning data unit is used to receive second-type input data that is marked as not requiring integerization and to infer the output data.

2. The deep learning machine of claim 1, wherein the inference performance value is derived by receiving the integerized input data and inferring the output data.

3. The deep learning machine according to claim 2, wherein the inference performance value is derived by inferring the output data based on third quantized learning data, wherein the third quantized learning data is obtained by performing deep learning and quantization on the input data.

4. The deep learning machine of claim 1, wherein the labeling criterion is provided based on fourth quantized learning data obtained through deep learning and quantization rules, the rules being rules concerning the difference between the inference performance value of the input data and the inference performance value of the integerized input data.

5. The deep learning machine according to claim 1, wherein the classification unit has at least one labeling criterion corresponding to the category of the input data.

6. The deep learning machine according to claim 1, wherein the first learning data unit infers the output data based on first learning data, the first learning data being derived by deep learning on input data labeled as the first type.

7. The deep learning machine according to claim 6, wherein the first learning data unit infers the output data based on the first quantized learning data, and the first quantized learning data is obtained by quantizing the first learning data.

8. The deep learning machine according to claim 1, wherein the second learning data unit infers the output data based on the second quantized learning data, the second quantized learning data being obtained by deep learning and quantization of input data labeled as the second type.

9. A deep learning machine computation method, comprising: A classification unit with a labeling standard receives input data, and the classification unit labels the input data into a first class that needs to be integerized and a second class that does not need to be integerized according to the labeling standard, wherein the labeling standard is provided based on inference performance values, and wherein the input data includes image data; The conversion unit converts the first type of input data that is marked as needing to be integerized in the input data labeled by the classification unit into integers. The first learning data unit receives the integerized first type of input data, and the first learning data unit infers output data; as well as The second learning data unit receives second type of input data that is marked as not requiring integerization, and after the first learning data unit infers the output data, the second learning data unit infers the output data.

10. The deep learning machine operation method according to claim 9, wherein the inference performance value is derived by receiving the integerized input data and inferring the output data based on the third quantization learning data, wherein the third quantization learning data is obtained by performing deep learning and quantization on the input data.

11. The deep learning machine operation method according to claim 9, wherein the labeling standard is provided by deep learning fourth quantization learning data, the fourth quantization learning data being obtained through deep learning and quantization rules, the rules being rules concerning the difference between the inference performance value of the input data and the inference performance value of the integerized input data.

12. The deep learning machine operation method according to claim 9, wherein when the classification unit is labeled, it is labeled according to at least one labeling standard corresponding to the category of the input data.

13. The deep learning machine operation method according to claim 9, wherein when the first learning data unit infers the output data, the output data is inferred based on the first quantized learning data, wherein the first quantized learning data is obtained by deep learning and quantization of the input data labeled as the first type.

14. The deep learning machine computation method according to claim 9, wherein, When the second learning data unit infers the output data, it infers the output data based on the second quantized learning data, which is obtained by deep learning and quantization of the input data labeled as the second type.

Images