Machine learning training apparatus and method, and computer readable recording medium

Abstract

The invention provides a machine learning training device. The machine learning training device comprises a virtual difficult anchor generation circuit, a classification circuit and a training circuit. The virtual difficult anchor generation circuit is used for generating a plurality of virtual difficult anchors according to a plurality of simple samples classified into a plurality of types. The virtual difficult anchors respectively correspond to one of a plurality of types. The classification circuit is used for classifying a plurality of difficult samples into a plurality of types according to the virtual difficult anchor. The partial difficult samples classified into the multiple types in the multiple difficult samples are multiple clean difficult samples. The other part of difficult samples which are not classified into the plurality of types in the plurality of difficult samples are a plurality of noise difficult samples. The training circuit is used for performing machine learning training according to a plurality of simple samples and a plurality of clean difficult samples.

Description

Technical Field

[0001] The present invention relates to a machine learning training device, a machine learning training method and a non-volatile computer-readable recording medium, and in particular to a machine learning training device, a machine learning training method and a non-volatile computer-readable recording medium for deep neural networks (deep learning). Background Art

[0002] In recent years, in the field of machine learning, when conducting research on deep neural networks (DNN), it is necessary to rely on sample sets containing labels or annotations. However, the sample set may contain noise samples (i.e., samples with incorrect labels or annotations), which will reduce the effectiveness of the machine learning model. Therefore, various noisy label learning (NLL) methods have been proposed. Noisy label learning methods aim to find clean samples (i.e., correctly labeled samples).

[0003] A method of performing sample selection using a small loss standard is proposed. The sample selection method using the small loss standard treats samples with small classification losses as clean samples and samples with large classification losses as noise samples. However, samples with large classification losses are not necessarily noise samples. They may be difficult to be learned by deep neural networks due to their complex visual patterns, which cause large classification losses. For example, in machine learning training for classifying aircraft images and ship images, it is more difficult for aircraft samples on the water surface to be learned by deep neural networks. If only the samples with large classification losses mentioned above are regarded as noise samples and discarded, it will cause inaccurate boundary decisions of machine learning, overfitting of the model and decreased performance. Therefore, a better method for screening noise samples is one of the problems to be solved in this field. Summary of the invention

[0004] One embodiment of the present invention discloses a machine learning training device. The machine learning training device includes a virtual difficulty anchor generating circuit, a classification circuit, and a training circuit. The virtual difficulty anchor generating circuit is used to generate multiple virtual difficulty anchors based on multiple simple samples classified into multiple types. The virtual difficulty anchors correspond to one of the multiple types respectively. The classification circuit is coupled to the virtual difficulty anchor generating circuit, and is used to classify multiple difficult samples into multiple types based on the virtual difficulty anchors. Some of the difficult samples classified into multiple types among the multiple difficult samples are multiple clean difficult samples. Another part of the difficult samples that are not classified into multiple types among the multiple difficult samples are multiple noise difficult samples. The training circuit is coupled to the classification circuit, and is used to perform machine learning training based on multiple simple samples and multiple clean difficult samples.

[0005] In some embodiments, the classification circuit is also used to classify the plurality of original samples into the plurality of simple samples and the plurality of difficult samples based on the plurality of losses of the plurality of original samples and a loss threshold, wherein the machine learning training device further includes: a feature extraction circuit coupled to the classification circuit, for extracting a plurality of feature vectors of the plurality of original samples.

[0006] In some embodiments, the virtual difficulty anchor generating circuit is further used to select at least two of the plurality of simple samples, and mix at least two of the plurality of simple samples according to at least one ratio value to generate one of the plurality of virtual difficulty anchors, wherein at least two of the plurality of simple samples are different types of the plurality of types.

[0007] In some embodiments, a first virtual difficulty anchor among the plurality of virtual difficulty anchors includes a first simple sample of a first ratio value and a second simple sample of a second ratio value, wherein the first simple sample is classified as a first type among the plurality of types, and the second simple sample is classified as a second type among the plurality of types, wherein when the first ratio value is higher than the second ratio value, the virtual difficulty anchor is set to the first type.

[0008] In some embodiments, the virtual difficulty anchor generation circuit is further used to update the virtual difficulty anchor generation circuit according to a plurality of loss functions of the plurality of virtual difficulty anchors, wherein the plurality of simple samples are divided into a plurality of batches, wherein the virtual difficulty anchor generation circuit is further used to generate the plurality of virtual difficulty anchors according to the plurality of batches and update the virtual difficulty anchor generation circuit.

[0009] In some embodiments, the classification circuit is further used to obtain at least one virtual difficulty anchor among the plurality of virtual difficulty anchors, at least one distance between which and a first difficult sample among the plurality of difficult samples is less than a distance threshold, and classify the first difficult sample into one of the plurality of types according to the at least one virtual difficulty anchor.

[0010] Another embodiment of the present invention discloses a machine learning training method. The machine learning training method comprises the following steps: generating a plurality of virtual difficulty anchors based on a plurality of simple samples classified into a plurality of types, wherein the plurality of virtual difficulty anchors respectively correspond to one of the plurality of types; classifying a plurality of difficult samples into a plurality of types based on the plurality of virtual difficulty anchors, wherein a portion of the plurality of difficult samples classified into a plurality of types are a plurality of clean difficult samples, wherein another portion of the plurality of difficult samples not classified into a plurality of types are a plurality of noise difficult samples; and performing machine learning training based on the plurality of simple samples and the plurality of clean difficult samples.

[0011] In some embodiments, the machine learning training method further includes the following steps: classifying the plurality of original samples into the plurality of simple samples and the plurality of difficult samples based on the plurality of losses of the plurality of original samples and a loss threshold; and extracting the plurality of feature vectors of the plurality of original samples.

[0012] In some embodiments, a first virtual difficulty anchor among the plurality of virtual difficulty anchors includes a first simple sample of a first ratio value and a second simple sample of a second ratio value, wherein the first simple sample is classified as a first type among the plurality of types, and the second simple sample is classified as a second type among the plurality of types, wherein the training method further includes the following steps: selecting at least two of the plurality of simple samples; mixing at least two of the plurality of simple samples according to at least one ratio value to generate one of the plurality of virtual difficulty anchors, wherein at least two of the plurality of simple samples are respectively different types among the plurality of types; and when the first ratio value is higher than the second ratio value, setting the virtual difficulty anchor to the first type.

[0013] In some embodiments, the plurality of simple samples are divided into a plurality of batches, wherein the machine learning training method further comprises the following steps: updating a virtual difficulty anchor generation circuit according to a plurality of loss functions of the plurality of virtual difficulty anchors; and generating the plurality of virtual difficulty anchors according to the plurality of batches and updating the virtual difficulty anchor generation circuit.

[0014] In some embodiments, the machine learning training method further includes the following steps: obtaining at least one virtual difficulty anchor among the plurality of virtual difficulty anchors, at least one distance between which is less than a distance threshold from a first difficult sample among the plurality of difficult samples; and classifying the first difficult sample into one of the plurality of types based on the at least one virtual difficulty anchor.

[0015] Another embodiment of the present invention discloses a non-volatile computer-readable recording medium for storing a computer program, wherein when the computer program is executed, one or more processing components will be caused to perform multiple operations including: generating multiple virtual difficulty anchors based on multiple simple samples classified into multiple types, wherein the multiple virtual difficulty anchors correspond to one of the multiple types respectively; classifying multiple difficult samples into multiple types based on the multiple virtual difficulty anchors, wherein a portion of the multiple difficult samples classified into multiple types are multiple clean difficult samples, wherein another portion of the multiple difficult samples not classified into multiple types are multiple noise difficult samples; and performing machine learning training based on the multiple simple samples and the multiple clean difficult samples.

[0016] In some embodiments, a first virtual difficulty anchor among the plurality of virtual difficulty anchors includes a first simple sample of a first ratio value and a second simple sample of a second ratio value, wherein the first simple sample is classified as a first type among the plurality of types, and the second simple sample is classified as a second type among the plurality of types, wherein the plurality of operations further include: selecting at least two of the plurality of simple samples; mixing at least two of the plurality of simple samples according to at least one ratio value to generate one of the plurality of virtual difficulty anchors, wherein at least two of the plurality of simple samples are respectively different types among the plurality of types; and when the first ratio value is higher than the second ratio value, setting the virtual difficulty anchor to the first type.

[0017] In some embodiments, the plurality of operations further include: obtaining at least one virtual difficulty anchor among the plurality of virtual difficulty anchors, at least one distance between which and a first difficult sample among the plurality of difficult samples is less than a distance threshold; and classifying the first difficult sample into one of the plurality of types according to the at least one virtual difficulty anchor. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a schematic diagram of a machine learning training device according to some embodiments of the present invention;

[0019] Figure 2 is a flow chart of a machine learning training method according to some embodiments of the present invention;

[0020] Figure 3 According to some embodiments of the present invention, Figure 2 Schematic diagram of the machine learning training method in;

[0021] Figure 4 is a schematic diagram of sample classification according to some embodiments of the present invention;

[0022] Figure 5 is a schematic diagram of generating a virtual difficulty anchor according to some embodiments of the present invention;

[0023] Figure 6 is a schematic diagram of generating a virtual difficulty anchor according to some embodiments of the present invention; and

[0024] Figure 7 is a schematic diagram of sample classification according to some embodiments of the present invention.

[0025] Component number description

[0026] 100 Machine Learning Training Devices

[0027] 110 Processor

[0028] 120 Memory

[0029] 112 Feature Extraction Circuit

[0030] 114 Classification Circuit

[0031] 116 Virtual difficulty anchor generation circuit

[0032] 118 Training Circuit

[0033] 200 Machine Learning Training Methods

[0034] S210, S230, S250 steps

[0035] So original sample

[0036] Sh difficult samples

[0037] Se Simple Sample

[0038] Shal Virtual Difficulty Anchor

[0039] Shn Noise Difficulty Sample

[0040] Shc clean hard sample

[0041] Loss Distance Loss

[0042] C1, C2 types

[0043] L Boundary

[0044] Lb1, Lb2 boundary

[0045] S11, S12, S13, S14, S15, S16, S17, S18 samples

[0046] S21, S22, S23, S24, S25, S26, S27, S28 samples

[0047] Lhal loss function

[0048] Su eigenvector

[0049] Sv eigenvector

[0050] Sa eigenvector

[0051] H1, H2, H3, H4, H5, H6 Virtual Difficulty Anchors

[0052] CL distance threshold DETAILED DESCRIPTION

[0053] See also Figure 1 . Figure 1 1 is a schematic diagram of a machine learning training device 100 according to some embodiments of the present invention. In some embodiments, the machine learning training device 100 includes a processor 110 and a memory 120. In terms of connection, the processor 110 is further coupled to the memory 120. The processor 110 includes a feature extraction circuit 112, a classification circuit 114, a virtual difficulty anchor generation circuit 116, and a training circuit 118. In terms of connection, the feature extraction circuit 112 is coupled to the classification circuit 114, the classification circuit 114 is coupled to the virtual difficulty anchor generation circuit 116, and the classification circuit 114 is also coupled to the training circuit 118.

[0054] like Figure 1 The machine learning training device 100 is only for illustration purposes. Figure 1 The machine learning training device 100 may further include other components required for operation and application. For example, the machine learning training device 100 may also include an output interface (e.g., a display panel for displaying information), an input interface (e.g., a touch panel, a keyboard, a microphone, a scanner, or a flash memory reader), and a communication circuit (e.g., a WiFi communication model, a Bluetooth communication model, a wireless telecommunication network communication model, etc.). In some embodiments, the machine learning training device 100 may be established by a computer, a server, or a processing center.

[0055] In some embodiments, the memory 120 may be a flash memory, a HDD, an SSD (solid state drive), a DRAM (dynamic random access memory) or an SRAM (static random access memory). In some embodiments, the memory 120 may be a non-volatile computer-readable recording medium storing at least one instruction associated with the machine learning training method. The processor 110 may access and execute the at least one instruction.

[0056] In some embodiments, the processor 110 may be, but is not limited to, a single processor or a collection of multiple microprocessors, such as a CPU or a GPU. The microprocessor is electrically coupled to the memory 120 to access and execute the machine learning training method according to at least one instruction. For ease of understanding and explanation, the details of the machine learning training method will be described in the following paragraphs.

[0057] Details of the embodiments of the present invention are described below. Figure 2 The machine learning training method 200 disclosed in Figure 2 Is applicable to Figure 1 FIG. 2 is a flow chart of a machine learning training method 200 of a machine learning training device 100 in FIG. However, the embodiments of the present invention are not limited thereto.

[0058] See also Figure 2. Figure 2 2 is a schematic diagram of a machine learning training method 200 according to some embodiments of the present invention. However, the embodiments of the present invention are not limited thereto.

[0059] It should be noted that machine learning training methods can be applied to Figure 1 The machine learning training device 100 in FIG. 1 is a system having the same or similar structure as that in FIG. 1 . Figure 1 The machine learning training method is described by taking as an example, however, the present invention does not Figure 1 The application is limited.

[0060] It should be noted that in some embodiments, the machine learning training method can also be implemented as a computer program and stored in a non-transitory computer-readable recording medium, so that a computer, an electronic device, or the above-mentioned Figure 1 The processor 110 in the processor 110 reads the recording medium and executes the operation method. The non-transient computer readable recording medium can be a read-only memory, a flash memory, a floppy disk, a hard disk, an optical disk, a flash disk, a magnetic tape, a database accessible by a network, or a non-transient computer readable recording medium with the same function that can be easily thought of by those skilled in the art.

[0061] In addition, it should be understood that the operations of the operation method mentioned in this embodiment, except for those whose order is specifically stated, can be adjusted in order according to actual needs, and can even be executed simultaneously or partially simultaneously.

[0062] Furthermore, in different embodiments, these operations may be adaptively increased, replaced, and / or omitted.

[0063] See also Figure 2 The machine learning training method includes the following steps. For the sake of convenience and clarity, the following also refers to Figure 1 as well as Figure 2 , the relationship between the components in the machine learning training device 100 is used to illustrate Figure 2 Detailed steps of the machine learning training method shown.

[0064] In step S210, a plurality of virtual difficulty anchors are generated according to a plurality of simple samples classified into a plurality of types, wherein the plurality of virtual difficulty anchors respectively correspond to one of the plurality of types. Figure 1 The illustrated virtual difficulty anchor generation circuit 114 is executed.

[0065] Please also read Figure 3 . Figure 3 According to some embodiments of the present invention, Figure 2 Schematic diagram of the machine learning training method 200 in FIG.

[0066] like Figure 3 In some embodiments, before step S210, Figure 1 The feature extraction circuit 112 in extracts the feature vectors of the multiple original samples So. Next, Figure 1 The classification circuit 114 in the embodiment classifies the original samples So into simple samples Se and difficult samples Sh according to the feature vectors of the original samples So. After the classification circuit 114 classifies the original samples So into simple samples Se and difficult samples Sh, step S210 is performed.

[0067] Please also read Figure 4 . Figure 4 is a schematic diagram of sample classification according to some embodiments of the present invention. Figure 4 As shown, samples S11 to S18 located on the left side of the classification boundary L should be classified as type C1, and samples S21 to S28 located on the right side of the boundary L should be classified as type C2.

[0068] Samples S15 to S18 and samples S25 to S28 are simple samples Se. Since simple samples Se are far from the boundary L, the features of the samples are relatively obvious, and the classification circuit 114 is less likely to misclassify them. In some embodiments, since simple samples Se are less likely to be misclassified, simple samples Se are all clean samples.

[0069] On the other hand, samples S11 to S14 are difficult samples Sh located between boundary Lb1 and boundary L, and samples S21 to S24 are difficult samples Sh located between boundary Lb2 and boundary L. Since samples S11 to S14 and samples S21 to S24 are closer to boundary L, when the classification circuit 114 performs classification, the classification circuit 114 may make a classification error because the features of the samples are relatively unclear. For example, the classification circuit 114 classifies samples S21 and S24 as type C1, and classifies samples S12 and S14 as type C2.

[0070] In some embodiments, for samples classified as type C1 by the classification circuit 114 , the classification circuit 114 sets the sample to include a label of type C1. Similarly, for samples classified as type C2 by the classification circuit 114 , the classification circuit 114 sets the sample to include a label of type C2.

[0071] The above-mentioned types C1 and C2 are distinguished by the classification circuit 114 through the image recognition algorithm. How to classify samples into types C1 and C2 through the image recognition algorithm is known to those skilled in the art and will not be described in detail here. In addition, three or more types are also within the scope of the implementation of the present case.

[0072] In some embodiments, the classification circuit 114 is further used to classify the original sample So into a simple sample Se and a difficult sample Sh according to a small loss standard. Specifically, the classification circuit 114 classifies the samples whose losses are lower than the loss threshold as simple samples Se, and classifies the samples whose losses are not lower than the loss threshold as difficult samples Sh according to the losses of the samples S11 to S18 and S21 to S28.

[0073] In some embodiments, the classification circuit 114 is further used to adjust the loss threshold according to a set ratio value. The set ratio value is the ratio of the simple samples Se to the original samples So. When the set ratio value is higher, the loss threshold is higher, so that the number of simple samples Se relative to the original samples So can reach the set ratio value. Conversely, when the set ratio value is lower, the loss threshold is lower, so that the number of simple samples Se relative to the number of original samples So can not exceed the set ratio value.

[0074] Please also read Figure 5 . Figure 5 FIG. 1 is a schematic diagram of generating a virtual difficulty anchor according to some embodiments of the present invention. Figure 5 As shown, in some embodiments, the virtual difficult anchor generation circuit 116 generates the virtual difficult anchor Shal according to the simple sample Se. After the virtual difficult anchor Shal is generated, the classification circuit 114 further classifies the virtual difficult anchor Shal to classify the virtual difficult anchor Shal into type C1 or type C2, and calculates the loss function Lhal of the virtual difficult anchor Shal.

[0075] Please also read Figure 6 . Figure 6 is a schematic diagram of generating a virtual difficulty anchor according to some embodiments of the present invention. In one embodiment, the virtual difficulty anchor generating circuit 116 is further configured to select at least two of the simple samples and mix the at least two of the simple samples according to at least one ratio value to generate one of the virtual difficulty anchors, wherein the at least two of the simple samples are of different types. Figure 6 As shown, the feature vector Su is the feature vector of the sample S25 of type C2, and the feature vector Sv is the feature vector of the sample S15 of type C1. The virtual difficulty anchor generation circuit 116 generates a feature vector Sa according to the feature vector Su and the feature vector Sv. The feature vector Sa corresponds to the virtual difficulty anchor H1.

[0076] In one embodiment, the virtual difficulty anchor H1 includes a sample S25 of a first ratio value and a sample S15 of a second ratio value. When the first ratio value is higher than the second ratio value, the virtual difficulty anchor generation circuit 116 sets the virtual difficulty anchor H1 to the second type C2, i.e., the same type as the sample S25. In other words, the virtual difficulty anchor generation circuit 116 sets the virtual difficulty anchor H1 to include a label of type C2.

[0077] In some other embodiments, the virtual difficulty anchor generation circuit 116 may further fuse three or more samples selected from different types to generate a virtual difficulty anchor, and use the type of the sample with the highest ratio value as the type of the virtual difficulty anchor.

[0078] In some embodiments, the virtual difficulty anchor generation circuit 116 calculates the loss function Lhal of the virtual difficulty anchor Shal according to the following formula.

[0079] Lha1=Lclo+H(Sa,Yu).

[0080] Lclo=-λ p <Sa,Su> -(1-λ p )<Sa,Sv> .

[0081] λ p ∈[0.5,1.0].

[0082] In the above formula,<Sa,Sv> is the cosine similarity (CosineDistance) between the feature vector Sa and the feature vector Sv,<Sa,Su> is the cosine similarity between the feature vector Sa and the feature vector Su, and H(Sa,Yu) is the cross entropy loss calculation function between the feature vector Sa and the label Yu (type C2) of the feature vector Sa. p is a random parameter between 0.5 and 1.0.

[0083] In some embodiments, the virtual difficulty anchor generation circuit 116 is further configured to update the virtual difficulty anchor generation circuit 116 according to the loss function Lhal of the virtual difficulty anchor Shal.

[0084] In some embodiments, the simple samples Se are divided into a plurality of batches. The virtual difficult anchor generation circuit 116 is further used to generate a virtual difficult anchor Shal according to the batches and update the virtual difficult anchor generation circuit 116.

[0085] For example, assume that the simple samples Se are divided into three batches, each batch contains part of the simple samples Se. The virtual difficult anchor generation circuit 116 first generates the first batch of virtual difficult anchors Shal based on the simple samples Se in the first batch, calculates the loss function Lhal, and then updates the virtual difficult anchor generation circuit 116 based on the loss function Lhal. Next, the virtual difficult anchor generation circuit 116 generates the second batch of virtual difficult anchors Shal based on the simple samples Se in the second batch, calculates the loss function Lhal, and then updates the virtual difficult anchor generation circuit 116 based on the loss function Lhal. Finally, the virtual difficult anchor generation circuit 116 generates the third batch of virtual difficult anchors Shal based on the simple samples Se in the third batch, calculates the loss function Lhal, and then updates the virtual difficult anchor generation circuit 116 based on the loss function Lhal.

[0086] Please refer back to Figure 2 In step S230, the plurality of difficult samples are classified into a plurality of types according to the plurality of virtual difficult anchors, wherein some of the plurality of difficult samples classified into a plurality of types are a plurality of clean difficult samples, and another portion of the plurality of difficult samples not classified into a plurality of types are a plurality of noise difficult samples. In some embodiments, step S230 is performed as follows Figure 1 The classification circuit 114 in is performed.

[0087] In some embodiments, the classification circuit 114 is further configured to obtain at least one virtual difficulty anchor among the plurality of virtual difficulty anchors, at least one distance between the virtual difficulty anchor and a first difficult sample among the plurality of difficult samples being less than a distance threshold, and classify the first difficult sample into one of the plurality of types according to the at least one virtual difficulty anchor.

[0088] Please also read Figure 7 . Figure 7 is a schematic diagram of sample classification according to some embodiments of the present invention. Figure 7 As shown, for sample S21, the virtual difficulty anchors whose distances from sample S21 are less than the distance threshold CL include virtual difficulty anchors H1, H5, and H6. Virtual difficulty anchors H1 and H5 are of type C2, and virtual difficulty anchor H6 is of type C1. Since more virtual difficulty anchors are of type C2 among virtual difficulty anchors H1, H5, and H6, the classification circuit 114 classifies sample S21 as type C2.

[0089] In some embodiments, if there is no virtual difficult anchor whose distance to the difficult sample Sh is less than the distance threshold CL, the difficult sample Sh will not be classified into type C1 or type C2. Alternatively, if the number of virtual difficult anchors whose distance to the difficult sample Sh is less than the distance threshold CL is less than the number threshold (e.g., less than 3), the difficult sample Sh will not be classified into type C1 or type C2. The classification circuit 114 determines the difficult sample Sh that will not be classified into type C1 or type C2 as a noise difficult sample Shn.

[0090] On the other hand, for a difficult sample Sh classified into type C1 or type C2, the classification circuit 114 determines the difficult sample as a clean difficult sample Shc.

[0091] In some embodiments, the distance mentioned above refers to the distance vector between feature vectors.

[0092] For example, Figure 7 Among the difficult samples S11 to S14 and S21 to S24, samples S11, S22, S14, S24 are noise difficult samples Shn, while samples S12, S13, S22, S23 are clean difficult samples Shc.

[0093] In step S250, machine learning training is performed based on a plurality of simple samples and a plurality of clean and difficult samples. In some embodiments, step S250 is performed as follows Figure 1 The training circuit 118 in is performed.

[0094] Please refer back to Figure 3 In some embodiments, the simple sample Sc and the clean and difficult sample Shc (and the correct labels contained therein) are input into the training circuit 118 for machine learning training. The training circuit 118 also calculates the distance loss Loss based on the simple sample Sc and the clean and difficult sample Shc (and the correct labels contained therein).

[0095] In some embodiments, the training circuit 118 is a semi-supervised learning model, and the training circuit further inputs the output of the semi-supervised learning model into a distance loss calculation model (not shown) to calculate the distance loss Loss. The distance loss Loss is used to determine the training efficiency of the training circuit 118.

[0096] In some embodiments, the feature extraction circuit 112 and the classification circuit 114 are further configured to update the feature extraction circuit 112 and the classification circuit 114 according to the simple sample Sc and the clean hard sample Shc (and the correct labels included therein).

[0097] In summary, the embodiments of the present invention provide a machine learning training device, a machine learning training method, and a non-volatile computer-readable recording medium. In addition to classifying the original samples into clean samples and noise samples based on a small loss standard, the implementation method of the present case classifies the original samples into simple samples and difficult samples. Based on the identified simple samples, a virtual difficult anchor is generated, and valuable clean difficult samples are identified from the difficult samples based on the virtual difficult anchor to better utilize the sample data for machine learning training. By using simple samples and clean difficult samples containing correct labels to perform machine learning (deep neural network) training, the implementation method of the present case achieves significant performance improvements compared to the prior art.

[0098] Although specific embodiments of the present invention have been disclosed in relation to the above embodiments, these embodiments are not intended to limit the present invention. Various substitutions and improvements may be performed in the present invention by a person of ordinary skill in the relevant art without departing from the principles and spirit of the present invention. Therefore, the protection scope of the present invention is determined by the scope of the attached patent application.

Claims

1. A machine learning training device, characterized in that: Include: a virtual difficulty anchor generation circuit for generating a plurality of virtual difficulty anchors according to a plurality of simple samples classified into a plurality of types, wherein the plurality of virtual difficulty anchors respectively correspond to one of the plurality of types; a classification circuit, coupled to the virtual difficulty anchor generation circuit, for classifying a plurality of difficult samples into the plurality of types according to the plurality of virtual difficulty anchors, wherein a portion of the plurality of difficult samples classified into the plurality of types is a plurality of clean difficult samples, and another portion of the plurality of difficult samples not classified into the plurality of types is a plurality of noise difficult samples; as well as A training circuit is coupled to the classification circuit and is used for performing machine learning training according to the plurality of simple samples and the plurality of clean and difficult samples.

2. The machine learning training device according to claim 1, characterized in that: The classification circuit is further used to classify the plurality of original samples into the plurality of simple samples and the plurality of difficult samples according to the plurality of losses of the plurality of original samples and a loss threshold, wherein the machine learning training device further comprises: A feature extraction circuit is coupled to the classification circuit and is used to extract a plurality of feature vectors of the plurality of original samples.

3. The machine learning training device according to claim 1, characterized in that: The virtual difficulty anchor generating circuit is further used to select at least two of the plurality of simple samples and mix at least two of the plurality of simple samples according to at least one ratio value to generate one of the plurality of virtual difficulty anchors, wherein at least two of the plurality of simple samples are respectively different types of the plurality of types.

4. The machine learning training device according to claim 3, characterized in that: A first virtual difficulty anchor among the plurality of virtual difficulty anchors comprises a first simple sample of a first ratio value and a second simple sample of a second ratio value, wherein the first simple sample is classified as a first type among the plurality of types, and the second simple sample is classified as a second type among the plurality of types, wherein when the first ratio value is higher than the second ratio value, the virtual difficulty anchor is set to the first type.

5. The machine learning training device according to claim 1, characterized in that: The virtual difficulty anchor generation circuit is also used to update the virtual difficulty anchor generation circuit according to the multiple loss functions of the multiple virtual difficulty anchors, wherein the multiple simple samples are divided into multiple batches, and the virtual difficulty anchor generation circuit is also used to generate the multiple virtual difficulty anchors according to the multiple batches and update the virtual difficulty anchor generation circuit.

6. The machine learning training device according to claim 1, characterized in that: The classification circuit is further configured to obtain at least one virtual difficulty anchor among the plurality of virtual difficulty anchors, at least one distance between which and a first difficult sample among the plurality of difficult samples is less than a distance threshold, and classify the first difficult sample into one of the plurality of types according to the at least one virtual difficulty anchor.

7. A machine learning training method, characterized in that: Include: generating a plurality of virtual difficulty anchors according to a plurality of simple samples classified into a plurality of types, wherein the plurality of virtual difficulty anchors respectively correspond to one of the plurality of types; Classifying a plurality of difficult samples into the plurality of types according to the plurality of virtual difficulty anchors, wherein a portion of the plurality of difficult samples classified into the plurality of types is a plurality of clean difficult samples, and another portion of the plurality of difficult samples not classified into the plurality of types is a plurality of noise difficult samples; and Machine learning training is performed according to the plurality of simple samples and the plurality of clean and difficult samples.

8. The machine learning training method according to claim 7, characterized in that: Also includes: According to a plurality of losses of a plurality of original samples and a loss threshold, classifying the plurality of original samples into the plurality of simple samples and the plurality of difficult samples; and Extract a plurality of feature vectors of the plurality of original samples.

9. The machine learning training method according to claim 7, characterized in that: A first virtual difficulty anchor among the plurality of virtual difficulty anchors comprises a first simple sample of a first ratio value and a second simple sample of a second ratio value, wherein the first simple sample is classified as a first type among the plurality of types, and the second simple sample is classified as a second type among the plurality of types, wherein the training method further comprises: Selecting at least two of the plurality of simple samples; mixing at least two of the plurality of simple samples according to at least one ratio value to generate one of the plurality of virtual difficulty anchors, wherein at least two of the plurality of simple samples are respectively of different types among the plurality of types; and When the first ratio value is higher than the second ratio value, the virtual difficulty anchor is set to the first type.

10. The machine learning training method according to claim 7, characterized in that: The plurality of simple samples are divided into a plurality of batches, wherein the machine learning training method further comprises: updating a virtual difficulty anchor generating circuit according to a plurality of loss functions of the plurality of virtual difficulty anchors; and The plurality of virtual difficulty anchors are generated according to the plurality of batches and the virtual difficulty anchor generating circuit is updated.

11. The machine learning training method according to claim 7, characterized in that: Also includes: Obtaining at least one virtual difficult anchor among the plurality of virtual difficult anchors, at least one distance between which and a first difficult sample among the plurality of difficult samples is less than a distance threshold; as well as The first difficult sample is classified into one of the plurality of types according to the at least one virtual difficulty anchor.

12. A non-volatile computer-readable recording medium for storing a computer program, characterized in that: When the computer program is executed, one or more processing components are caused to perform a plurality of operations including: generating a plurality of virtual difficulty anchors according to a plurality of simple samples classified into a plurality of types, wherein the plurality of virtual difficulty anchors respectively correspond to one of the plurality of types; Classifying a plurality of difficult samples into the plurality of types according to the plurality of virtual difficulty anchors, wherein a portion of the plurality of difficult samples classified into the plurality of types is a plurality of clean difficult samples, and another portion of the plurality of difficult samples not classified into the plurality of types is a plurality of noise difficult samples; and Machine learning training is performed according to the plurality of simple samples and the plurality of clean and difficult samples.

13. The non-volatile computer readable recording medium according to claim 12, wherein: A first virtual difficulty anchor among the plurality of virtual difficulty anchors comprises a first simple sample of a first ratio value and a second simple sample of a second ratio value, wherein the first simple sample is classified as a first type among the plurality of types, and the second simple sample is classified as a second type among the plurality of types, wherein the plurality of operations further comprises: Selecting at least two of the plurality of simple samples; mixing at least two of the plurality of simple samples according to at least one ratio value to generate one of the plurality of virtual difficulty anchors, wherein at least two of the plurality of simple samples are respectively of different types among the plurality of types; and When the first ratio value is higher than the second ratio value, the virtual difficulty anchor is set to the first type.

14. The non-volatile computer readable recording medium according to claim 12, wherein: The complex number operation also includes: Obtaining at least one virtual difficult anchor among the plurality of virtual difficult anchors, at least one distance between which and a first difficult sample among the plurality of difficult samples is less than a distance threshold; as well as The first difficult sample is classified into one of the plurality of types according to the at least one virtual difficulty anchor.

Images