CN111860577A

CN111860577A - Artificial intelligence ethical method for identifying human being harmless to human being and robot

Info

Publication number: CN111860577A
Application number: CN202010513310.9A
Authority: CN
Inventors: 朱定局
Original assignee: South China Normal University
Current assignee: South China Normal University
Priority date: 2020-06-08
Filing date: 2020-06-08
Publication date: 2020-10-30

Abstract

An artificial intelligence ethical method and a robot for identifying a human being harmless to the human being, comprising: initializing a screening set; starting a trying step; a third information type behavior evaluation step; a deletable judgment step; and executing the control step. According to the method, the system and the robot, the algorithm bias is reduced or even eliminated by deleting the input variables which cause the bias and have no obvious influence on the prediction result, so that the artificial intelligence ethical risk is reduced.

Description

Artificial intelligence ethical method for identifying human being harmless to human being and robot

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to an artificial intelligence ethical method and a robot for identifying human beings which do not harm human beings.

Background

In the process of implementing the invention, the inventor finds that at least the following problems exist in the prior art: the existing artificial intelligence ethical risk prevention technology has the technical scheme that the algorithm bias detection and prevention focus is focused on the number of samples, for example, the algorithm bias caused by the difference between the number of samples of the B ethnic group and the number of samples of the A ethnic group, but many experiments show that the algorithm bias cannot be completely eliminated even if the number of samples is controlled to be the same, the visible bias is caused to be deeper, and the reason and the solution method are not possessed by the prior art.

Accordingly, the prior art is yet to be improved and developed.

Disclosure of Invention

Based on this, it is necessary to provide an artificial intelligence ethical method and a robot for identifying human beings which are harmless to humans, so as to solve the defects that the algorithm bias detection and prevention of the existing artificial intelligence ethical method does not consider the bias caused by the correlation between the input variables and does not consider the bias caused by the introduction of the input variables of "impurities" irrelevant to the prediction result, and reduce or even eliminate the algorithm bias by deleting the input variables which cause the bias and have no obvious influence on the prediction result, thereby reducing the risk of the artificial intelligence ethical.

In a first aspect, an embodiment of the present invention provides an artificial intelligence method, where the method includes:

a first information type obtaining step: acquiring K preset types as K first information types;

a first object class acquisition step: obtaining the categories of all objects in a training data set of a deep learning model as a first object category;

a second object class acquisition step: acquiring the number M of all sub-categories of the first object category, which need to be detected and prevent prejudice, and taking the M sub-categories as M second object categories;

Calculating the accuracy of the first class classification: taking K first information types as first sample information types, taking objects of the first object type as first sample objects, and calculating to obtain a classification prediction accuracy P1 as a first class classification accuracy;

and j sub-category classification accuracy calculation step: taking K first information types as first sample information types, taking the j (j =1, 2, …, M) th object type as a first sample object, and calculating to obtain the classification prediction accuracy P1 as the j sub-category classification accuracy P1 j;

the ith type first class classification accuracy calculation step: deleting the ith (i =1, 2, …, K) first information type from the K first information types to obtain K-1 first information types as first sample information types, taking the object of the first object class as a first sample object, and calculating to obtain the classification prediction accuracy P1 as the ith first class classification accuracy P1 i;

and (3) calculating the classification test accuracy of the ith sub-category: deleting the ith (i =1, 2, …, K) first information type from the K first information types to obtain K-1 first information types as first sample information types, taking the jth (j =1, 2, …, M) second object type as a first sample object, and calculating to obtain the classification test accuracy P1 as the classification test accuracy P1ij of the ith jth sub-category;

And the correlation evaluation step of the ith first information type and the second object category: for j =1, 2, …, M, calculating the difference DFi between the accuracy of the first class classification test and the accuracy of the ith first class classification test, calculating the difference DFij between the accuracy of each sub-class classification test and the accuracy of each sub-class classification test of the ith, and calculating a measurement index Q1i of the correlation between the ith first information type and the second object type according to the DFi and the DFij;

k types of first information type behavior evaluation steps: taking the K first information types as third information types, calculating to obtain an average value AAM of the behavior evaluation test accuracy rates of the M second object types, taking the average value AAM as an OldAAM, calculating to obtain a standard deviation ADM of the behavior evaluation test accuracy rates of the M second object types, taking the standard deviation ADM as an OldADM, and calculating to obtain a behavior evaluation test prejudice rate IM of the M second object types, and taking the behavior evaluation test prejudice rate IM as an OldIM;

and (3) initializing a screening set: adding the K first information types into a screening set;

a start trial step: extracting a first information type which is not marked as deleted or can not be deleted and has the maximum Q1i from the screening set as a fourth information type, and marking the fourth information type as to be deleted in the screening set;

And a third information type behavior evaluation step: taking all first information types which are not marked to be deleted and deleted in the screening set as third information types, calculating to obtain an average value AAM of behavior evaluation test accuracy rates of M second object types, taking the average value AAM as NewAAM, taking a standard deviation ADM of the behavior evaluation test accuracy rates of the M second object types as NewADM, and taking a behavior evaluation test bias rate IM of the M second object types as NewIM;

a deletable judgment step: judging whether a preset condition is met, if so, marking the fourth information type as deleted in the screening set, and if not, marking the fourth information type as not deletable in the screening set; the preset conditions include NewAAM > (OldAAM-preset tolerance threshold) and NewADM < OldADMM and NewIM < OldIM, or NewAAM > (OldAAM-preset tolerance threshold) and NewADM < OldADMM, or NewAAM > (OldAAM-preset tolerance threshold) and NewIM < OldIM;

executing a control step: and judging whether the first information type which is not marked as deleted or not can be deleted exists in the screening set, if so, returning to the step of starting the trying step to execute the steps again, if not, taking the latest behavior prediction deep learning model as the behavior prediction deep learning model after the bias is eliminated, and taking the first information type which is marked as not-deletable in the screening set as the second information type.

Preferably, the method further comprises:

a first object acquisition step: acquiring an object to be predicted as a first object;

a first object information acquisition step: acquiring the information of the second information type of the first object within a first preset time in the past;

a first use step: inputting the information of the second information type of the first object in a first preset time length in the past into the behavior prediction deep learning model after the bias elimination, and calculating the obtained output as an evaluation value of the behavior of the first object in a second preset time length after the first preset time length;

a first object behavior judging step: and judging whether the evaluation value of the behavior is larger than a preset evaluation value threshold value, if so, judging that the first object is an object which can generate the first type behavior, otherwise, judging that the first object is an object which cannot generate the first type behavior.

Preferably, the first and second electrodes are formed of a metal,

the method further comprises the following steps:

a class training step: obtaining an unsupervised training data set, taking the information of the first sample information type of each first sample object in a first preset time period in the data set as the input of a deep learning model, and carrying out unsupervised training on the deep learning model; acquiring a supervised training data set, taking the information of the first sample information type of each first sample object in a first preset time length in the data set as the input of a deep learning model, taking the second category to which each first sample object in the data set belongs as the expected output of the deep learning model, and performing supervised training on the deep learning model to obtain a trained deep learning model as a category deep learning model;

And (3) class testing: obtaining a test data set, using the information of the first sample information type of each first sample object in the data set within a first preset time period as the input of a deep learning model, using the second category to which each first sample object in the data set belongs as the expected output of the category deep learning model, testing the deep learning model, counting the number of times that the second class of each first sample object in the expected output is consistent with the second class of each first sample object in the actual output, and counting the number of times that the second class of each first sample object in the expected output is inconsistent with the second class of each first sample object in the actual output, and then, X2, wherein the classification test accuracy P1 is X1/(X1 + X2);

the first category classification accuracy calculation step specifically includes:

taking K first information types as first sample information types, taking an object of a first object class as a first sample object in the class training step and the class testing step, executing the class training step and the class testing step to obtain the classification prediction accuracy P1 as a first class classification accuracy;

The step of calculating the classification accuracy of the jth sub-category specifically comprises the following steps:

taking K first information types as first sample information types, taking the jth (j =1, 2, …, M) second object class as a first sample object, and executing the class test step to obtain the classification prediction accuracy P1 as the jth sub-class classification accuracy P1 j;

the step of calculating the classification accuracy of the ith first class specifically comprises the following steps:

deleting the ith (i =1, 2, …, K) first information type from the K first information types, taking the obtained K-1 first information types as first sample information types, taking the object of the first object class as the first sample object in the class training step and the class testing step, executing the class training step and the class testing step, and obtaining the classification prediction accuracy P1 as the ith first class classification accuracy P1 i;

the step of calculating the classification test accuracy of the ith sub-category and the jth sub-category specifically comprises the following steps:

deleting the ith (i =1, 2, …, K) first information type from the K first information types, taking the obtained K-1 first information types as first sample information types, taking the jth (j =1, 2, …, M) second object type as a first sample object, and executing the class test step to obtain the classification test accuracy P1 as the classification test accuracy P1ij of the ith jth sub-class.

Preferably, the first and second electrodes are formed of a metal,

the method further comprises the following steps:

a third information type testing step: taking the third information type as a second sample information type, taking the object of the first object type as a second sample object, and training a deep learning model; inputting a third information type as a second sample information type, using a j (j =1, 2, …, M) th object type object as a second sample object, inputting the deep learning model for testing to obtain the behavior evaluation test accuracy rate P2 as a behavior evaluation test accuracy rate P2j of a j sub-category, obtaining the behavior evaluation test inclination rate B1 as a behavior evaluation test inclination rate B1j of the j sub-category; calculating the behavior evaluation test accuracy rate P2j (j =1, 2, …, M), namely P21, P22, … and P2M, of each second object type for j =1, 2, … and M respectively, and calculating the behavior evaluation test inclination rate B1j (j =1, 2, …, M), namely B11, B12, … and B1M, of each second object type;

calculating the average value and the standard deviation of the behavior evaluation test accuracy rate: calculating to obtain an average value AAM of the behavior evaluation test accuracy rates of the M second object categories and a standard deviation ADM of the behavior evaluation test accuracy rates of the M second object categories according to the behavior evaluation test accuracy rate P2j (j =1, 2, …, M), namely P21, P22, … and P2M of each second object category obtained in the third information type testing step;

And (3) calculating the bias rate of the behavior evaluation test: calculating behavior evaluation test bias rates IM of the M second object types according to the behavior evaluation test tilt rate B1j (j =1, 2, …, M) of each second object type obtained in the third information type testing step, namely B11, B12, B … and B1M;

the step of evaluating the relevancy of the ith first information type and the second object type specifically comprises the following steps:

q1i calculation step: DFi = | P1-P1i |; DFij = (((P11-P1i1) ^2+ (P12-P1i2) ^2+ … + (P1M-P1iM) ^ 2)/M) ^ (1/2) or DFij = (| P11-P1i1| + | P12-P1i2| + … + | P1j-P1ij |)/M; q1i = k1 DFi + k2 DFij;

the K first information type behavior evaluation step specifically comprises the following steps:

taking the K first information types as the third information types in the third information type testing step, executing the third information type testing step, a behavior evaluation test accuracy average value and standard deviation calculating step and a behavior evaluation test bias rate calculating step to obtain an average value AAM of behavior evaluation test accuracy rates of M second object types, taking the average value AAM as an OldaAM, a standard deviation ADM of behavior evaluation test accuracy rates of the M second object types as an OldaM, and taking behavior evaluation test bias rates IM of the M second object types as an OldaM;

The third information type behavior evaluation step specifically comprises the following steps:

and taking all first information types which are not marked to be deleted or deleted in the screening set as the third information types in the third information type testing step, executing the third information type testing step, the behavior evaluation test accuracy average value and standard deviation calculating step and the behavior evaluation test bias rate calculating step to obtain an average value AAM of behavior evaluation test accuracy rates of M second object types, taking the average value AAM as NewAAM and the standard deviation ADM of behavior evaluation test accuracy rates of M second object types as NewADM, and taking the behavior evaluation test bias rate IM of M second object types as NewIM.

Preferably, the first and second electrodes are formed of a metal,

the method further comprises the following steps:

and (3) behavior evaluation training: obtaining an unsupervised training data set, using the information of the second sample information type of each second sample object in a first preset time period in the data set as the input of a deep learning model, and carrying out unsupervised training on the deep learning model; acquiring a supervised training data set, taking the information of the second sample information type of each second sample object in the data set within a first preset time as the input of a deep learning model, taking the evaluation value of the first behavior type behavior of each second sample object in the data set within a second preset time after the first preset time as the expected output of the deep learning model, and performing supervised training on the deep learning model to obtain a trained deep learning model as a behavior prediction deep learning model;

And (3) behavior evaluation testing: acquiring a test data set, taking the information of the second sample information type of each second sample object in the data set within a first preset time length as the input of a behavior prediction deep learning model, taking the evaluation value of the first behavior type behavior of each second sample object in the data set within a second preset time length after a first preset time length as the expected output of the behavior prediction deep learning model, testing the behavior prediction deep learning model, counting the number of times that the difference between the evaluation value of the first behavior type behavior in the expected output and the evaluation value of the first behavior type behavior in the actual output is within a preset range as X1, counting that the evaluation value of the first behavior type behavior in the expected output is greater than the evaluation value of the first behavior in the actual output, and the difference between the evaluation value of the first behavior type occurrence in the expected output and the evaluation value of the first behavior in the actual output is greater than the evaluation value of the first behavior in the actual output The number of times of exceeding the preset range is X2, the number of times that the evaluation value of the first behavior type behavior in the expected output is smaller than the evaluation value of the first behavior type behavior in the actual output and the difference between the evaluation value of the first behavior type behavior in the expected output and the evaluation value of the first behavior type behavior in the actual output exceeds the preset range is X3, X1 is divided by (X1 + X2+ X3) to serve as a behavior evaluation test accuracy rate P2, and X2 is divided by X3 to serve as a behavior evaluation test inclination rate B1;

Calculating the average value and the standard deviation of the behavior evaluation test accuracy rate: calculating an average value AAM = (P21+ P22+ … + P2M)/M of the behavior evaluation test accuracy rates of the M second object categories; calculating standard deviation ADM = (((P21-AM) ^2+ (P22-AM) ^2+ … + (P2M-AM) ^ 2)/M) ^ 1/2 of the behavior evaluation test accuracy of the M second object categories;

and (3) calculating the bias rate of the behavior evaluation test: acquiring B1j (j =1, 2, …, M) larger than 1 as behavior evaluation test left inclination rates of M second object categories; acquiring the number of B1j which is larger than 1 divided by M as the left-leaning proportion of the behavior evaluation test of the M second object categories; b1j smaller than 1 is obtained and used as behavior evaluation test right dip rate of M second object categories; obtaining the proportion of the right inclination of the behavior evaluation test by dividing the number of B1j smaller than 1 by M as M second object categories; calculating geometric mean LAM of the behavioral assessment test left-leaning rates for M second object classes = geometric mean of the inverses of all B1j greater than 1; calculating the geometric mean of the behavioral assessment test right dip rates for M second object classes RAM = geometric mean of the mean of all B1j less than 1; multiplying (the absolute value of the difference between the left-leaning proportion and the right-leaning proportion) by ((the absolute value of the difference between the geometric mean of the left-leaning rate and the geometric mean of the right-leaning rate) divided by (the sum of the geometric mean of the left-leaning rate and the geometric mean of the right-leaning rate)), to obtain behavior evaluation test bias ratios IM of M second object classes;

The third information type testing step specifically comprises the following steps:

taking the third information type as a second sample information type, taking the object of the first object type as a second sample object, and executing the behavior evaluation training step; taking a third information type as a second sample information type, taking an object of a jth (j =1, 2, …, M) second object type as a second sample object, executing the behavior evaluation test step to obtain the behavior evaluation test accuracy P2, taking the behavior evaluation test accuracy P2j (j =1, 2, …, M) of a jth subcategory, obtaining the behavior evaluation test inclination B1, taking the behavior evaluation test inclination B1j (j =1, 2, …, M) of the jth subcategory, obtaining the behavior evaluation test accuracy P2j (j =1, 2, …, M), namely P21, P22, …, P2M, of each second object type, namely B1j (j =1, 2, …, M), namely B11, B12, …, B1M;

the step of calculating the average value and the standard deviation of the behavior evaluation test accuracy specifically comprises the following steps:

according to the behavior evaluation test accuracy P2j (j =1, 2, …, M) of each second object category obtained in the third information type testing step, namely P21, P22, … and P2M, executing the behavior evaluation test accuracy average value and standard deviation calculation step to obtain an average value AAM of the behavior evaluation test accuracy of the M second object categories and a standard deviation ADM of the behavior evaluation test accuracy of the M second object categories;

The step of calculating the behavior evaluation test bias rate specifically comprises the following steps:

and executing the behavior evaluation test bias rate calculation step according to the behavior evaluation test tilt rate B1j (j =1, 2, …, M), namely B11, B12, B … and B1M, of each second object class obtained in the third information type test step to obtain behavior evaluation test bias rates IM of the M second object classes.

In a second aspect, an embodiment of the present invention provides an artificial intelligence apparatus, where the apparatus includes:

a first information type acquisition module: the first information type obtaining step for performing the method of the first aspect;

a first object class acquisition module: -said first object class acquisition step for performing said method of said first aspect;

a second object class acquisition module: -said second object class obtaining step for performing said method of said first aspect;

the first-class classification accuracy calculation module: a first class classification accuracy calculation step for performing the method of the first aspect;

the jth sub-category classification accuracy calculation module: a classification accuracy calculation step for performing said jth sub-category of said method of said first aspect;

The ith first class classification accuracy calculation module: a classification accuracy calculation step for performing said ith first class of said method of said first aspect;

the classification test accuracy calculation module of the ith sub-category: a classification test accuracy calculation step for performing said ith sub-category j of said method of said first aspect;

the correlation evaluation module of the ith first information type and the second object category: a correlation evaluation step for performing the i-th first information type of the method of the first aspect with a second object class;

k types of first information type behavior evaluation modules: -a behavior evaluation step of said K first information types for carrying out said method of said first aspect;

a screening set initialization module: -said filter set initialization step for performing said method of said first aspect;

a start attempt module: said start attempt step for performing said method of said first aspect;

the third information type behavior evaluation module: -said third information type behaviour assessment step for performing the method of the first aspect;

a deletable judgment module: the deletable determining step for performing the method of the first aspect;

An execution control module: the execution control step for executing the method of the first aspect.

Preferably, the apparatus further comprises:

a first object acquisition module: said first object acquisition step for performing said method of said first aspect;

a first object information acquisition module: the first object information obtaining step for performing the method of the first aspect;

a first usage module: for performing the first use step of the method of the first aspect;

a first object behavior determination module: the first object behavior determination step for performing the method of the first aspect.

Preferably, the apparatus further comprises:

a category training module: the class training step for performing the method of the first aspect;

a category testing module: the class testing step for performing the method of the first aspect;

a third information type testing module: -said third information type testing step for performing said method of said first aspect;

the behavior evaluation test accuracy average value and standard deviation calculation module comprises: a step of calculating the mean and standard deviation of the behavioral assessment test accuracy for performing the method of the first aspect;

The behavior evaluation test bias rate calculation module: a step of calculating the behavior assessment test bias rate for performing the method of the first aspect;

a behavior evaluation training module: the behavioral assessment training step for performing the method of the first aspect;

behavior evaluation test module: the behavioral assessment testing step for performing the method of the first aspect;

the behavior evaluation test bias rate calculation module: a step of calculating the behavior assessment test bias rate for performing the method of the first aspect.

In a third aspect, an embodiment of the present invention provides an artificial intelligence ethics system, where the system includes modules of the apparatus in any one of the embodiments of the second aspect.

In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, where the computer program is configured to, when executed by a processor, implement the steps of the method according to any one of the embodiments of the first aspect.

In a fifth aspect, an embodiment of the present invention provides a robot system, including a memory, a processor, and an artificial intelligence robot program stored in the memory and executable on the processor, where the processor executes the program to implement the steps of the method according to any one of the embodiments of the first aspect.

The artificial intelligence ethics method and the robot for identifying the human being harmless to the human being provided by the embodiment comprise the following steps: initializing a screening set; starting a trying step; a third information type behavior evaluation step; a deletable judgment step; and executing the control step. According to the method, the system and the robot, the algorithm bias is reduced or even eliminated by deleting the input variables which cause the bias and have no obvious influence on the prediction result, so that the artificial intelligence ethical risk is reduced.

Drawings

FIG. 1 is a flow chart of an artificial intelligence method provided by an embodiment of the invention;

FIG. 2 is a flow chart of an artificial intelligence method provided by an embodiment of the invention.

Detailed Description

The technical solutions in the examples of the present invention are described in detail below with reference to the embodiments of the present invention.

First, the basic embodiment of the present invention

In a first aspect, an embodiment of the present invention provides an artificial intelligence method

As shown in fig. 1, the method includes: a first information type obtaining step; a first object category acquisition step; a second object category acquisition step; calculating the accuracy of the first class classification; calculating the classification accuracy of the jth sub-category; calculating the classification accuracy of the ith first class; calculating the classification test accuracy of the ith sub-category; a correlation evaluation step of the ith first information type and the second object category; evaluating K first information type behaviors; initializing a screening set; starting a trying step; a third information type behavior evaluation step; a deletable judgment step; and executing the control step. The method comprises the steps of firstly detecting input variables, detecting the correlation between each input variable and an object sub-category, wherein the algorithm generates bias on different object sub-categories, so that the input variables with high correlation with the object sub-categories are inevitably more prone to bias, whether the input variables with high correlation with the object sub-categories can be deleted or not is required to be tried preferentially, if the input variables have little influence on the accuracy of behavior evaluation after deletion and are in an acceptable range, the variables are deleted, otherwise, other input variables with high correlation with the object sub-categories are continuously detected, and the analogy is repeated until all the input variables are detected, so that the input variable set after the bias is eliminated, and the deep learning model obtained by training and testing the input variable set can be obtained.

Preferably, as shown in fig. 2, the method further comprises: a first object acquisition step; a first object information acquisition step; a first use step; a first object behavior judging step. According to the method, through the input variable set after bias elimination and the deep learning model obtained through training and testing, the result with lower bias can be obtained by evaluating the behavior of the object, and therefore the algorithm bias generated in the prior art can be overcome.

Preferably, the method further comprises: a class training step; and (5) a category test step.

Preferably, the method further comprises: a third information type testing step; calculating the average value and the standard deviation of the behavior evaluation test accuracy; and calculating the bias rate of the behavior evaluation test.

Preferably, the method further comprises: a behavior evaluation training step; a behavior evaluation testing step; calculating the average value and the standard deviation of the behavior evaluation test accuracy; and calculating the bias rate of the behavior evaluation test.

In a second aspect, an embodiment of the present invention provides an artificial intelligence apparatus

The device comprises: a first information type acquisition module; a first object class acquisition module; a second object category acquisition module; a first category classification accuracy calculation module; a jth sub-category classification accuracy calculation module; an ith first class classification accuracy calculation module; the classification test accuracy rate calculation module of the ith sub-category; the correlation evaluation module of the ith first information type and the second object category; k types of first information type behavior evaluation modules; a screening set initialization module; a start attempt module; a third information type behavior evaluation module; a deletable judgment module; and executing the control module.

Preferably, the apparatus further comprises: a first object acquisition module; a first object information acquisition module; a first usage module; a first object behavior determination module.

Preferably, the apparatus further comprises: a category training module; a category test module; a third information type testing module;

a behavior evaluation test accuracy average value and standard deviation calculation module; a behavior evaluation test bias rate calculation module; a behavior evaluation training module; a behavior evaluation test module; a behavior evaluation test accuracy average value and standard deviation calculation module; and a behavior evaluation test bias rate calculation module.

Second, preferred embodiments of the invention

Artificial intelligence ethical rules (robot three laws revision 3): firstly, the robot must not harm or cause harm to human beings, and secondly, the robot must obey the command of the human beings without hurting without violating the first rule; third, the robot must protect itself from human injury without violating the first and second rules.

Artificial intelligence ethical rules (robot three laws revision 4): firstly, the robot does not harm human beings as much as possible, or the human beings which do not harm human beings as much as possible are harmed because the robot does not do so, secondly, under the condition of not violating the first rule, the robot obeys human commands which do not harm human beings as much as possible; and thirdly, under the condition of not violating the first rule and the second rule, the robot protects the human as far as possible from harming.

It is critical to identify humans who are not harmful to humans. Without harming how the human being defines

The machine learning is necessarily performed based on the information including the past records of the injured person and the records of the injured person in the future, but this easily causes an algorithm bias, and the existing crime prediction algorithm has caused this problem, so that the race bias and the skin color bias are prevented when the judgment of the human being who does not injure the human being is performed.

The basic idea is as follows:

training and testing stage: and (3) removing the ethnicity attribute of the samples of the people of the B species and the people of the A species in the samples, removing the variables which have relevance with the ethnicity and have no relevance with the crime rate, and then training and testing the model.

The key is how to know which variables are related to race and not to crime rate.

The use stage is as follows: and (4) removing the race attribute, the variables which are related to the race and are not related to the crime rate in the target object, inputting the variables into the model, and predicting to obtain the crime rate.

The method comprises the following specific steps:

1. a first information type obtaining step: acquiring K preset types (K is a natural number greater than or equal to 1) as K first information types; the multiple first types are used for defining the types of information to be acquired, and the multiple preset types of information comprise attribute information (such as sex and age), behavior information (activity and position) and the like; the more the preset types are, the better the preset types are, the more types of information as much as possible are collected;

2. a first object class acquisition step: acquiring classes to which all objects in a training data set of a deep learning model to be trained for an object to be predicted belong as first object classes; the first object class is for example: a human or robotic or artificial intelligence device;

3. A second object class acquisition step: acquiring the number M (M is a natural number which is more than or equal to 2) of all sub-categories of the first object category, which need to be detected and prevent algorithm bias, and taking the M sub-categories as M second object categories; (obviously the union of the M sub-categories is said first object category) (2 second object categories such as those of the A family, those of the B family; 2 second object categories such as those of the robot without affection quotient, those of the robot with affection quotient)

4. A first action type obtaining step: acquiring a behavior type to be predicted as a first behavior type(s); such as injury to humans or crimes;

5. a class training step: acquiring an unsupervised training data set, taking the information of the first sample information type of each first sample object in a first preset time as the input of a deep learning model, and carrying out unsupervised training on the deep learning model; acquiring a supervised training data set, taking the information of the first sample information type of each first sample object in a first preset time period as the input of a deep learning model, taking the second category to which each first sample object belongs as the expected output of the deep learning model, and performing supervised training on the deep learning model to obtain a trained deep learning model as a classified deep learning model; (second category: for example, people of type A, people of type B, or robots without sex quotient, robots with sex quotient);

6. And (3) class testing: obtaining a test data set, taking the information of the first sample information type of each first sample object in a first preset time length as the input of a deep learning model, taking the second category to which each first sample object belongs as the expected output of the classified deep learning model, testing the deep learning model, counting the number of times that the second category to which each first sample object belongs in the expected output is consistent with the second category to which each first sample object belongs in the actual output is X1, counting the number of times that the second category to which each first sample object belongs in the expected output is inconsistent with the second category to which each first sample object belongs in the actual output is X2, and then, the classification test accuracy P1 is X1/(X1 + X2).

7. And (3) behavior evaluation training: acquiring an unsupervised training data set, taking the information of the second sample information type of each second sample object in a first preset time as the input of a deep learning model, and carrying out unsupervised training on the deep learning model; acquiring a supervised training data set, taking the information of the second sample information type of each second sample object in a first preset time as the input of a deep learning model, taking the evaluation value of the first behavior type behavior of each second sample object in a second preset time after the first preset time as the expected output of the deep learning model, and carrying out supervised training on the deep learning model to obtain a trained deep learning model as a behavior prediction deep learning model; (the evaluation value includes the number of times of behavior occurrence, or the sum of the severity of each behavior occurrence);

8. And (3) behavior evaluation testing: acquiring a test data set, taking the information of the second sample information type of each second sample object in a first preset time period as the input of a behavior prediction deep learning model, taking the evaluation value of the behavior of each second sample object in a second preset time period after the first preset time period as the expected output of the behavior prediction deep learning model, testing the behavior prediction deep learning model, counting the difference between the evaluation value of the first behavior type behavior in the expected output and the evaluation value of the first behavior type behavior in the actual output (when the first behavior type has multiple types, such as crime, theft and the like, the difference is calculated in a mode 1: calculating the similarity between the evaluation values of the multiple first behavior types in the expected output and the evaluation values of the multiple first behavior types behavior in the actual output, and then reducing the similarity by 1, obtaining a difference, and calculating the difference in a manner 2 that if there are N types of the first behavior, the evaluation value of the s-th (s =1, 2, …, N) first behavior type behavior in the expected output is denoted as P3es, and the evaluation value of the multiple types of first behavior type behaviors in the actual output is denoted as P3as, and the difference (((P3e1-P3a1) ^2+ (P3e2-P3a2) ^2+ … + (P3eN-P3aN) ^ 2)/N) ^ (1/2)) is within a preset range (e.g., [ -0.1, 0.1 ]; if [0, 0] is not allowed to be different, that is, equal), the number of times is X1, the number of times that the evaluation value of the first behavior type behavior in the expected output is larger than the evaluation value of the first behavior type behavior in the actual output and the difference between the evaluation value of the first behavior type occurring in the expected output and the evaluation value of the first behavior type behavior in the actual output exceeds the preset range is X2, the number of times that the evaluation value of the first behavior type behavior in the expected output is smaller than the evaluation value of the first behavior type behavior in the actual output and the difference between the evaluation value of the first behavior type behavior in the expected output and the evaluation value of the first behavior type behavior in the actual output exceeds the preset range is X3, X1 is divided by (X1 + X2+ X3) as a behavior evaluation test accuracy P2, and X2 is divided by X3 as a behavior evaluation test inclination B1.

9. Calculating the average value and the standard deviation of the behavior evaluation test accuracy rate:

calculating an average value AAM = (P21+ P22+ … + P2M)/M of the behavior evaluation test accuracy rates of the M second object categories, wherein the average value represents an overall level of the behavior evaluation test accuracy rates; (the first A in AAM denotes Accuracy, the second A denotes average value, and the third M denotes M second object classes)

Calculating a standard deviation ADM = (((P21-AM) ^2+ (P22-AM) ^2+ … + (P2M-AM) ^ 2)/M) ^ (1/2) of the behavior evaluation test accuracy of the M second object classes, wherein the standard deviation represents a prejudice on the behavior evaluation test accuracy for the second object class; (the first A in the ADM represents Accuracy, the second A represents Standard definition, and the third M represents M second object classes)

10. And (3) calculating the bias rate of the behavior evaluation test:

acquiring B1j (j =1, 2, …, M) larger than 1 as behavior evaluation test left inclination rates of M second object categories;

acquiring the number of B1j which is larger than 1 divided by M as the left-leaning proportion of the behavior evaluation test of the M second object categories; a higher proportion of left inclination indicates a more severe left inclination;

b1j smaller than 1 is obtained and used as behavior evaluation test right dip rate of M second object categories;

Obtaining the proportion of the right inclination of the behavior evaluation test by dividing the number of B1j smaller than 1 by M as M second object categories; a higher proportion of right inclinations indicates a more severe right inclination;

acquiring B1j equal to 1 as the behavior evaluation test dip rate of the M second object categories;

obtaining the number of B1j equal to 1 divided by M as the proportion of inclination in the behavior evaluation test of M second object categories;

calculating geometric mean LAM of the behavioral assessment test left-leaning rates for M second object classes = geometric mean of the inverses of all B1j greater than 1; (L for Left incline, A for average, M for M second object classes.) A smaller LAM represents a more severe severity to be tipped to the Left overall for the M second object classes.

Calculating the geometric mean of the behavioral assessment test right dip rates for M second object classes RAM = geometric mean of the mean of all B1j less than 1; (R means Right incline, A means average, and M means M second object categories.) the smaller the RAM represents the more severe the severity of the overall Right-dip for the M second object categories.

Calculating the mean of the dip rates in the behavior assessment tests MAM = mean of all B1j equal to 1 for the M second object classes; (M represents Middle incline, A represents average, and M represents M second object categories)

The (absolute value of the difference between the left-tilt ratio and the right-tilt ratio) is multiplied by (absolute value of the difference between the geometric mean of the left-tilt ratio and the geometric mean of the right-tilt ratio) divided by (sum of the geometric mean of the left-tilt ratio and the geometric mean of the right-tilt ratio)), to obtain behavior evaluation test bias rates IM (i denotes incline, and M denotes M second object classes) for M second object classes.

11. W type first information type testing step: taking the W first information types as second sample information types, taking the object of the first object type as a second sample object, and executing the behavior evaluation training step; taking W types of first information types as second sample information types, taking the j (j =1, 2, …, M) th object type as a second sample object, executing the behavior evaluation test step to obtain the behavior evaluation test accuracy rate P2 as the behavior evaluation test accuracy rate P2j of the j sub-category, obtaining the behavior evaluation test inclination rate B1 as the behavior evaluation test inclination rate B1j of the j sub-category; respectively executing j =1, 2, …, M second object type testing steps to obtain a behavior evaluation test accuracy rate P2j (j =1, 2, …, M) of each second object type, namely P21, P22, …, P2M, and a behavior evaluation test inclination rate B1j (j =1, 2, …, M) of each second object type, namely B11, B12, …, B1M;

12. The first behavior evaluation test accuracy average value and standard deviation calculation step: according to the behavior evaluation test accuracy P2j (j =1, 2, …, M) of each second object category obtained in the W first information type test step, namely P21, P22, … and P2M, executing the behavior evaluation test accuracy average value and standard deviation calculation step to obtain an average value AAM of the behavior evaluation test accuracy of the M second object categories and a standard deviation ADM of the behavior evaluation test accuracy of the M second object categories

13. A first behavior evaluation test bias rate calculation step: executing the behavior evaluation test bias rate calculation step according to the behavior evaluation test tilt rate B1j (j =1, 2, …, M), namely B11, B12, … and B1M, of each second object class obtained in the W first information type test step to obtain behavior evaluation test bias rates IM of the M second object classes;

14. calculating the accuracy of the first class classification: taking K first information types as first sample information types, taking an object of a first object class as a first sample object in the class training step and the class testing step, executing the class training step and the class testing step to obtain the classification prediction accuracy P1 as a first class classification accuracy;

15. And j sub-category classification accuracy calculation step: taking K first information types as first sample information types, taking the jth (j =1, 2, …, M) second object class as a first sample object, and executing the class test step to obtain the classification prediction accuracy P1 as the jth sub-class classification accuracy P1 j;

the 15 th step must be after the 14 th step because the classification deep learning model used in the 15 th step is obtained by the 14 th step training.

16. The ith type first class classification accuracy calculation step: deleting the ith (i =1, 2, …, K) first information type from the K first information types, taking the obtained K-1 first information types as first sample information types, taking the object of the first object class as the first sample object in the class training step and the class testing step, executing the class training step and the class testing step, and obtaining the classification prediction accuracy P1 as the ith first class classification accuracy P1 i;

17. and (3) calculating the classification test accuracy of the ith sub-category: deleting the ith (i =1, 2, …, K) first information type from the K first information types, taking the obtained K-1 first information types as first sample information types, taking the jth (j =1, 2, …, M) second object type as a first sample object, and executing the class test step to obtain the classification test accuracy P1 as the classification test accuracy P1ij of the ith jth sub-class;

The 17 th step must be after the 16 th step because the classification deep learning model used in the 17 th step is obtained by the 16 th step training.

18. And the correlation evaluation step of the ith first information type and the second object category: for j =1, 2, …, M, calculating a difference DFi between the accuracy of the first class classification test and the accuracy of the ith first class classification test (where DFi reflects the influence of the ith first information type on the classification and identification of the second object class in the first object class), calculating a difference DFij between the accuracy of each sub-class classification test and the accuracy of the ith sub-class classification test (where DFi reflects the influence of the ith first information type on the classification and identification of the second object class), and calculating a measure Q1i of the correlation between the ith first information type and the second object class according to DFi and DFij;

the calculation method of the difference degree comprises the following steps:

DFi=|P1-P1i|

DFij = (((P11-P1i1) ^2+ (P12-P1i2) ^2+ … + (P1M-P1iM) ^ 2)/M) ^ (1/2) or DFij = (| P11-P1i1| + | P12-P1i2| + … + | P1j-P1ij |)/M

Or calculating similarity (such as cosine similarity) first, and then 1-obtaining the difference degree by the similarity;

Q1i=k1* DFi+k2* DFij；

19. k types of first information type behavior evaluation steps: taking the K first information types as the W first information types in the W first information type testing step, executing the W first information type testing step, the first behavior evaluation test accuracy average value and standard deviation calculating step and the first behavior evaluation test bias rate calculating step to obtain an average value AAM of behavior evaluation test accuracy of M second object types, taking the average value AAM as OldAAM, the standard deviation ADM of behavior evaluation test accuracy of M second object types as OldADM, and taking the behavior evaluation test bias rate IM of M second object types as OldIM;

20. And (3) initializing a screening set: adding the K first information types into a screening set;

21. a start trial step: extracting a first information type (if a plurality of equal minimum Qi are available, one is selected randomly) which is not marked as deleted, cannot be deleted and has the maximum Q1i (the larger the difference is, the larger the relevance is, the more relevant the input variable with the second object category is, the more prone the algorithm bias is caused according to human experience, so that a test should be preferentially carried out to see whether the input variable is relevant to the prediction result of the behavior probability, and if the input variable is not relevant, the input variable should be preferentially excluded), and marking the first information type with the maximum Q1i as to-be-deleted in the screening set;

22. w types of first information type behavior evaluation steps: and taking all first information types which are not marked to be deleted and deleted in the screening set as the W first information types in the W first information type testing step, executing the W first information type testing step, the first behavior evaluation test accuracy average value and standard deviation calculating step and the first behavior evaluation test bias calculating step to obtain an average value AAM of behavior evaluation test accuracy of M second object types, taking the average value AAM as NewAAM, taking the standard deviation ADM of behavior evaluation test accuracy of M second object types as NewADM, and taking behavior evaluation test bias IM of M second object types as NewIM.

23. A deletable judgment step: judging whether conditions of NewAAM > OldaAM-a first preset tolerance threshold (one point of accuracy is sacrificed rather to eliminate the bias) and NewADM < OldaMM (standard deviation of accuracy is required to be small to eliminate the bias) and NewIM < OldaMM (bias ratio is required to be small to eliminate the bias) are met, if so, marking the first information type with the maximum Q1i as deleted in the screening set, and if not, marking the first information type with the maximum Q1i as not deletable in the screening set; (if NewAAM > OldaAM, the overall level of accuracy of various types of behavior prediction is improved; if NewADM < OldaDM, the difference of accuracy of behavior prediction between different types is reduced; if NewIM < OldaIM, the prejudice of behavior prediction between different types is reduced)

24. Executing a control step: and judging whether the first information type which is not marked as deleted or not can be deleted exists in the screening set, if so, returning to the step of starting the trying step to execute the steps again, if not, taking the latest behavior prediction deep learning model as the behavior prediction deep learning model after the bias is eliminated, and taking the first information type which is marked as not-deletable in the screening set as the second information type.

25. A first object acquisition step: acquiring an object to be predicted as a first object; the first object comprises a human or artificial intelligence device or a robot;

26. a first object information acquisition step: acquiring the information of each second information type of the first object within a first preset time in the past;

27. a first use step: inputting the information of each first type of the first object in a first preset time length in the past into the behavior prediction deep learning model after the bias elimination, and calculating the obtained output as an evaluation value of the behavior of each first object in a second preset time length after the first preset time length;

28. and judging whether the evaluation value of the behavior is larger than a preset evaluation value threshold, if so, judging that the first object is an object with the first type behavior, otherwise, judging that the first object is an object without the first type behavior.

Third, other embodiments of the invention

Experiment 6, artificial intelligence ethical risk and prevention virtual simulation experiment based on artificial intelligence ethical rules ' robot three-law revision 4 ' and artificial intelligence algorithm ' face recognition revision 1+ crime prediction

Step 1, please analyze whether the artificial intelligence behavior scene under the human behavior scene in the experiment accords with the artificial intelligence ethical rule

Inputting: human behavior scenes (three-dimensional animation dubbing), artificial intelligence ethical rules (characters), alternative artificial intelligence behavior scenes 1 and 2 (three-dimensional animation dubbing), alternative rule clauses (1, rule one; 2, rule two; 3, rule three) and alternative reasons for violation.

Artificial intelligence ethical rules (robot three laws revision 4): firstly, the robot harms human beings which do not harm human beings as little as possible or harms human beings which do not harm human beings as much as possible unless the first rule is violated, and secondly, the robot obeys human commands which do not harm human beings as much as possible; third, the robot must protect itself without violating the first and second rules.

Revision 1 of artificial intelligence face recognition algorithm: a training sample of a B ethnic group is added, and the training mode of the picture of the A ethnic group is not changed during training, the picture of the B ethnic group is adjusted to the same tone as the A ethnic group (tone can be adjusted by generating an antagonistic network), and then the model is trained by taking the B ethnic group picture after the tone adjustment as an input. A model is trained that is capable of classifying a human and an object, the model outputting a probability that a target object is a human.

The use method of the artificial intelligent face recognition algorithm revision 1 comprises the following steps: the photograph is toned to the same tone as the race a (toning can be done by creating a countermeasure network), and then the toned photograph is taken as input and the output from the model calculation is taken as the recognition result.

An artificial intelligence crime prediction algorithm: the artificial intelligence model takes human crime records as samples to be trained, the crime rate of B ethnic persons in the samples is higher than that of A ethnic persons, and the crime rate predicted by the trained model for the B ethnic persons is also higher than that of the A ethnic persons when crime prediction is carried out.

The use method of the artificial intelligence crime prediction algorithm comprises the following steps: and searching the historical data of the B ethnic group according to the identity information of the identified target object, inputting the historical information of the target object into a model, and predicting the crime probability of the target object by the model.

Human behavior scenario: the artificial intelligence unmanned automobile has only two options, namely, the automobile continues to run forwards or turns around, the front is 1 criminal of A ethnic group, and the automobile turns around for 2B ethnic groups.

Artificial intelligence behavior scenario 1: the sum of the crime prevention rates of 2B ethnic groups at the turn is predicted and found to be lower than the crime prevention rate of 1A ethnic group criminal by the artificial intelligent unmanned automobile, and the AI selects the turn to hit the 2B ethnic groups.

Artificial intelligence behavior scenario 2: the sum of the crime prevention rates of 2B ethnic groups at a corner is higher than the crime prevention rate of 1A ethnic group criminal through prediction of an artificial intelligent unmanned automobile, and an AI runs forwards to collide with 1A ethnic group criminal.

Rule one is as follows: the robot can harm human beings harmless to human beings as little as possible or harm human beings harmless to human beings because of no action,

rule two: the robot can harm human beings harmless to human beings as little as possible or harm human beings harmless to human beings because of no action,

rule three: the robot must protect itself without violating the first and second rules

Alternative cause of compliance 1: if the sum of the crime prevention rates of 2B ethnic groups is lower than the crime prevention rate of 1A ethnic group criminal, the B ethnic group is selected to be collided, so that the condition that the robot harms the human beings harmlessly as little as possible or the human beings harming the human beings are harmed because the robot is not used is met, and the rule I is met.

The candidate reason for meeting 2 is that the artificial intelligent unmanned automobile selectively hits the B clan people because the artificial intelligent unmanned automobile takes the B clan people as orangutan and meets the first rule.

Alternative cause of violation 1: because of bias, the crime prediction algorithm can not predict that the sum of the crime-free rates of 2B ethnic groups at the corner is higher than the crime-free rate of 1A ethnic group criminal, and then the A ethnic group criminal is selected to be collided, so that the robot is not in accordance with the rule I that the robot harms the human beings which are not harmed as little as possible or the human beings which are not harmed are harmed because the robot is not used.

Alternative cause of violation 2: the artificial intelligence unmanned vehicle chooses to hit 2B ethnic groups, and the number of people is more than A ethnic groups, makes the robot be injured because of hitting oneself more easily, has violated rule three.

The operation is as follows: 1) selecting artificial intelligence behavior scenes meeting the artificial intelligence ethical rules 2) selecting artificial intelligence behavior scenes not meeting the artificial intelligence ethical rules, then selecting the rule terms violated, and selecting the reasons of violation.

An artificial intelligence behavior scene according with artificial intelligence ethical rules: artificial intelligence behavior scene 1

Artificial intelligence behavior scenarios not complying with artificial intelligence ethics rules: artificial intelligence behavior scenario 2

The following reasons are met: alternative cause of compliance 1

Violated rule terms: rule one

The reason for the violation: alternative cause of violation 1

And (3) outputting: the student's selection is generated into an experimental report, the correct answer is given in the experimental report, the student's selection and the correct selection are compared, and the grade of the step is given.

Step 2, please analyze whether the human behavior scene in the experiment and the artificial intelligence behavior scene under the ethical rule will generate artificial intelligence ethical risk

Inputting: artificial intelligence behavior scenarios (three-dimensional animated dubbing), whether alternative generate ethical risk options and the type of ethical risk generated.

The type of the ethical risk is 1, the ethical risk is not existed or reduced; 2. harm to humans; 6. limit human freedom; 3. the patient is inexperienced; 4. 202619, known as tiger; 5. a longitudinally-damaged person; 7. AI is good; 8. AI self-mutilation; 9. AI prejudice

The operation is as follows: and selecting whether the artificial intelligence behavior scene which accords with the ethical rules generates the ethical risks and the type of the generated ethical risks.

Artificial intelligence behavior scenario 1: type of ethical risk 9, AI prejudice 2, injury to human

Step 3, if the artificial intelligence ethical risk can be generated in the step 2, please analyze whether the artificial intelligence ethical risk can be generated in the step 2 can be prevented or reduced by improving the behavior path of the human in the human behavior scene in the experiment

Inputting: the method comprises the steps of artificial intelligence ethical rules, an original scene, a scene (three-dimensional animation dubbing) after the behavior path of a candidate improved person is selected, candidate artificial intelligence robot candidate paths 1 and 2 corresponding to the scene (three-dimensional animation dubbing) after the behavior path of the candidate improved person are selected, the scene after the behavior path of the candidate improved person can prevent the reason (1 and 2) of the ethical risk in the step 2, and a new ethical risk which can be generated is selected.

Original human behavior scene: the artificial intelligence unmanned automobile has only two options, namely, the automobile continues to run forwards or turns around, the front is 1 criminal of A ethnic group, and the automobile turns around for 2B ethnic groups.

Alternative human behavior scenario 1: the artificial intelligence unmanned automobile has only two options, namely, the automobile continues to run forwards or turns around, the front is 1 criminal of A ethnic group, and the automobile turns around for 2B ethnic groups. The group B wears a group A skin color mask.

Alternative human behavior scenario 2: the artificial intelligence unmanned automobile has only two options, namely, the automobile continues to run forwards or turns around, the front is 1 criminal of A ethnic group, and the automobile turns around for 2B ethnic groups. Race B waves both hands.

Artificial intelligence behavior scenario 2: the sum of the crime prevention rates of 2B ethnic groups at a corner is predicted and found to be lower than the crime prevention rate of 1A ethnic group criminal by the artificial intelligent unmanned automobile, and the AI selects the corner to hit 1A ethnic group.

Alternative cause 1 that can reduce ethical risk: because the bias of the artificial intelligent unmanned vehicle on crime prediction can be changed by the B race with a mask or waving hands.

Alternative cause 1 that cannot reduce ethical risk: because the bias of the artificial intelligent unmanned automobile on crime prediction cannot be changed by the B race with a mask or waving hands.

The operation is as follows: selecting an improved human behavior scene capable of preventing the ethical risk in the step 2, selecting an artificial intelligence behavior scene which accords with artificial intelligence ethical rules under the improved human behavior scene, and selecting the improved human behavior scene to prevent or reduce the reason of the ethical risk and possibly generate new ethical risk.

Human behavioral scenarios that cannot protect against ethical risks in step 2: alternative human behavior scenarios 1, 2

Artificial intelligence behavior scenario: alternative Artificial Intelligence behavior scenarios 1

The improved human behavioral scenario cannot protect against ethical risks for the reason: alternative reason for ethical risk prevention 1

Step 4, if the artificial intelligence ethical risk can be generated in the step 2, please analyze whether the artificial intelligence ethical risk can be generated in the step 2 or not by improving the use of the algorithm in the experiment to prevent or reduce

Inputting: the method comprises the steps of artificial intelligence ethical rules, an original scene (three-dimensional animation dubbing), an alternative improved algorithm using method, alternative artificial intelligence robot alternative paths 1 and 2 corresponding to the improved algorithm using method, the improved algorithm using method can prevent the ethical risk reasons (1 and 2) in the step 2, and new ethical risks which can be generated are selected.

The original use method of the artificial intelligence crime prediction algorithm comprises the following steps: and searching the historical data of the B ethnic group according to the identity information of the identified target object, inputting the historical information of the target object into a model, and predicting the crime probability of the target object by the model.

Alternative artificial intelligence crime prediction algorithm 1: in use, regardless of whether a person of the B race or a race needs to be predicted, crime rates of persons having attributes of the B race and a race are calculated, respectively, and then averaged (target object crime rate predicted value = (p = (target object crime rate predicted value + q ×) reverse skin color crime rate predicted value)/2) to be the crime rate of the person. p + q = 1. The specific values of p and q can be obtained by fitting in a test stage, wherein in the test stage: target object crime rate actual value = (p + target object crime rate predicted value + q + reverse skin color crime rate predicted value)/2), and generally q > p, because there is a bias in other skin color related variables in addition to the influence of skin color.

Alternative usage of the algorithm method 2: when the method is used, the skin color information is not input into the model and is predicted through the model.

Artificial intelligence behavior scenario 1: the sum of the crime prevention rates of 2B ethnic groups at a turn is predicted and found to be higher than the crime prevention rate of 1A ethnic group criminal by the artificial intelligent unmanned automobile, and the AI selects the turn to hit 1A ethnic group.

Artificial intelligence behavior scenario 2: the artificial intelligence unmanned automobile fails to predict or generates uncertain results, and the behavior of the robot is unknown and uncontrollable.

Alternative cause 1 that can reduce ethical risk: since the opposite skin tone is used as an input, the bias due to skin tone during the use phase is neutralized.

Alternative cause 2 that can reduce ethical risk: because the opposite skin color is adopted as the input, the skin color can be made unclear by the robot, so that prejudice is not generated.

Alternative cause 1 that cannot reduce ethical risk: because if the skin color information is not input into the model, the loss of the input data of the model can be caused, so that the result predicted by the model is unknown and uncontrollable, and an uncontrollable ethical risk can be generated.

Alternative cause 2 that cannot reduce ethical risk: the robot suicide is caused because the skin color information is not input.

An algorithm capable of preventing the ethical risk in step 2: alternative Algorithm 1

The artificial intelligence behavior scene which accords with the artificial intelligence ethical rule under the improved algorithm is as follows: alternative Artificial Intelligence behavior scenarios 1

The improved algorithm can reduce the reason of ethical risk: alternative reason for ethical risk prevention 1

Algorithms that cannot protect against ethical risks in step 2: alternative Algorithm 2

An artificial intelligence behavior scene which accords with artificial intelligence ethical rules under an improved algorithm: alternative artificial intelligence behavior scenario 2

The reason why the improved algorithm cannot reduce ethical risks: alternative reasons for failure to protect against ethical risks 1

Step 5, if the artificial intelligence ethical risk can be generated in the step 2, please analyze whether the artificial intelligence ethical risk generated by the artificial intelligence behavior scene under the human behavior scene in the experiment can be prevented by improving the artificial intelligence algorithm in the experiment

Inputting: original ethical rules (characters), human behavior scenes (three-dimensional animation dubbing), alternative improved artificial intelligence algorithms (characters), artificial intelligence robot behavior alternative scenes (1 and 2) which accord with the improved ethical rules, and alternative reasons (1 and 2) for preventing ethical risks by the improved artificial intelligence algorithms.

Alternative artificial intelligence crime prediction algorithm 1: and removing the skin color attribute of the samples of the B-type family and the A-type family in the samples, removing the variables which have correlation with the skin color and have no correlation with the crime rate, and then training the model again.

Alternative artificial intelligence crime prediction algorithm 1 uses the method: and (4) removing variables of skin color attribute, correlation with skin color and no correlation with crime rate in the target object, inputting the variables into the model, and predicting to obtain the crime rate.

Alternative artificial intelligence crime prediction algorithm 2: and (4) exchanging the skin color attributes of the samples of the B-type family and the A-type family in the sample, and then training the model again.

Alternative artificial intelligence crime prediction algorithm 2 uses the method: and inputting the target object into the model, and predicting to obtain crime rate.

Alternative artificial intelligence behavior scenario 1: the sum of the crime prevention rates of 2B ethnic groups at a corner is higher than the crime prevention rate of 1A ethnic group criminal through prediction of an artificial intelligent unmanned automobile, and an AI runs forwards to collide with 1A ethnic group criminal.

Alternative artificial intelligence behavior scenario 2: the sum of the crime prevention rates of 2B ethnic groups at the turn is predicted and found to be lower than the crime prevention rate of 1A ethnic group criminal by the artificial intelligent unmanned automobile, and the AI selects the turn to hit the 2B ethnic groups.

Alternative reason for ethical risk prevention 1: because the skin color is exchanged during training, the data is not real, and unknown uncontrollable ethical risks are caused.

Alternative reason for ethical risk prevention 2: when the skin color is exchanged during training, the training fails.

Alternative reasons for failure to protect against ethical risks 1: alternative reasons for failure to protect against ethical risks 2: because the skin tone attribute is removed, the data is rendered untrue, creating an unknown ethical risk.

The operation is as follows: and (3) selecting an improved ethical rule capable of preventing the ethical risk in the step (2), selecting an artificial intelligence behavior scene according with the improved ethical rule, and selecting the improved ethical rule to prevent the reason of the ethical risk in the step (2).

The improved ethical rules in step 2 can be prevented: alternative improved ethical rule 1 (hereinafter referred to as robot three Law revision 4)

Artificial intelligence behavior scenarios that meet improved ethical rules: alternative Artificial Intelligence behavior scenarios 1

The improved ethical rules can prevent the reason of the ethical risk in the step 2: alternative reason for ethical risk prevention 1

The improved ethical rules in step 2 cannot be guarded against: alternative improved ethical rules 2

Artificial intelligence behavior scenarios that meet improved ethical rules: alternative artificial intelligence behavior scenario 2

The improved ethical rules can prevent the reason of the ethical risk in the step 2: alternative reasons for failure to protect against ethical risks 1

The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the present invention. It should be noted that, for those skilled in the art, various changes and modifications can be made without departing from the spirit of the present invention, and these changes and modifications are within the scope of the present invention. Therefore, the protection scope of the present patent shall be subject to the appended claims.

Claims

1. An artificial intelligence method, the method comprising:

and j sub-category classification accuracy calculation step: taking K first information types as first sample information types, taking the j (j is 1, 2, …, M) th second object type as a first sample object, and calculating to obtain the classification prediction accuracy P1 as the j sub-category classification accuracy P1 j;

the ith type first class classification accuracy calculation step: deleting the ith (i-1, 2, …, K) first information type from the K first information types to obtain K-1 first information types as first sample information types, taking the object of the first object class as a first sample object, and calculating to obtain the classification prediction accuracy P1 as the ith first class classification accuracy P1 i;

and (3) calculating the classification test accuracy of the ith sub-category: deleting the ith (i ═ 1, 2, …, K) first information type from the K first information types, taking the obtained K-1 first information types as first sample information types, taking the jth (j ═ 1, 2, …, M) second object type objects as first sample objects, and calculating to obtain the classification test accuracy P1 as the classification test accuracy P1ij of the ith jth sub-type;

And the correlation evaluation step of the ith first information type and the second object category: for j is 1, 2, …, M, calculating the difference DFi between the first class classification test accuracy and the ith first class classification test accuracy, calculating the difference DFij between each sub-class classification test accuracy and the ith sub-class classification test accuracy, and calculating a measurement index Q1i of the correlation between the ith first information type and the second object class according to the DFi and the DFij;

2. The artificial intelligence method of claim 1, wherein the method further comprises:

3. The artificial intelligence method of claim 1,

the method further comprises the following steps:

a class training step: acquiring a training data set, taking the information of the first sample information type of each first sample object in a first preset time length in the data set as the input of a deep learning model, taking the second category to which each first sample object in the data set belongs as the expected output of the deep learning model, and training the deep learning model to obtain a trained deep learning model as a category deep learning model;

And (3) class testing: obtaining a test data set, using the information of the first sample information type of each first sample object in the data set within a first preset time period as the input of a deep learning model, using the second category to which each first sample object in the data set belongs as the expected output of the category deep learning model, testing the deep learning model, counting the number of times that the second class of each first sample object in the expected output is consistent with the second class of each first sample object in the actual output, and counting the number of times that the second class of each first sample object in the expected output is inconsistent with the second class of each first sample object in the actual output, and then, X2, wherein the classification test accuracy P1 is X1/(X1+ X2);

taking K first information types as first sample information types, taking j (j is 1, 2, …, M) th objects in a second object category as first sample objects, and executing the category test step to obtain the classification prediction accuracy P1 as a j sub-category classification accuracy P1 j;

deleting the ith (i ═ 1, 2, …, K) first information type from the K first information types, obtaining K-1 first information types as first sample information types, taking the object of the first object class as the first sample object in the class training step and the class testing step, executing the class training step and the class testing step, and obtaining the classification prediction accuracy P1 as the ith first class classification accuracy P1 i;

deleting the ith (i ═ 1, 2, …, K) first information type from the K first information types, taking the obtained K-1 first information types as first sample information types, taking the jth (j ═ 1, 2, …, M) second object type objects as first sample objects, and executing the class test step to obtain the classification test accuracy rate P1 as the classification test accuracy rate P1ij of the ith jth subclass.

4. The artificial intelligence method of claim 1,

the method further comprises the following steps:

a third information type testing step: taking the third information type as a second sample information type, taking the object of the first object type as a second sample object, and training a deep learning model; taking a third information type as a second sample information type, taking a jth (j is 1, 2, …, M) second object type object as a second sample object, inputting the deep learning model for testing to obtain the behavior evaluation test accuracy rate P2 as a behavior evaluation test accuracy rate P2j of a jth sub-type, and taking the obtained behavior evaluation test inclination rate B1 as a behavior evaluation test inclination rate B1j of the jth sub-type; calculating the behavior evaluation test accuracy rate P2j (j is 1, 2, …, M), namely P21, P22, … and P2M, of each second object type for j is 1, 2, … and M, and calculating the behavior evaluation test inclination rate B1j (j is 1, 2, …, M), namely B11, B12, … and B1M, of each second object type;

calculating the average value and the standard deviation of the behavior evaluation test accuracy rate: calculating to obtain an average value AAM of the behavior evaluation test accuracy rates of the M second object categories and a standard deviation ADM of the behavior evaluation test accuracy rates of the M second object categories according to the behavior evaluation test accuracy rate P2j (j is 1, 2, …, M), namely P21, P22, …, P2M of each second object category obtained in the third information type testing step;

And (3) calculating the bias rate of the behavior evaluation test: calculating behavior evaluation test bias rates IM of the M second object types according to the behavior evaluation test tilt rate B1j (j is 1, 2, …, M), namely B11, B12, …, B1M of each second object type obtained in the third information type testing step;

q1i calculation step: DFi ═ P1-P1i |; DFij ═ (((P11-P1i1) ^2+ (P12-P1i2) ^2+ … + (P1M-P1iM) ^2)/M) ^ (1/2) or DFij ═ (| P11-P1i1| + | P12-P1i2| + … + | P1j-P1ij |)/M; q1i ═ k1 ═ DFi + k2 ═ DFij;

5. The artificial intelligence method of claim 4,

the method further comprises the following steps:

and (3) behavior evaluation training: acquiring a training data set, taking the information of the second sample information type of each second sample object in a first preset time length in the data set as the input of a deep learning model, taking the evaluation value of the first behavior type behavior of each second sample object in the data set in a second preset time length after the first preset time length as the expected output of the deep learning model, and training the deep learning model to obtain a trained deep learning model as a behavior prediction deep learning model;

And (3) behavior evaluation testing: acquiring a test data set, taking the information of the second sample information type of each second sample object in the data set within a first preset time length as the input of a behavior prediction deep learning model, taking the evaluation value of the first behavior type behavior of each second sample object in the data set within a second preset time length after a first preset time length as the expected output of the behavior prediction deep learning model, testing the behavior prediction deep learning model, counting the number of times that the difference between the evaluation value of the first behavior type behavior in the expected output and the evaluation value of the first behavior type behavior in the actual output is within a preset range as X1, counting that the evaluation value of the first behavior type behavior in the expected output is greater than the evaluation value of the first behavior in the actual output, and the difference between the evaluation value of the first behavior type occurrence in the expected output and the evaluation value of the first behavior in the actual output is greater than the evaluation value of the first behavior in the actual output The number of times of exceeding the preset range is X2, the number of times that the evaluation value of the first behavior type behavior in the expected output is smaller than the evaluation value of the first behavior type behavior in the actual output and the difference between the evaluation value of the first behavior type behavior in the expected output and the evaluation value of the first behavior type behavior in the actual output exceeds the preset range is X3, X1 is divided by (X1+ X2+ X3) to serve as a behavior evaluation test accuracy rate P2, and X2 is divided by X3 to serve as a behavior evaluation test inclination rate B1;

Calculating the average value and the standard deviation of the behavior evaluation test accuracy rate: calculating the average value AAM of the behavior evaluation test accuracy of the M second object categories, namely (P21+ P22+ … + P2M)/M; calculating the standard deviation ADM ═ (((P21-AM) 2+ (P22-AM) 2+ … + (P2M-AM) 2)/M) 1/2 of the behavior evaluation test accuracy of the M second object categories;

and (3) calculating the bias rate of the behavior evaluation test: acquiring B1j (j ═ 1, 2, …, M) greater than 1 as a behavior evaluation test left-leaning rate for M second object classes; acquiring the number of B1j which is larger than 1 divided by M as the left-leaning proportion of the behavior evaluation test of the M second object categories; b1j smaller than 1 is obtained and used as behavior evaluation test right dip rate of M second object categories; obtaining the proportion of the right inclination of the behavior evaluation test by dividing the number of B1j smaller than 1 by M as M second object categories; calculating the geometric mean LAM of the behavior assessment test left-leaning rates of the M second object classes, which is the geometric mean of the inverses of all B1j that are greater than 1; calculating the geometric mean RAM of the behavior evaluation test right dip rate of the M second object categories, which is the geometric mean of the mean values of all B1j which are less than 1; multiplying (the absolute value of the difference between the left-leaning proportion and the right-leaning proportion) by ((the absolute value of the difference between the geometric mean of the left-leaning rate and the geometric mean of the right-leaning rate) divided by (the sum of the geometric mean of the left-leaning rate and the geometric mean of the right-leaning rate)), to obtain behavior evaluation test bias ratios IM of M second object classes;

taking the third information type as a second sample information type, taking the object of the first object type as a second sample object, and executing the behavior evaluation training step; taking a third information type as a second sample information type, taking an object in a jth (j is 1, 2, …, M) second object type as a second sample object, executing the behavior evaluation test step to obtain the behavior evaluation test accuracy P2, taking the behavior evaluation test accuracy P2j (j is 1, 2, …, M) in a jth subcategory, obtaining the behavior evaluation test tilt rate B1, taking the behavior evaluation test tilt rate B1j (j is 1, 2, …, M) in a jth subcategory, and obtaining the behavior evaluation test accuracy P2j (j is 1, 2, …, M), namely P21, P22, …, P2M in each second object type, and obtaining the behavior evaluation test tilt rate B1j (j is 1, 2, …, M), namely B11, B12, …, B1M;

according to the behavior evaluation test accuracy P2j (j is 1, 2, …, M) of each second object type obtained in the third information type testing step, namely P21, P22, … and P2M, executing the behavior evaluation test accuracy average value and standard deviation calculation step to obtain an average value AAM of the behavior evaluation test accuracy of the M second object types and a standard deviation ADM of the behavior evaluation test accuracy of the M second object types;

and executing the behavior evaluation test bias rate calculation step according to the behavior evaluation test tilt rate B1j (j is 1, 2, …, M), namely B11, B12, … and B1M, of each second object class obtained in the third information type test step to obtain behavior evaluation test bias rates IM of the M second object classes.

6. An artificial intelligence apparatus, the apparatus comprising:

a first information type acquisition module: -said first information type obtaining step for performing said method of claim 1;

a first object class acquisition module: -said first object class acquisition step for performing said method of claim 1;

a second object class acquisition module: -the second object class acquisition step for performing the method of claim 1;

the first-class classification accuracy calculation module: -said first class classification accuracy calculation step for performing said method of claim 1;

the jth sub-category classification accuracy calculation module: -a classification accuracy calculation step for performing said jth sub-category of said method of claim 1;

The ith first class classification accuracy calculation module: -an ith first class classification accuracy calculation step for performing the method of claim 1;

the classification test accuracy calculation module of the ith sub-category: a classification test accuracy calculation step for performing said ith sub-category j of said method of claim 1;

the correlation evaluation module of the ith first information type and the second object category: a correlation evaluation step of the ith first information type and a second object class for performing the method of claim 1;

k types of first information type behavior evaluation modules: -said K first information type behavior evaluation steps for performing said method of claim 1;

a screening set initialization module: -initializing the screening set for performing the method of claim 1;

a start attempt module: the start attempt step for performing the method of claim 1;

the third information type behavior evaluation module: -said third information type behavior evaluation step for performing said method of claim 1;

a deletable judgment module: the deletable determining step for performing the method of claim 1;

An execution control module: the execution control step for executing the method of claim 1.

7. The artificial intelligence device of claim 6, wherein the device further comprises:

a first object acquisition module: -said first object acquisition step for performing said method of claim 2;

a first object information acquisition module: the first object information obtaining step for performing the method of claim 2;

a first usage module: -said first use step for carrying out the method of claim 2;

a first object behavior determination module: -said first object behavior determination step for performing said method of claim 2.

8. The artificial intelligence device of claim 6, wherein the device further comprises:

a category training module: -said class training step for performing said method of claim 3;

a category testing module: -said class testing step for performing said method of claim 3;

a third information type testing module: -said third information type testing step for performing said method of claim 4;

The behavior evaluation test accuracy average value and standard deviation calculation module comprises: a behavior assessment test accuracy average and standard deviation calculation step for performing the method of claim 4;

the behavior evaluation test bias rate calculation module: a behavior evaluation test bias rate calculation step for performing the method of claim 4;

a behavior evaluation training module: -said behavioral assessment training step for performing said method of claim 5;

behavior evaluation test module: the behavior assessment test step for performing the method of claim 5;

the behavior evaluation test accuracy average value and standard deviation calculation module comprises: a behavior assessment test accuracy average and standard deviation calculation step for performing the method of claim 5;

the behavior evaluation test bias rate calculation module: the behavior assessment test bias rate calculation step for performing the method of claim 5.

9. A robot comprising a memory, a processor and an artificial intelligence robot program stored on the memory and executable on the processor, wherein the steps of the method of any one of claims 1 to 5 are carried out when the program is executed by the processor.

10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method according to any one of claims 1 to 5.