Patent Application
20230186047
The present application is based on, and claims priority from JP Application Serial Number 2021-200534, filed Dec. 10, 2021, the disclosure of which is hereby incorporated by reference herein in its entirety.
The present disclosure relates to a technique for evaluating a trained machine learning model.
U.S. Pat. No. 5,210,798 and WO 2019/083553 disclose a so-called capsule network as a vector neural network type machine learning model using a vector neuron. The vector neuron means a neuron whose input and output are vectors. The capsule network is a machine learning model in which a vector neuron called a capsule is a node of the network. The vector neural network type machine learning model such as the capsule network can be used for class discrimination of input data.
In the related art, there may occur an event in which a trained machine learning model does not match an initial target, such as not reaching a desired discrimination accuracy. Evaluation of the trained machine learning model, such as identification of the cause of this event, is likely to differ between users depending on user experience, etc. Therefore, there has been a demand for a technique capable of evaluating the trained machine learning model without causing a difference between users. The evaluation of the trained machine learning model includes an evaluation of evaluation data used for an evaluation of the trained machine learning model in addition to the evaluation of the trained machine learning model itself.
According to a first aspect of the present disclosure, there is provided an evaluation method for a trained machine learning model. The machine learning model is a vector neural network model including a plurality of vector neuron layers, and is trained by using a plurality of pieces of training data including input data and a prior label associated with the input data. The evaluation method includes the steps of (a) inputting evaluation data to the trained machine learning model to generate first explanatory information used for an evaluation of the machine learning model, (b) using a value indicated by each piece of information included in the first explanatory information to generate second explanatory information indicating an evaluation of the trained machine learning model, and (c) outputting the generated second explanatory information, wherein the step (a) includes the steps of (a1) inputting the evaluation data to the trained machine learning model to obtain a feature spectrum from an output of a specific layer of the trained machine learning model, (a2) obtaining a spectral similarity that is a similarity between the feature spectrum and a known feature spectrum included in a known feature spectrum group obtained from an output of the specific layer by inputting the plurality of pieces of training data to the trained machine learning model again, for each of a plurality of the known feature spectra included in the known feature spectrum group, (a3) obtaining a data similarity that is a similarity between the input data and the evaluation data, and (a4) generating the first explanatory information including spectral similarity information related to the spectral similarity and data similarity information related to the data similarity.
According to a second aspect of the present disclosure, there is provided an evaluation device for a trained machine learning model. The evaluation device includes a memory configured to store the machine learning model, the machine learning model being a vector neural network model including a plurality of vector neuron layers, the machine learning model being trained by using a plurality of pieces of training data including input data and a prior label associated with the input data, and a processor, wherein the processor is configured to execute processing of (a) inputting evaluation data to the trained machine learning model to generate first explanatory information used for an evaluation of the machine learning model, (b) using a value indicated by each piece of information included in the first explanatory information to generate second explanatory information indicating an evaluation of the trained machine learning model, and (c) outputting the generated second explanatory information, wherein the processing (a) includes processing of (a1) inputting the evaluation data to the trained machine learning model to obtain a feature spectrum from an output of a specific layer of the trained machine learning model, (a2) obtaining a spectral similarity that is a similarity between the feature spectrum and a known feature spectrum included in a known feature spectrum group obtained from an output of the specific layer by inputting the plurality of pieces of training data to the trained machine learning model again, for each of a plurality of the known feature spectra included in the known feature spectrum group, (a3) obtaining a data similarity that is a similarity between the input data and the evaluation data, and (a4) generating the first explanatory information including spectral similarity information related to the spectral similarity and data similarity information related to the data similarity.
According to a third aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing a program for causing a computer to execute evaluation of a trained machine learning model. The machine learning model is a vector neural network model including a plurality of vector neuron layers, and is trained by using a plurality of pieces of training data including input data and a prior label associated with the input data. The computer program causes the computer to execute the following functions (a) inputting evaluation data to the trained machine learning model to generate first explanatory information used for an evaluation of the machine learning model, (b) using a value indicated by each piece of information included in the first explanatory information to generate second explanatory information indicating an evaluation of the trained machine learning model, and (c) outputting the generated second explanatory information, wherein the function (a) includes the following functions (a1) inputting the evaluation data to the trained machine learning model to obtain a feature spectrum from an output of a specific layer of the trained machine learning model, (a2) obtaining a spectral similarity that is a similarity between the feature spectrum and a known feature spectrum included in a known feature spectrum group obtained from an output of the specific layer by inputting the plurality of pieces of training data to the trained machine learning model again, for each of a plurality of the known feature spectra included in the known feature spectrum group, (a3) obtaining a data similarity that is a similarity between the input data and the evaluation data, and (a4) generating the first explanatory information including spectral similarity information related to the spectral similarity and data similarity information related to the data similarity.
A. Exemplary Embodiment:
The evaluation device 100 includes a processor 110, a memory 120, an interface circuit 130, and an input device 140 and a display unit 150 coupled to the interface circuit 130. The evaluation device 100 is, for example, a personal computer. The imaging device 400 is also coupled to the interface circuit 130. Although not limited thereto, for example, the processor 110 not only has a function of executing processes described in detail below but also has a function of displaying data obtained by the process and data generated in the course of the process on the display unit 150.
The processor 110 includes a training execution unit 112, a class discrimination unit 113, and an evaluation processing unit 114 by executing various programs stored in the memory 120. The training execution unit 112 executes the training process of the machine learning model 200 using a training data group TDG. The class discrimination unit 113 inputs data IM to the trained machine learning model 200 to execute class discrimination of the input data IM. The evaluation processing unit 114 inputs evaluation data ED to the trained machine learning model 200 to generate first explanatory information FEI, and generates second explanatory information SEI indicating an evaluation of the machine learning model 200 from the first explanatory information FEI. The generated secondary explanatory information SEI is output to the display unit 150.
The evaluation processing unit 114 includes a similarity calculation unit 310 and an evaluation unit 330. The similarity calculation unit 310 inputs the evaluation data ED to the trained machine learning model 200 to generate the first explanatory information FEI including spectral similarity information IRSp and data similarity information IDa. The spectral similarity information IRSp is information indicating a degree of similarity between a known feature spectrum KSp obtained by inputting the training data TD to the trained machine learning model 200 and a feature spectrum Sp obtained by inputting the evaluation data ED to the trained machine learning model 200. The data similarity information IDa is information indicating a data similarity Da between the training data TD of a generation source of a specific known feature spectrum KSp specified based on a spectral similarity RSp between the known feature spectrum KSp and the feature spectrum Sp, and the evaluation data ED. The first explanatory information FEI is used to evaluate the trained machine learning model 200. The evaluation unit 330 generates the second explanatory information SEI indicating an evaluation of the trained machine learning model 200 by using values indicated by various kinds of information included in the first explanatory information FEI.
In the above description, at least a part of the functions of the training execution unit 112, the class discrimination unit 113, and the evaluation processing unit 114 may be realized by a hardware circuit. The processor in this specification is a term including such a hardware circuit. The processor that executes the class discrimination process may be one or more processors included in one or more remote computers coupled to the evaluation device 100 via a network.
The memory 120 stores the machine learning model 200, the training data group TDG, an evaluation data group EDG, and a known feature spectrum group KSpG. The machine learning model 200 is used for class discrimination of the input data IM. Each machine learning model 200 is a vector neural network type machine learning model having a plurality of vector neuron layers. The machine learning model 200 is trained using a plurality of pieces of the training data TD. A configuration example and an operation of the machine learning model 200 will be described later.
The training data group TDG is a set of the training data TD that is supervised data. In the present exemplary embodiment, each training data TD of the training data group TDG includes input data IM and a prior label LB associated with the input data IM. In the present exemplary embodiment, the input data IM is a captured image of a target object captured by the imaging device 400. In the present exemplary embodiment, the prior label LB is a label indicating a type of the target object. In the present exemplary embodiment, the “label” and “class” have the same meaning.
The evaluation data group EDG is a set of the evaluation data ED used for evaluating the trained machine learning model 200. The evaluation data ED is at least one type of data of: the training data TD; verification data VD; and abnormality data AD. The training data TD is data of the training data group TDG used for training of the machine learning model 200. The verification data VD is data that is not used for training the machine learning model 200 and includes the input data IM and the prior label LB associated with the input data IM. In the present exemplary embodiment, the verification data VD is data generated by cross-validation on the training data group TDG. That is, a part of the plurality of piece of training data TD prepared for training is used as the verification data VD. The abnormal data AD is the input data IM with which the prior label LB is not associated. The abnormal data AD is the input data IM assumed to be classified as an unknown-class different from the class corresponding to the prior label LB by the machine learning model 200.
The known feature spectrum group KSpG is a set of the feature spectra Sp obtained when the training data group TDG is input to the trained machine learning model 200. The feature spectrum Sp will be described later. Note that feature spectrum Sp of the known feature spectrum group KSpG is also referred to as the known feature spectrum KSp.
Although two convolution vector neuron layers 240, 250 are used in the example of
A captured image having a size of 28 x 28 pixels is input to the input layer 210. The configuration of each layer other than the input layer 210 can be described as follows.
Conv layer 220: Conv [32, 5, 2]
PrimeVN layer 230: PrimeVN [16, 1, 1]
ConvVN1 layer 240: ConvVN1 [12, 3, 2]
ConvVN2 layer 250: ConvVN2 [6, 3, 1]
ClassVN layer 260: ClassVN [M, 3, 1]
Vector dimension VD: VD=16
In the description of each layer, a character string before parentheses is a layer name, and numbers in the parentheses are the number of channels, the surface size of the kernel, and the stride in this order. For example, the layer name of the Conv layer 220 is “Conv”, the number of channels is 32, the surface size of the kernel is 5×5, and the stride is 2. In
The input layer 210 and the Conv layer 220 are layers composed of scalar neurons. The other layers 230 to 260 are layers composed of vector neurons. The vector neuron is a neuron whose input and output are vectors. In the above description, the dimension of the output vector of the individual vector neuron is constant at 16. Hereinafter, the term “node” is used as a superordinate concept of the scalar neuron and the vector neuron.
In
As is well known, a post-convolution resolution W1 is given by:
W1=Ceil{(W0−Wk+1)/S} (A1)
Where, W0 is a pre-convolution resolution, Wk is a surface size of the kernel, S is a stride, and Ceil {X} is a function for rounding up the fractional part of X.
The resolution of each layer illustrated in
The ClassVN layer 260 has M channels. M is the number of classes discriminated by the machine learning model 200. In the present exemplary embodiment, M is 2, and two class determination values Class_1 and Class_2 are output. The number of channels M of the ClassVN layer 260 can be set to an arbitrary integer of 2 or more. When both of the two class determination values Class_1 and Class_2 are less than a threshold, an unknown-class indicating that the data belongs to a class different from that of the training data TD used for training is output as the class discrimination result.
In
As illustrated in
In the present disclosure, the vector neuron layer used for calculating the spectral similarity RSp is also referred to as a “specific layer”. As the specific layer, a vector neuron layer other than the ConvVN2 layer 250 may be used, and an arbitrary number of one or more vector neuron layers can be used. Further, a configuration of the feature spectrum Sp, a generation method of the spectral similarity information IRSp, and a generation method of the data similarity information IDa will be described later.
As illustrated in
In step S14, the training execution unit 112 generates the known feature spectrum group KSpG by inputting the training data group TDG used for training again to the trained machine learning model 200. The generated known feature spectrum group KSpG is stored in the memory 120. The known feature spectrum group KSpG is a set of the known feature spectra KSp to be described below.
The vertical axis of
The number of feature spectra Sp obtained from the output of the ConvVN2 layer 250 for one piece of the input data IM is equal to the number of planar positions (x, y) of the ConvVN2 layer 250, i.e., the number of partial regions R250, and is therefore 9.
Each record of the known feature spectrum group KSpG includes a parameter k indicating an order of the partial region Rn in the layer, a parameter c indicating the class, a parameter q indicating the data number, and the known feature spectrum KSp. The known feature spectrum KSp is the same as the feature spectrum Sp in
The parameter k of the partial region Rn takes a value indicating one of the plurality of partial regions Rn included in the specific layer, that is, one of the planar positions (x, y). In the ConvVN2 layer 250, the number of the partial regions R250 is 9, therefore k=1 to 9. The parameter c representing the class takes a value indicating one of the M classes that can be discriminated by the machine learning model 200. In the present exemplary embodiment, M=2, therefore c=1 to 2. The parameter q of the data number indicates the serial number of the training data belonging to each class, takes a value from 1 to max1 for c=1, and takes a value from 1 to max2 for c=2. As described above, the feature spectrum Sp is associated with the class c and the data number q of the training data. The feature spectra Sp are classified into classes.
A-2. Summary description of an evaluation process:
As illustrated in
The spectral similarity RSp includes a self-class maximum spectral similarity RSp_maxA and a different-class maximum spectral similarity RSp_maxB. The self-class maximum spectral similarity RSp_maxA is a similarity indicating a maximum value among a plurality of self-class spectral similarities RSp_A. The self-class spectral similarity RSp_A is the spectral similarity RSp between the feature spectrum Sp of the evaluation data ED and a self-class known feature spectrum KSp_A which is the known feature spectrum of the same class as the evaluation class indicated by the prior label LB associated with the evaluation data ED in the known feature spectrum group KSpG. The different-class maximum spectral similarity RSp_maxB is a similarity indicating a maximum value among a plurality of different-class spectral similarities RSp_B. The different-class spectral similarity RSp_B is the spectral similarity RSp between the feature spectrum Sp of the evaluation data ED and a different-class known feature spectrum KSp_B which is the known feature spectrum of a class different from the evaluation class in the known feature spectrum group KSpG.
The distribution representative value PRSp includes a self-class spectrum representative value PRSp_A and a different-class spectrum representative value PRSp_B. The self-class spectrum representative value PRSp_A is a representative value of the distribution of the plurality of self-class spectral similarities RSp_A. The different-class spectrum representative value PRSp_B is a representative value of the distribution of the plurality of different-class spectral similarities RSp_B. A detailed method of generating the representative value PRSp will be described later.
The spectral similarity information IRSp includes self-class spectral similarity information IRSp_A related to the self-class maximum spectral similarity RSp_maxA and different-class spectral similarity information IRSp_B related to the different-class maximum spectral similarity RSp_maxB. The self-class spectral similarity information IRSp_A includes the self-class maximum spectral similarity RSp_maxA and the self-class spectral representative value PRSp_A described above. In other exemplary embodiments, the self-class spectral similarity information IRSp_A may include at least one of the self-class maximum spectral similarity RSp_maxA and the self-class spectral representative value PRSp_A. The different-class spectral similarity information IRSp_B includes the different-class maximum spectral similarity RSp_maxB and the different-class spectral representative value PRSp_B described above. In other exemplary embodiments, the different-class spectral similarity information IRSp_B may include at least one of the different-class maximum spectral similarity RSp_maxB and the different-class spectral representative value PRSp_B.
The data similarity information IDa is information related to the data similarity Da. In the present exemplary embodiment, the data similarity information IDa is information indicating the data similarity Da. The data similarity Da includes a self-class maximum data similarity Da_A and a different-class maximum data similarity Da_B. The self-class maximum data similarity Da_A is a similarity between the input data IM associated with the self-class known feature spectrum KSp_A that is a calculation source of the self-class maximum spectral similarity RSp_maxA and the evaluation data ED. The different-class maximum data similarity Da_B is a similarity between the input data IM associated with the different-class known feature spectrum KSp_B that is a calculation source of the different-class maximum spectral similarity RSp_maxB and the evaluation data ED.
The data similarity information IDa includes self-class data similarity information IDa_A and different-class data similarity information IDa_B. The self-class data similarity information IDa_A is information related to the above-described self-class maximum data similarity Da_A, and is information indicating the self-class maximum data similarity Da_A in the present exemplary embodiment. The different-class data similarity information IDa_B is information related to the above-described different-class maximum data similarity Da_B, and is information indicating the different-class maximum data similarity Da_B in the present exemplary embodiment.
As illustrated in
A-3. Calculation method for the spectral similarity:
As a calculation method for the above-described spectral similarity, for example, any one of the following methods can be adopted.
(1) a first calculation method M1 for obtaining the spectral similarity RSp without considering a correspondence of the partial regions in the known feature spectrum KSp of the training data TD and in the feature spectrum Sp of the evaluation data ED.
(2) a second calculation method M2 for obtaining the spectral similarity RSp between the corresponding partial regions Rn with respect to the known feature spectrum KSp of the training data TD and the feature spectrum Sp of the evaluation data ED.
(3) a third calculation method M3 for obtaining the spectral similarity RSp without considering the partial region Rn at all.
In the following, a method of calculating the spectral similarity RSp from the output of the ConvVN2 layer 250 according to these calculation methods M1, M2, and M3 will be sequentially described.
In the first calculation method M1, the local spectral similarity S (i, j, k) is calculated using the following equation.
S (i, j, k)=max [G{Sp (j, k), KSp (i, j, k=all, q=all)}] (c1)
Where
i is a parameter indicating the self-class or the different-class;
j is a parameter indicating the specific layer;
k is a parameter indicating the partial region Rn;
q is a parameter indicating the data number;
G {a, b} is a function for obtaining the spectral similarity between a and b;
Sp (j, k) is the feature spectrum obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED;
KSp (i, j, k=all, q=all) is the known feature spectrum of the all data number q in all partial regions Rn of the specific layer j associated with a self-class i or different-class i in the known feature spectrum group KSpG illustrated in
max [X] is a logical operation taking the maximum value of the values of X.
As the function G {a, b} for obtaining the local spectral similarity, for example, an equation for obtaining the cosine similarity or an equation for obtaining the similarity according to the distance can be used. Further, when the evaluation data ED is the abnormal data AD, the above parameter i specifies all classes, so KSp (i, j, k=all, q=all) in the above equation (c1) becomes KSp (i=all, j, k=all, q=all).
The three types of maximum spectral similarities RSp_max illustrated on the right side of
As described above, in the first calculation method Ml of the maximum spectral similarity, the maximum spectral similarities RSp_maxA, RSp_maxB, and RSp_max are calculated by the following method.
(1) obtaining the local spectral similarity S (i, j, k) that is a spectral similarity between the feature spectrum Sp obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED and all the known feature spectra KSp associated with the specific layer j and the same class as the prior label LB of the evaluation data ED, and
(2) obtaining the self-class maximum spectral similarity RSP_maxA by taking the maximum value, the average value, or the minimum value of the local spectral similarity S (i, j, k) for the plurality of partial regions Rn.
(1) obtaining the local spectral similarity S (i, j, k) that is a spectral similarity between the feature spectrum Sp obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED and all the known feature spectra KSp associated with the specific layer j and a class different from the prior label LB of the evaluation data ED, and
(2) obtaining the different-class maximum spectral similarity RSP_maxB by taking the maximum value, the average value, or the minimum value of the local spectral similarity S (i, j, k) for the plurality of partial regions Rn. Maximum spectral similarity RSp_max
(1) obtaining the local spectral similarity S (i, j, k) that is a spectral similarity between the feature spectrum Sp obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED and all the known feature spectra KSp associated with the specific layer j, and
(2) obtaining the maximum spectral similarity RSP_max by taking the maximum value, the average value, or the minimum value of the local spectral similarity S (i, j, k) for the plurality of partial regions Rn.
According to the first calculation method M1, the maximum spectral similarities RSp_max, RSp_maxA, and RSp_maxB can be obtained by a relatively simple calculation and procedure.
Where,
KSp (i, j, k, q=all) is the known feature spectrum of the all data number q in the specific partial region Rn of the specific layer j associated with the self-class i or different-class i in the known feature spectrum group KSpG illustrated in
In the first calculation method M1 described above, the known feature spectra KSp (i, j, k=all, q=all) in all the partial regions Rn of the specific layer j are used, whereas in the second calculation method M2, only the known feature spectra KSp (i, j, k, q=all) for the same partial region Rn as the partial region Rn of the feature spectra Sp (j, k) are used. Other methods in the second calculation method M2 are the same as those in the first calculation method M1.
As described above, in the second calculation method M2 of the maximum spectral similarity, the maximum spectral similarities RSp_maxA, RSp_maxB, and RSp_max are calculated by the following method.
(1) obtaining the local spectral similarity S (i, j, k) that is a spectral similarity between the feature spectrum Sp obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED and all the known feature spectra KSp associated with the specific partial region Rn of the specific layer j and the same class as the prior label LB of the evaluation data ED, and
(2) obtaining the self-class maximum spectral similarity RSP_maxA by taking the maximum value, the average value, or the minimum value of the local spectral similarity S (i, j, k) for the plurality of partial regions Rn. Different-class maximum spectral similarity RSp_maxB
(1) obtaining the local spectral similarity S (i, j, k) that is a spectral similarity between the feature spectrum Sp obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED and all the known feature spectra KSp associated with the specific partial region Rn of the specific layer j and a class different from the prior label LB of the evaluation data ED, and
(2) obtaining the different-class maximum spectral similarity RSP_maxB by taking the maximum value, the average value, or the minimum value of the local spectral similarity S (i, j, k) for the plurality of partial regions Rn. Maximum spectral similarity RSp_max
(1) obtaining the local spectral similarity S (i, j, k) that is a spectral similarity between the feature spectrum Sp obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED and all the known feature spectra KSp associated with the specific layer j and the specific partial region Rn, and
(2) obtaining the maximum spectral similarity RSP_max by taking the maximum value, the average value, or the minimum value of the local spectral similarity S (i, j, k) for the plurality of partial regions Rn.
According to the second calculation method M2, the maximum spectral similarities RSp_max, RSp_maxA, and RSp_maxB can be obtained by a relatively simple calculation and procedure.
The maximum spectral similarity RSp (i, j) obtained by the third calculation method M3 is calculated using the following equation.
RSP (i, j)=max [G {Sp (j, k=all), KSp (i, j, k=all, q=all)}] (c3)
Where,
Sp (j, k=all) is a feature spectrum obtained from the output of all the partial regions Rn of the specific layer j in accordance with the evaluation data ED. Further, when the evaluation data ED is the abnormal data AD, the above parameter i specifies all classes, so KSp (i, j, k=all, q=all) in the above equation (c3) becomes KSp (i=all, j, k=all, q=all).
As described above, in the third calculation method M3 of the maximum spectral similarity, the maximum spectral similarities RSp_maxA, RSp_maxB, and RSp_max are calculated by the following method.
(1) obtaining the self-class spectral similarity RSp_A that is a similarity between all the feature spectra Sp obtained from the output of the specific layer j in accordance with the evaluation data and all the known feature spectra KSp associated with the specific layer j and the self-class i, respectively, and
(2) setting a maximum value among the plurality of self-class spectral similarities RSp_A as the self-class maximum spectral similarity RSp_maxA.
(1) obtaining the different-class spectral similarity RSp_B that is a similarity between all the feature spectra Sp obtained from the output of the specific layer j in accordance with the evaluation data and all the known feature spectra KSp associated with the specific layer j and the different-class i, respectively, and
(2) setting a maximum value among the plurality of different-class spectral similarities RSp_B as the different-class maximum spectral similarity RSp_maxB. Maximum spectral similarity RSp_max
(1) obtaining the spectral similarity RSp_A that is a spectral similarity between the feature spectrum Sp obtained from the output of the specific partial region Rn of the specific layer j in accordance with the evaluation data ED and all the known feature spectra KSp associated with the specific layer j, and
(2) setting a maximum value among the plurality of spectral similarities RSp as the maximum spectral similarity RSp_max.
According to the third calculation method M3, the maximum spectral similarities RSp_maxA, RSp_maxB, and RSp_max can be obtained by a simpler calculation and procedure.
A-4. For the distribution and representative value of the spectral similarities:
For example, any one of the following three methods of a first representative value calculation method MR1 to a third representative value calculation method MR3 can be adopted as a calculation method of the distribution of the spectral similarities and the representative values described above.
(1) First representative value calculation method MR1:
An individual local spectral similarity SI_max, which is a maximum value of a spectral similarity RSp between the known feature spectrum KSp obtained from one training data TD and the feature spectrum Sp of the evaluation data ED in each partial region Rn, is obtained for each of the plurality of pieces of training data TD to create a histogram as a distribution. A partial representative value that is a representative value of the histogram created for each partial region Rn is determined. As the partial representative value, a maximum value, a median value, or a mode value of the histogram can be used. In the first representative value calculation method MR1, a correspondence between the known feature spectrum KSp and the partial region Rn in the feature spectrum Sp of the evaluation data ED is considered.
Alternatively, the partial representative value may be determined by the following method. In this method, first, at least one unimodal distribution is obtained by fitting a histogram created for each partial region Rn with a mixed Gaussian distribution using an expectation-maximization algorithm (EM algorithm). When a plurality of the unimodal distributions are generated, one representative unimodal distribution is determined by using the following selection conditions C1 and C2. When one unimodal distribution is obtained, this unimodal distribution is set as the representative unimodal distribution.
Condition C1: A ratio of the area of one unimodal distribution to the entire area of the histogram of the individual local spectral similarity SI_max is equal to or greater than an area threshold.
Condition C2: A mean value of the individual local spectral similarities SI_max is the largest in the unimodal distributions satisfying the condition C1.
The area threshold in the condition C1 is set to a value of, for example, about 5 to 10%. The representative value of the representative unimodal distribution is the partial representative value of the histogram created for each partial region Rn. The representative value of the representative unimodal distribution is, for example, a mode value of the representative unimodal distribution.
Then, each partial representative value is subjected to an integration process to generate the representative value PRSp of the distribution of the spectral similarity RSp. The integration process may be, for example, a process of setting the maximum value of the plurality of partial representative values as the representative value PRSp or a process of setting the average value of the plurality of partial representative values as the representative value PRSp.
(2) Second representative value calculation method MR2:
An individual local spectral similarity SI_max, which is a maximum value of a spectral similarity RSp between the known feature spectrum KSp obtained from one training data TD and the feature spectrum Sp of the evaluation data ED in each partial region Rn, is obtained for each of the plurality of pieces of training data TD to create a histogram as a distribution. A partial representative value that is a representative value of the histogram created for each partial region Rn is determined. As the partial representative value, a maximum value, a median value, or a mode value of the histogram can be used. The second representative value calculation method MR2 is different from the first representative value calculation method MR1 in that corresponding partial regions Rn in the known feature spectrum KSp and the feature spectrum Sp of the evaluation data ED are compared with each other. As in the case of the first representative value calculation method MR1, the representative value PRSp may be determined from at least one unimodal distribution obtained by fitting with the mixed Gaussian distribution using the expectation-maximization algorithm (EM algorithm).
(3) Third representative value calculation method MR3:
Without considering the partial region Rn at all, an individual spectral similarity Sa_max, which is a maximum value of the spectral similarity between the known feature spectrum KSp obtained from one piece of training data TD and the spectral similarity RSp with the feature spectrum Sp of the evaluation data ED, is obtained for each of the plurality of pieces of training data TD to create a histogram as a distribution. The maximum value, the median value, or the mode value of the created histogram is used as the representative value PRSp. As in the case of the first representative value calculating method MR1, the representative value PRSp may be determined from at least one unimodal distribution obtained by fitting with the mixed Gaussian distribution using the expectation-maximization algorithm (EM algorithm).
In the first to third representative value calculation methods MR1 to MR3, when the evaluation data ED is the training data TD or the verification data VD, the representative value PRSp_A of the self-class spectral similarity RSp_A and the representative value PRSp_B of the different-class spectral similarity RSp_B are calculated separately for the self-class and the different-class. On the other hand, when the evaluation data ED is the abnormal data AD, the representative value PRSp is calculated without distinguishing between the self-class and the different-class.
In the first representative value calculation method MR1, the individual local spectral similarity SI_max (i, j, k, q) is calculated using the following equation.
SI_max (i, j, k, q)=max [G {Sp (j, k), KSp (i, j, k=all, q)}] (d1)
Further, when the evaluation data ED is the abnormal data AD, the above parameter i specifies all classes, so KSp (i, j, k=all, q) in the above equation (d1) becomes KSp (i=all, j, k=all, q). The individual local spectral similarity SI_max in the above equation (d1) can be calculated in the process of calculating the local spectral similarity S (i, j, k) using the first calculation method M1.
In the second representative value calculation method MR2, the individual local spectral similarity SI_max (i, j, k, q) is calculated using the following equation.
SI_max (i, j, k, q)=max [G {Sp (j, k), KSp (i, j, k, q)}] (d2)
Further, when the evaluation data ED is the abnormal data AD, the above parameter i specifies all classes, so KSp (i, j, k, q) in the above equation (d2) becomes KSp (i=all, j, k, q). The individual local spectral similarity SI_max in the above equation (d2) can be calculated in the process of calculating the local spectral similarity S (i, j, k) using the second calculation method M2.
For example, when the evaluation data ED is the training data TD or the verification data VD, the similarity calculation unit 310 obtains the representative value PRSp from each partial representative value calculated from the histogram corresponding to the partial region Rn for each of the self-class and the different-class.
In the third representative value calculation method MR3, the individual spectral similarity Sa_max (i, j, q) is calculated using the following equation.
Sa_max (i, j, q)=max [G {Sp (j, k=all), KSp (i, j, k=all, q)}] (d3) Further, when the evaluation data ED is the abnormal data AD, the above parameter i specifies all classes, so KSp (i, j, k=all, q) in the above equation (d3) becomes KSp (i=all, j, k=all, q). The individual spectral similarity Sa_max can be calculated in the process of calculating the spectral similarity RSp using the third calculation method M3.
A-5. Data similarity calculation method:
As described above, the step illustrated in
A-6. Detailed description of the evaluation process:
As illustrated in
When the evaluation data ED is the training data TD, in step S32 illustrated in
When the determination result in step S32 is “Yes”, in step S33, the evaluation unit 330 determines whether or not a value DaBm indicated by the different-class data similarity information IDa_B is equal to or greater than a first different-class data similarity threshold thDaBm1. The value DaBm indicated by the different-class data similarity information IDa_B is the different-class maximum data similarity Da_B in the present exemplary embodiment. As described above, in step S33, a second training comparison result that is a comparison result between the value DaBm indicated by the different-class data similarity information IDa_B and the first different-class data similarity threshold thDaBm1 is generated.
When the determination result in step S33 is “Yes”, that is, when the value SpBr is equal to or greater than the first different-class spectral similarity threshold thSpBr1 and the value DaBm is equal to or greater than the first different-class data similarity threshold thDaBm1, it is assumed that the following event occurs. The trained machine learning model 200 determines that the evaluation input data IM that is input data of the training data TD as the evaluation data ED, and the input data IM of the training data TD having a class different from that of the evaluation data ED and having the known feature spectrum KSp similar to the feature spectrum Sp of the evaluation input data IM, are similar as data. Therefore, originally, it may be preferable that the input data IM of the different-class similar to the evaluation input data IM is classified to be the self-class. Originally, it may be preferable that the evaluation input data IM and the input data IM of the training data TD are determined not to be similar to each other as data. Therefore, in this case, in step S34, the evaluation unit 330 generates, as the second explanatory information SEI, information including at least one of a fact that there is inappropriate incomplete data in the training data TD as the plurality of pieces of evaluation data ED, and a fact that information of the training data TD as the evaluation data ED used for an evaluation is insufficient as information necessary for class discrimination. The incomplete data means data having a feature remarkably similar to normal training data of a different class. The fact that the information of the training data TD is insufficient as the information necessary for the class discrimination is assumed to be, for example, a case where the resolution of the input data IM of the training data TD is too low or a case where the input data IM is in a region different from the region necessary for the class discrimination.
When the determination result in step S33 is “No”, that is, when the value SpBr is equal to or greater than the first different-class spectral similarity threshold thSpBr1 and the value DaBm is less than the first different-class data similarity threshold thDaBm1, it is assumed that the following event occurs. When the different-class spectrum representative value PRSp_B as the value SpBr is large, it is assumed that the trained machine learning model 200 determines that the evaluation input data IM of the evaluation data ED is similar in feature to the training data TD of a different-class that is different from the class indicated by the prior label LB associated with the evaluation input data IM. On the other hand, the fact that the different-class maximum data similarity Da_B, which is the value DaBm, is low, means that the evaluation input data IM and the input data IM of the different-class training data TD determined to have similar feature by the trained machine learning model 200 are not similar in terms of data in terms of an index for calculating data similarity such as a mean square error (MSE). In this case, it is considered that the trained machine learning model 200 may not be able to correctly perform class discrimination of the evaluation data ED. Therefore, in step S35, the evaluation unit 330 generates, as the second explanatory information SEI, information indicating that there is a possibility that the machine learning model 200 lacks capability of correctly performing class discrimination of the evaluation data ED.
In step S34 and step S35, the secondary explanatory information SEI is generated by using the first training comparison result generated in step S32 and the second training comparison result generated in step S33. Note that step S34 and step S35 may be executed regardless of the number of pieces of the evaluation data ED satisfying the condition, or may be executed when the number of pieces of the evaluation data ED satisfying the condition is equal to or greater than a predetermined threshold.
When the determination result in step S32 is “No”, in step S36, the evaluation unit 330 determines whether or not a value SpAr indicated by the self-class spectral similarity information IRSp_A is less than a first self-class spectral similarity threshold thSpAr1. In the present exemplary embodiment, the value SpAr is the self-class spectrum representative value PRSp_A. The evaluation unit 330 executes step S36 for each of the plurality of pieces of evaluation data ED, and counts a number NmSpAr of pieces of the evaluation data ED satisfying the condition that the value SpAr is less than the first self-class spectral similarity threshold thSpAr1. Next, in step S37, the evaluation unit 330 determines whether or not the number NmSpAr is equal to or greater than a predetermined first data-threshold thNm1. When the determination result in step S37 is “Yes”, it is assumed that the following event occurs. That is, when the number of pieces of evaluation data ED satisfying the condition is large, it is assumed that the machine learning model 200 determines that the features of the evaluation input data IM and the input data IM of the training data group TDG of the same class as the class indicated by the prior label LB associated with the evaluation input data IM are not similar to each other. Therefore, in this case, there is a possibility that the features of a large number of pieces of the evaluation input data IM are deviated from the features of the input data IM of the other training data TD belonging to the same class. Therefore, in step S38, the evaluation unit 330 generates first training evaluation information indicating that there is a large variation between the features of the plurality of pieces of input data IM included in the plurality of pieces of training data TD used as the evaluation data ED.
When the determination result in step S37 is “No”, step S39 is executed. That is, when the number of pieces of evaluation data ED satisfying the condition is small, the evaluation unit 330 generates, as the second explanatory information SEI, second training evaluation information indicating that there is outlier data in the plurality of pieces of input data IM included in the plurality of pieces of training data TD used as the evaluation data ED. The outlier data means data having characteristics significantly different from those of a set of normal training data in general.
When the determination result in step S36 is “No” for each evaluation data ED, in step S40, the evaluation unit 330 generates, as the second explanatory information SEI, information indicating that the evaluation of the machine learning model 200 is normal.
As illustrated in
When the determination result in step S52 is “Yes”, in step S53, the evaluation unit 330 determines whether or not a value DaBm indicated by the different-class data similarity information IDa_B is equal to or greater than a second different-class data similarity threshold thDaBm2, as in step S33. In step S53, a second verification comparison result that is a comparison result between the value DaBm indicated by the different-class data similarity information IDa_B and the different-class data similarity threshold thDaBm is generated. The second different-class data similarity threshold thDaBm2 may be the same as or different from the first different-class data similarity threshold thDaBm1.
When the determination result in step S53 is “Yes”, that is, when the value SpBr is equal to or greater than the second different-class spectral similarity threshold thSpBr2 and the value DaBm is equal to or greater than the second different-class data similarity threshold thDaBm2, it is assumed that an event similar to the case where the evaluation-data ED is the training data TD has occurred. Therefore, in step S54, the evaluation unit 330 generates, as the second explanatory information SEI, information including at least one of a fact that there is inappropriate incomplete data in the plurality of pieces of verification data ED, and a fact that information of the verification data ED is insufficient as information necessary for class discrimination.
When the determination result in step S53 is “No”, that is, when the value SpBr is equal to or greater than the second different-class spectral similarity threshold thSpBr2 and the value DaBm is less than the second different-class data similarity threshold thDaBm2, it is assumed that an event similar to the case where the evaluation-data ED is the training data TD has occurred. Therefore, in step S55, the evaluation unit 330 generates, as the second explanatory information SEI, information indicating that there is a possibility that the machine learning model 200 lacks capability of correctly performing class discrimination of the verification data VD that is the evaluation data ED.
In step S54 and step S55, the secondary explanatory information SEI is generated using the first verification comparison result generated in step S52 and the second verification comparison result generated in step S53. Note that step S54 and step S55 may be executed regardless of the number of pieces of the evaluation data ED satisfying the condition, or may be executed when the number of pieces of the evaluation data ED satisfying the condition is equal to or greater than a predetermined threshold.
When the determination result in step S52 is “No”, in step S56, the evaluation unit 330 determines whether or not the value SpAr indicated by the self-class spectral similarity information IRSp_A is less than a second self-class spectral similarity threshold thSpAr2. When the verification data VD is used as the evaluation data ED, the value SpAr is, for example, the self-class maximum spectral similarity RSp_maxA. When the determination result in step S56 is “No” for all the evaluation data ED, in step S58, the evaluation unit 330 generates information indicating that the evaluation of the machine learning model 200 is normal as the second explanatory information SEI. The second self-class spectral similarity threshold thSpAr2 may be the same as or different from the first self-class spectral similarity threshold thSpAr1.
When the determination result in step S56 is “Yes”, in step S57, the evaluation unit 330 determines whether or not the value DaAm indicated by the self-class data similarity information IDa_A is equal to or greater than a self-class data similarity threshold thDaAm. The value DaAm indicated by the self-class data similarity information IDa_A is the self-class maximum data similarity Da_A in the present exemplary embodiment. When the determination result in step S57 is “Yes” in each the evaluation data ED, that is, when the value SpAr indicated by the self-class spectral similarity information IRSp_A is less than the second self-class spectral similarity threshold thSpAr2 and the value DaAm indicated by the self-class data similarity information IDa_A is equal to or greater than the self-class data similarity threshold thDaAm, it is assumed that the following events have occurred. In other words, the machine learning model 200 determines that the feature of the evaluation input data IM which is the input data of the verification data VD is not similar to the feature of the input data IM of the training data TD of the same self-class as the class indicated by the prior class LB of the verification data VD. On the other hand, the input data IM of the evaluation input data IM and the evaluation input data IM, which are determined to have the most similar feature by the trained machine learning model 200, are similar as data in the index for calculating the similarity of data such as the mean square error (MSE). As described above, the machine learning model 200 may determine that even the evaluation input data IM similar to the input data IM used for training is in a different class if it is not used for training. Therefore, in step S59, the evaluation unit 330 generates information indicating that over-training of the machine learning model 200 occurs as the second explanatory information SEI.
When the determination result in step S57 is “No”, the evaluation unit 330 executes step S56 and step S57 for each of the plurality of pieces of evaluation data ED, and counts a number NmCa of pieces of evaluation data ED satisfying the conditions that the value SpAr is less than the second self-class spectral similarity threshold thSpAr2 and the value DaAm is less than the self-class data similarity threshold thDaAm. Next, in step S60, the evaluation unit 330 determines whether or not the number NmCa is equal to or greater than a predetermined second data-threshold thNm2. When “Yes” in step S60, i.e. when the number NmCa is equal to or greater than the predetermined second data-threshold thNm2, it is assumed that the following events have occurred. In other words, the machine learning model 200 determines that the feature of many pieces of evaluation input data IM is not similar to the feature of the input data IM of the training data TD of the same self-class as the class indicated by the prior class LB of the verification data VD. In addition, the machine learning model 200 determines that the evaluation input data IM and the input data IM determined to be most similar as a feature are not similar to each other in terms of data similarity. When the number of pieces of verification data VD for which the above-described determination has been performed is large, it is assumed that there is a large difference in feature between the evaluation input data IM of the verification data VD and the input data IM of the training data TD. Therefore, in step S62, the evaluation unit 330 generates, as the second explanatory information SEI, the first verification evaluation information indicating that there is a large difference in feature between each evaluation input data IM as each input data included in the plurality of pieces of verification data and the input data IM included in the training data TD.
When “No” in step S60, i.e. when the number NmCa is less than the predetermined second data-threshold thNm2, it is assumed that the following events have occurred. That is, it is assumed that there is the outlier data in the plurality of pieces of input data IM included in the plurality of pieces of verification data VD. Therefore, in step S64, the evaluation unit 330 generates, as the second explanatory information SEI, second verification evaluation information indicating that there is the outlier data in the plurality of pieces of input data IM included in the plurality of pieces of verification data VD.
When the evaluation data ED illustrated in
As illustrated in
When the determination result in step S72 is “Yes”, the evaluation unit 330 determines whether or not the maximum data similarity Da_max is equal to or greater than a predetermined abnormal-data similarity threshold thDa. When the determination result in step S73 is “Yes”, that is, when the maximum spectral similarity RSp_max is equal to or greater than the abnormal spectral threshold thRSp and the maximum data similarity Da_max is equal to or greater than the abnormal-data similarity threshold thDa, it is assumed that the following events have occurred. That is, when the evaluation data ED is the abnormal data AD, it is expected that the maximum data similarity Da_max is low. However, when the maximum data similarity Da_max is equal to or greater than the abnormal data similarity threshold thDa, it means that the abnormal data AD does not have information indicating abnormality. Therefore, in step S74, the evaluation unit 330 generates, as the second explanatory information SEI, information indicating that information of the abnormal data AD is insufficient as information necessary for class discrimination.
When the determination result in step S73 is “No”, that is, when the maximum spectral similarity RSp_max is equal to or greater than the abnormal spectral threshold thRSp and the maximum data similarity Da_max is less than the abnormal-data similarity threshold thDa, it is assumed that the following events have occurred. That is, the same event occurs as step S35 is executed in the case where the evaluation data ED is the training data TD or as step S55 is executed in the case where the evaluation data ED is the verification data VD. Therefore, in step S75, the evaluation unit 330 generates, as the second explanatory information, information indicating that there is a possibility that the machine learning model 200 lacks capability of correctly performing class discrimination of the abnormal data AD, that is, the capability of discriminating an unknown-class.
The various thresholds used in the evaluation process described above with reference to
A-7. Addressing method:
An addressing method for the generated secondary explanatory information SEI will be described below. When step S35 and step S55 are executed and information indicating that there is a possibility that the machine learning model 200 cannot correctly perform class discrimination of the evaluation data ED is generated as the second explanatory information SEI, the following addressing method may be taken.
The network configuration of the machine learning model 200 is reviewed.
In the addressing method 1A, for example, the network configuration is reviewed by increasing the number of vector neuron layers. In the addressing method 1A, for example, the number of layers of the network of the machine learning model 200 is increased, or the specific layer from which the feature spectrum Sp is acquired is changed.
When step S75 is executed and information indicating that there is a possibility that the machine learning model 200 cannot correctly perform class discrimination of the evaluation data ED is generated as the second explanatory information SEI, the following addressing method may be taken.
The machine learning model 200 is trained by correcting the training data TD to training data more suitable for training for class discrimination.
In the addressing method 1B, for example, the input data IM of the training data TD is subjected to data processing for deleting elements indicating simple features. Further, for example, the prior label LB is further subdivided and associated with the input data IM. Note that the addressing method 1B may include the same addressing method as the addressing method 1A.
When step S59 is executed and information indicating that the over-training of the machine learning model 200 occurs is generated, the following addressing method may be taken.
The training parameter of the machine learning model 200 is reviewed.
In the addressing method 1C, for example, the number of epochs is reduced, or the batch size in mini-batch training is reduced.
When information indicating that the information of the evaluation data ED is insufficient as information necessary for class discrimination is generated at step S34, step S54, or step S74, the following addressing methods may be taken.
The data resolution of the input data IM is changed.
For example, the data resolution of a characteristic region for class discrimination is increased.
Pre-processing such as average difference is introduced to original data of data IM for input.
The acquisition condition of the input data IM is reviewed.
For example, the distance between the target object and the imaging device is changed, or the target object is irradiated with light.
When information indicating that there is incomplete data among the plurality of pieces of evaluation data ED is generated in step S34 or step S54, the following addressing methods may be taken.
The prior label LB of the training data TD is reviewed.
For example, when a different prior label LB is associated with the plurality of pieces of input data IM having similar feature, the same prior label LB is newly associated with the plurality of pieces of input data IM.
The training data TD is reduced.
For example, when a different prior label LB is associated with the plurality of pieces of input data IM having similar feature, only the input data IM associated with one prior label LB is left, and the remaining pieces of input data IM are deleted.
When information indicating that there is outlier data is generated in step S39 or step S64, the following addressing methods may be taken.
The training data TD is extended.
For example, the input data IM having a feature close to the outlier data is added as the training data TD.
The outlier data is deleted from the training data TD.
Pre-processing is executed on the training data TD and the verification data VD.
Examples of the pre-processing include a smoothing process and a normalization process when it is assumed that there is outlier data due to noise.
When the first training evaluation information is generated in step S38, the following addressing methods may be taken.
The training data TD is extended.
For example, the input data IM for reducing the variation is newly added as the training data TD.
The input data IM causing the increase in the variation is deleted from the training data TD.
Pre-processing is executed on the training data TD.
Examples of the pre-processing include the smoothing process and normalization process when it is assumed that the variation is large due to noise.
In step S62, when the first verification evaluation information is generated, the following addressing methods may be taken.
The training data TD is extended.
The training data TD is newly added so that the deviation of the feature between each evaluation input data IM and the input data IM included in the training data TD becomes small.
Pre-processing is executed on the training data TD.
Examples of the pre-processing include the smoothing process and normalization process.
In step S74, when information indicating that the information of the abnormal data AD is insufficient as the information necessary for class discrimination is generated, the following addressing method may be taken.
The specific layer is changed.
For example, when a feature that affects the class discrimination appears in a fine shape, a lower layer is set as the specific layer. For example, the specific layer is changed from the ConvVN2 layer 250 to the ConvVN1 layer 240.
The evaluation unit 330 may display the contents of the addressing methods 1A to 1K described above on the display unit 150. Accordingly, it is possible to easily grasp the addressing method regardless of the experience of the user.
A-8. Specific examples of the addressing method:
A-8-1. First specific example:
The plurality of pieces of training data TD are input as the evaluation data ED to the trained machine learning model 200, and class discrimination is executed using an activation value corresponding to a determination value of each class output from the ClassVN layer 260. As a result, a correct answer rate of the class discrimination is lower than the desired value. In this example, the ClassVN layer 260 is set as the specific layer. The evaluation unit 330 generates the secondary explanatory information SEI using the first explanatory information FEI. In this case, as the second explanatory information SEI, information indicating that there is a possibility that the machine learning model 200 cannot correctly perform class discrimination of the training data TD as the evaluation data ED is generated in step S55 illustrated in
A-8-2. Second specific example:
The plurality of pieces of verification data VD are input as the evaluation data ED to the trained machine learning model 200, and class discrimination is executed using an activation value corresponding to a determination value of each class output from the ClassVN layer 260. As a result, the correct answer rate of the class discrimination is lower than the desired value. In this example, the ClassVN layer 260 is set as the specific layer. The evaluation unit 330 generates the secondary explanatory information SEI using the first explanatory information FEI. In this case, the first verification evaluation information is generated in step S62 illustrated in
A-8-3. Third specific example:
The abnormal data AD expected to be class-discriminated as an unknown-class is input as the evaluation data ED to the trained machine learning model 200, and the evaluation process is executed. The class discrimination is executed using an activation value corresponding to a determination value of each class output from the ClassVN layer 260. In this example, the ClassVN layer 260 is set as the specific layer. The evaluation unit 330 generates the secondary explanatory information SEI using the first explanatory information FEI. In this case, as the second explanatory information SEI, information indicating that the information of the abnormal data AD is insufficient as information necessary for class discrimination is generated in step S74 illustrated in
A-9. Calculation method for an output vector of each layer of the machine learning model:
A calculation method of the output of each layer in the machine learning model 200 illustrated in
Each node of the PrimeVN layer 230 regards the scalar outputs of the 1×1×32 nodes of the Conv layer 220 as a 32-dimensional vector, and multiplies this vector by a transformation matrix to obtain vector outputs of the node. The transformation matrix is an element of a kernel having a surface size of 1×1 and is updated by training of the machine learning model 200. It is also possible to integrate the processes of the Conv layer 220 and the PrimeVN layer 230 into one primary vector neuron layer.
When the PrimeVN layer 230 is referred to as a “lower layer L” and the ConvVN1 layer 240 adjacent to the upper side thereof is referred to as an “upper layer L+1”, the output of each node of the upper layer L+1 is determined using the following equation.
Where
MLi is an output vector of the i-th node in the lower layer L;
ML+1j is an output vector of the j-th node in the upper layer L+1;
vij is a prediction vector of the output vector ML+1j;
WLij is a prediction matrix for calculating the prediction vector vij from the output vector MLi of the lower layer L;
uj is a sum vector which is the sum, i.e., linear combination, of the prediction vectors vij;
aj is an activation value which is a normalization coefficient obtained by normalizing a norm |uj| of the sum vector uj; and
F (X) is a normalization function for normalizing X.
As the normalizing function F (X), for example, the following equation (E3a) or equation (E3b) can be used.
Where,
k is an ordinal number for all nodes of the upper layer L+1; and
β is an adjustment parameter which is an arbitrary positive coefficient, for example, β=1.
In the above equation (E3a), the activation value aj is obtained by normalizing the norm |uj| of the sum vector uj with respect to all nodes of the upper layer L+1 by the softmax function. On the other hand, in the (E3b) equation, the activation value aj is obtained by dividing the norm |uj| of the sum vector uj by the sum of the norms |uj| of all the nodes of the upper layer L+1. As the normalizing function F (X), the (E3a) equation or a function other than the (E3b) equation may be used.
The ordinal number i in the above equation (E2) is conveniently assigned to a node in the lower layer L used to determine the output vector ML+1j of the j-th node in the upper layer L+1, and takes a value of 1 to n. An integer n is the number of nodes in the lower layer L used to determine the output vector ML+1j of the j-th node in the upper layer L+1. Thus, the integer n is given by:
N=Nk×Nc (E5)
Where, Nk is a surface size of the kernel, and Nc is a number of channels of the PrimeVN layer 230, which is the lower layer. In the example of
One kernel used to obtain the output vector of the ConvVN1 layer 240 has 3×3×16=144 elements with a kernel size of 3×3 and a depth of 16 channels in the lower layer, and each of these elements is a prediction matrix WLij. Also, 12 sets of these kernels are required to generate output vectors for the 12 channels of the ConvVN1 layer 240. Therefore, the number of prediction matrices WLij of the kernel used to obtain the vectors outputted from the ConvVN1 layer 240 is 144×12=1728. These prediction matrices WLij are updated by training of the machine learning model 200.
As can be seen from the above-described (E1) to (E4) equations, the output vectors ML+1j of the individual nodes of the upper layer L+1 are obtained by the following calculations.
(a) An output vector MLi of each node of the lower layer L is multiplied by a prediction matrix WLij to obtain a prediction vector vij;
(b) A sum vector uj of the prediction vectors vij obtained from each node of the lower layer L, i.e. a linear combination, is obtained;
(c) A norm |uj| of the sum vector uj is normalized to obtain an activation value aj; and
(d) The sum vector uj is divided by the norm |uj| and further multiplied by the activation value aj.
The activation value aj is a normalization coefficient obtained by normalizing the norm |uj| for all the nodes in the upper layer L+1. Therefore, the activation value aj can be considered as an index indicating the relative output intensity of each node among all nodes in the upper layer L+1. The norm used in equations (E3), (E3a), (E3b), and (4) is typically the L2 norm representing the vector length. In this case, the activation value aj corresponds to the vector length of the output vector ML+1j. Since the activation value aj is only used in the above-described (E3) equation and (E4) equation, it does not need to be output from the node. However, the upper layer L+1 can also be configured to output the activation value aj to the outside.
The configuration of the vector neural network is substantially the same as the configuration of the capsule network, and the vector neuron of the vector neural network correspond to the capsule of the capsule network. However, the calculation by the above-described equations (E1) to (E4) used in the vector neural network is different from the calculation used in the capsule network. The most significant difference therebetween is that, in the capsule network, the prediction vector vij on the right side of the above equation (E2) is each multiplied by a weight, and the weight is searched by repeating dynamic routing a plurality of times. On the other hand, in the vector neural network of the present exemplary embodiment, since the output vector ML+1j is obtained by sequentially calculating the above-described equations (E1) to (E4) once, there is no need to repeat the dynamic routing, and there is an advantage that the calculation is faster. In addition, the vector neural network of the present exemplary embodiment has an advantage that a memory amount required for calculation is less than that of the capsule network, and according to an experiment of the inventor of the present disclosure, the required memory amount is only about ½ to ⅓.
The vector neural network is the same as the capsule network in that it uses nodes that input and output vectors. Thus, the advantages of using vector neurons are in common with capsule networks. In addition, the plurality of layers 210 to 250 are the same as the ordinary convolutional neural network in that a higher layer expresses a feature of a greater region and a lower layer expresses a feature of a smaller region. Here, the “feature” means a characteristic portion included in input data to the neural network. Vector neural networks and capsule networks are superior to ordinary convolutional neural networks in that the output vector of a certain node contains spatial information representing the spatial information of the feature expressed by that node. That is, the vector length of an output vector of a certain node represents the existence probability of a feature expressed by the node, and the vector direction represents the spatial information such as the direction and scale of the feature. Therefore, the vector directions of the output vectors of two nodes belonging to the same layer represent the positional relationship of the respective features. Alternatively, the vector direction of the output vector of the two nodes represents the variation of the feature. For example, for a node corresponding to an “eye” feature, the direction of the output vector may represent variations in the thinness of the eye, how it is lifted, etc. In the ordinary convolutional neural network, it is said that spatial information of a feature is lost by a pooling process. As a result, the vector neural network and the capsule network have an advantage that they are excellent in the performance of identifying input data as compared with the ordinary convolutional neural network.
The advantage of the vector neural network can be considered as follows. That is, in a vector neural network, it is advantageous that the output vector of the node expresses the feature of the input data as coordinates in a continuous space. Therefore, the output vector can be evaluated such that if the vector directions are close to each other, the features are similar. In addition, even when a feature included in input data cannot be covered by supervised data, the feature can be discriminated by interpolation. On the other hand, the ordinary convolutional neural network has a drawback in that the feature of input data cannot be expressed as coordinates in a continuous space because random compression is applied by the pooling processing.
Since the output of each node of the ConvVN2 layer 250 and the ClassVN layer 260 is determined in the same manner using the above-described (E1) to (E4) equations, a detailed description thereof will be omitted. The resolution of the ClassVN layer 260, which is the uppermost layer, is 1×1, and the channel number thereof is M.
The output of the ClassVN layer 260 is converted into a plurality of determination values Class1 to Class2 for a known class. These determination values are normally values normalized by the softmax function. Specifically, for example, the determination value for each class can be obtained by calculating a vector length of the output vector from the output vector of each node of the ClassVN layer 260, and further normalizing the vector length of each node by the softmax function. As described above, the activation value aj obtained by the above equation (E3) is a value corresponding to the vector length of the output vector ML+1j and is normalized. Therefore, the activation value aj in each node of the ClassVN layer 260 may be output and used as it is as the determination value for each class.
In the above-described exemplary embodiment, as the machine learning model 200, a vector neural network that obtains an output vector by calculation of the above-described (E1) equation to (E4) equation is used. However, instead of this, the capsule network disclosed in U.S. Pat. No. 5,210,798 or WO 2009/083553 may be used.
According to the above-described exemplary embodiment, it is possible to generate and output the second explanatory information SEI indicating the evaluation of the trained machine learning model 200 using the first explanatory information FEI including the spectral similarity information IRSp and the data similarity information IDa. Accordingly, it is possible to evaluate the trained machine learning model without causing a difference between users. In addition, based on the evaluation of the machine learning model 200, it is possible to efficiently improve the machine learning model 200, such as increasing the correct answer rate.
B. Other Aspects:
The present disclosure is not limited to the embodiments described above, and may be implemented in various aspects without departing from the spirits of the disclosure. For example, the present disclosure can be realized by the following aspects. Appropriate replacements or combinations may be made to the technical features in the above-described embodiments which correspond to the technical features in the aspects described below to solve some or all of the problems of the disclosure or to achieve some or all of the advantageous effects of the disclosure. Additionally, when the technical features are not described herein as essential technical features, such technical features may be deleted appropriately.
(1) According to a first aspect of the present disclosure, there is provided an evaluation method for a trained machine learning model. The machine learning model is a vector neural network model including a plurality of vector neuron layers, and is trained by using a plurality of pieces of training data including input data and a prior label associated with the input data. The evaluation method includes the steps of (a) inputting evaluation data to the trained machine learning model to generate first explanatory information used for an evaluation of the machine learning model, (b) using a value indicated by each piece of information included in the first explanatory information to generate second explanatory information indicating an evaluation of the trained machine learning model, and (c) outputting the generated second explanatory information, wherein the step (a) includes the steps of (a1) inputting the evaluation data to the trained machine learning model to obtain a feature spectrum from an output of a specific layer of the trained machine learning model, (a2) obtaining a spectral similarity that is a similarity between the feature spectrum and a known feature spectrum included in a known feature spectrum group obtained from an output of the specific layer by inputting the plurality of pieces of training data to the trained machine learning model again, for each of a plurality of the known feature spectra included in the known feature spectrum group, (a3) obtaining a data similarity that is a similarity between the input data and the evaluation data, and (a4) generating the first explanatory information including spectral similarity information related to the spectral similarity and data similarity information related to the data similarity. According to this aspect, it is possible to generate and output the second explanatory information indicating the evaluation of the trained machine learning model using the first explanatory information including the spectral similarity information and the data similarity information. Therefore, it is possible to evaluate the trained machine learning model without causing a difference between users.
(2) In the above aspect, when at least one of (i) the training data and (ii) verification data is used as the evaluation data in the step (a), the verification data being not used for training of the machine learning model and including the input data and the prior label associated with the input data, the step (a2) may include the steps of obtaining a self-class spectral similarity that is the spectral similarity between the feature spectrum and a self-class known feature spectrum of the same class as an evaluation class indicated by the prior label associated with the evaluation data among the known feature spectrum group, for each of a plurality of the self-class known feature spectra, and obtaining a different-class spectral similarity that is the spectral similarity between the feature spectrum and a different-class known feature spectrum of a class different from the evaluation class among the known feature spectrum group, for each of a plurality of the different-class known feature spectra, the step (a3) may include the steps of obtaining a self-class maximum data similarity that is the similarity between the input data associated with the self-class known feature spectrum that is a calculation source of a self-class maximum spectral similarity indicating a maximum value among a plurality of the self-class spectral similarities and the evaluation data, and obtaining a different-class maximum data similarity that is the similarity between the input data associated with the different-class known feature spectrum that is a calculation source of a different-class maximum spectral similarity indicating a maximum value among a plurality of the different-class spectral similarities and the evaluation data, and the step (a4) may include the step of generating the first explanatory information including self-class spectral similarity information related to the self-class maximum spectral similarity, self-class data similarity information related to the self-class maximum data similarity, different-class spectral similarity information related to the different-class maximum spectral similarity, and different-class data similarity information related to the different-class maximum data similarity. According to this aspect, it is possible to generate and output the second explanatory information indicating a more detailed evaluation of the trained machine learning model using the first explanatory information including more types of information. This makes it possible to efficiently improve the machine learning model based on the evaluation of the machine learning model.
(3) In the above aspect, when the training data is used as the evaluation data in the step (a), the step (b) may include the step of (b1) generating the second explanatory information using a first training comparison result between a value indicated by the different-class spectral similarity information and a predetermined first different-class spectral similarity threshold and a second training comparison result between a value indicated by the different-class data similarity information and a predetermined first different-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information using the first training comparison result and the second training comparison result.
(4) In the above aspect, the step (b1) may include the step of generating, as the second explanatory information, information indicating at least one of a fact that there is inappropriate incomplete data in the training data and a fact that information of the training data is insufficient as information necessary for class discrimination, when the value indicated by the different-class spectral similarity information is equal to or greater than the first different-class spectral similarity threshold and the value indicated by the different-class data similarity information is equal to or greater than the first different-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating a specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(5) In the above aspect, the step (b1) may include the step of generating, as the second explanatory information, information indicating that there is a possibility that the machine learning model lacks capability of correctly performing class discrimination of the evaluation data, when the value indicated by the different-class spectral similarity information is equal to or greater than the first different-class spectral similarity threshold and the value indicated by the different-class data similarity information is less than the first different-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating a specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(6) In the above aspect, when the plurality of pieces of training data are used as the evaluation data in the step (a), the step (b) may include the step of (b2) generating, as the second explanatory information, at least one of first training evaluation information indicating that a variation in the input data included in the plurality of pieces of training data used as the evaluation data is large and second training evaluation information indicating that the input data included in the plurality of pieces of training data used as the evaluation data includes outlier data, when a value indicated by the self-class spectral similarity information is less than a predetermined first self-class spectral similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating a specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(7) In the above aspect, the step (b2) may include the steps of generating the first training evaluation information as the second explanatory information when a number of pieces of the training data as the evaluation data satisfying that the value indicated by the self-class spectral similarity information is less than the first self-class spectral similarity threshold is equal to or greater than a predetermined first data threshold, and generating the second training evaluation information as the second explanatory information when a number of pieces of the training data as the evaluation data satisfying that the value indicated by the self-class spectral similarity information is less than the first self-class spectral similarity threshold is less than the first data threshold. According to this aspect, it is possible to generate the second explanatory information indicating more specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(8) In the above aspect, when the verification data is used as the evaluation data in the step (a), the step (b) may include the step of (b3) generating the second explanatory information using a first verification comparison result between a value indicated by the different-class spectral similarity information and a predetermined second different-class spectral similarity threshold and a second verification comparison result between a value indicated by the different-class data similarity information and a predetermined second different-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information using the first verification comparison result and the second verification comparison result.
(9) In the above aspect, the step (b3) may include the step of generating, as the second explanatory information, information indicating at least one of a fact that there is inappropriate incomplete data in the verification data and a fact that information of the verification data is insufficient as information necessary for class discrimination, when the value indicated by the different-class spectral similarity information is equal to or greater than the second different-class spectral similarity threshold and the value indicated by the different-class data similarity information is equal to or greater than the second different-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating a specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(10) In the above aspect, the step (b3) may include the step of generating, as the second explanatory information, information indicating that there is a possibility that the machine learning model lacks capability of correctly performing class discrimination of the evaluation data, when the value indicated by the different-class spectral similarity information is equal to or greater than the second different-class spectral similarity threshold and the value indicated by the different-class data similarity information is less than the second different-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating more specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(11) In the above aspect, when a plurality of pieces of the verification data are used as the evaluation data in the step (a), the step (b) may include the step of (b4) generating, as the second explanatory information, at least one of first verification evaluation information indicating that a feature difference between the input data included in the plurality of pieces of verification data and the input data included in the training data is large and second verification evaluation information indicating that there is outlier data in a plurality of pieces of the input data included in the plurality of pieces of verification data, when the value indicated by the self-class spectral similarity information is less than a predetermined second self-class spectral similarity threshold and the value indicated by the self-class data similarity information is less than a predetermined self-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating more specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(12) In the above aspect, the step (b4) may include the steps of generating the first verification evaluation information as the second explanatory information when a number of pieces of the verification data as the evaluation data satisfying that the value indicated by the self-class spectral similarity information is less than the second self-class spectral similarity threshold and the value indicated by the self-class data similarity information is less than the self-class data similarity threshold is equal to or greater than a predetermined second data threshold, and generating the second verification evaluation information as the second explanatory information when a number of pieces of the verification data as the evaluation data satisfying that the value indicated by the self-class spectral similarity information is less than the second self-class spectral similarity threshold and the value indicated by the self-class data similarity information is less than the self-class data similarity threshold is less than the second data threshold. According to this aspect, it is possible to generate the second explanatory information indicating more specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(13) In the above aspect, the step (b3) may include the step of generating, as the second explanatory information, information indicating that over-training of the machine learning model occurs when the value indicated by the self-class spectral similarity information is less than the second self-class spectral similarity threshold and the value indicated by the self-class data similarity information is equal to or greater than the self-class data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating a specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(14) In the above aspect, in the step (a4), the self-class spectral similarity information may include information of at least one of a representative value of a distribution of the plurality of self-class spectral similarities and the self-class maximum spectral similarity, and the different-class spectral similarity information may include information of at least one of a representative value of a distribution of the plurality of different-class spectral similarities and the different-class maximum spectral similarity. According to this aspect, the distribution of the spectral similarity or the maximum spectral similarity can be used as the self-class spectral similarity information or the different-class spectral similarity information.
(15) In the above aspect, when abnormal data is used as the evaluation data in the step (a), the abnormal data being not associated with the prior label and being assumed to be classified as an unknown class different from a class corresponding to the prior label, the step (a2) may include the step of specifying a maximum spectral similarity of a maximum value among the spectral similarities obtained for each of the plurality of known feature spectra, the step (a3) may include the step of obtaining a maximum data similarity that is a similarity between the input data associated with the known feature spectrum that is a calculation source of the maximum spectral similarity specified in the step (a2) and the abnormal data, and the step (a4) may include the step of generating the first explanatory information including spectral similarity information related to the spectral similarity and the maximum data similarity. According to this aspect, it is possible to generate and output the second explanatory information using the abnormality data.
(16) In the above aspect, when the abnormal data is used as the evaluation data in the step (a), the step (b) may include the step of generating, as the second explanatory information, information indicating that there is a possibility that the machine learning model lacks capability of correctly performing class discrimination of the abnormal data when the maximum spectral similarity is equal to or greater than a predetermined abnormal spectrum threshold and the maximum data similarity is less than a predetermined abnormal data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating a specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(17) In the above aspect, when the abnormal data is used as the evaluation data in the step (a), the step (b) may include the step of generating, as the second explanatory information, information indicating that information of the abnormal data is insufficient as information necessary for class discrimination when the maximum spectral similarity is equal to or greater than a predetermined abnormal spectrum threshold and the maximum data similarity is equal to or greater than a predetermined abnormal data similarity threshold. According to this aspect, it is possible to generate the second explanatory information indicating a specific evaluation. This makes it possible to improve the machine learning model, the evaluation data, and the training data more efficiently based on the evaluation of the machine learning model.
(18) According to a second aspect of the present disclosure, there is provided an evaluation device for a trained machine learning model. The evaluation device includes a memory configured to store the machine learning model, the machine learning model being a vector neural network model including a plurality of vector neuron layers, the machine learning model being trained by using a plurality of pieces of training data including input data and a prior label associated with the input data, and a processor, wherein the processor is configured to execute the following processes (a) inputting evaluation data to the trained machine learning model to generate first explanatory information used for an evaluation of the machine learning model, (b) using a value indicated by each piece of information included in the first explanatory information to generate second explanatory information indicating an evaluation of the trained machine learning model, and (c) outputting the generated second explanatory information, wherein the process (a) includes the following processes (a1) inputting the evaluation data to the trained machine learning model to obtain a feature spectrum from an output of a specific layer of the trained machine learning model, (a2) obtaining a spectral similarity that is a similarity between the feature spectrum and a known feature spectrum included in a known feature spectrum group obtained from an output of the specific layer by inputting the plurality of pieces of training data to the trained machine learning model again, for each of a plurality of the known feature spectra included in the known feature spectrum group, (a3) obtaining a data similarity that is a similarity between the input data and the evaluation data, and (a4) generating the first explanatory information including spectral similarity information related to the spectral similarity and data similarity information related to the data similarity. According to this aspect, it is possible to generate and output the second explanatory information indicating the evaluation of the trained machine learning model using the first explanatory information including the spectral similarity information and the data similarity information. Therefore, it is possible to evaluate the trained machine learning model without causing a difference between users.
(19) According to a third aspect of the present disclosure, there is provided a non-transitory computer-readable storage medium storing a program for causing a computer to execute evaluation of a trained machine learning model. The machine learning model is a vector neural network model including a plurality of vector neuron layers, and is trained by using a plurality of pieces of training data including input data and a prior label associated with the input data. The computer program causes the computer to execute the following functions (a) inputting evaluation data to the trained machine learning model to generate first explanatory information used for an evaluation of the machine learning model, (b) using a value indicated by each piece of information included in the first explanatory information to generate second explanatory information indicating an evaluation of the trained machine learning model, and (c) outputting the generated second explanatory information, wherein the function (a) includes the following functions (a1) inputting the evaluation data to the trained machine learning model to obtain a feature spectrum from an output of a specific layer of the trained machine learning model, (a2) obtaining a spectral similarity that is a similarity between the feature spectrum and a known feature spectrum included in a known feature spectrum group obtained from an output of the specific layer by inputting the plurality of pieces of training data to the trained machine learning model again, for each of a plurality of the known feature spectra included in the known feature spectrum group, (a3) obtaining a data similarity that is a similarity between the input data and the evaluation data, and (a4) generating the first explanatory information including spectral similarity information related to the spectral similarity and data similarity information related to the data similarity. According to this aspect, it is possible to generate and output the second explanatory information indicating the evaluation of the trained machine learning model using the first explanatory information including the spectral similarity information and the data similarity information. Therefore, it is possible to evaluate the trained machine learning model without causing a difference between users.
The present disclosure can be realized in various forms other than the above. For example, the present disclosure can be realized in a form of a non-transitory storage medium, etc. in which the computer program is recorded.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2021-200534 | Dec 2021 | JP | national |
