Machine learning apparatus, machine learning method, and computer-readable recording medium
A machine learning apparatus 100 is an apparatus for constructing, by transfer learning, a second identification dictionary to be used in a second task from a first identification dictionary to be used in a first task. The machine learning apparatus 100 includes: a pseudo data generation unit 10 that generates pseudo data by processing one of real data in the first task and real data in the second task or both pieces of real data; and a learning unit 20 that constructs a third identification dictionary by performing, using the first identification dictionary, first transfer learning using the pseudo data as training data, and furthermore, constructs the second identification dictionary by performing, using the third identification dictionary, second transfer learning using the real data in the second task as training data.
1. A machine learning apparatus for constructing, by transfer learning, a second identification dictionary to be used in a second task from a first identification dictionary to be used in a first task, the machine learning apparatus comprising:
a processor; and
a memory storing executable instructions that, when executed by the processor, causes the processor to perform as:
a pseudo data generation unit configured to generate pseudo data by processing one of real data in the first task and real data in the second task or both pieces of real data; and
a learning unit configured to construct a third identification dictionary by performing, using the first identification dictionary, first transfer learning using the pseudo data as training data, and furthermore, constructing the second identification dictionary by performing, using the third identification dictionary, second transfer learning using the real data in the second task as training data,
wherein, in a case where the first task is a task for identifying a motion of a person, and the second task is a task for identifying a motion that is different from the motion to be identified by the first task,
the pseudo data generation unit generates the pseudo data by specifying, from video data of a person that is the real data in the second task, a joint of the person, and executing image processing in which a perturbation is given to the specified joint with respect to the video data so as to change the posture of the person.
2. The machine learning apparatus according to claim 1 ,
wherein the pseudo data generation unit generates data similar to the real data in the second task as the pseudo data.
3. The machine learning apparatus according to claim 1 ,
wherein the pseudo data generation unit acquires a loss value calculated at the time of second transfer learning, and generates the pseudo data while adjusting the degree of perturbation to be given to the joint based on the acquired loss value.
4. The machine learning apparatus according to claim 1 ,
wherein the pseudo data generation unit acquires a loss value calculated at the time of first transfer learning, and generates the pseudo data while adjusting the degree of perturbation to be given to the joint based on the acquired loss value.
5. The machine learning apparatus according to claim 1 , further comprising:
a pseudo data selection unit configured to acquire a loss value calculated at the time of second transfer learning, selecting a specific piece of pseudo data from the generated pseudo data based on the acquired loss value, and outputting only the selected piece of pseudo data to the learning unit.
6. The machine learning apparatus according to claim 1 , further comprising:
a pseudo data selection unit configured to acquire a loss value calculated at the time of first transfer learning, selecting a specific piece of pseudo data from the generated pseudo data based on the acquired loss value, and outputting only the selected piece of pseudo data to the learning unit.
7. A machine learning method for constructing, by transfer learning, a second identification dictionary to be used in a second task from a first identification dictionary to be used in a first task, the machine learning method comprising:
generating pseudo data by processing one of real data in the first task and real data in the second task or both pieces of real data; and
constructing a third identification dictionary by performing, using the first identification dictionary, first transfer learning using the pseudo data as training data, and furthermore, constructing the second identification dictionary by performing, using the third identification dictionary, second transfer learning using the real data in the second task as training data,
wherein, in a case where the first task is a task for identifying a motion of a person, and the second task is a task for identifying a motion that is different from the motion to be identified by the first task,
in the generating, the pseudo data is generated by specifying, from video data of a person that is the real data in the second task, a joint of the person, and executing image processing in which a perturbation is given to the specified joint with respect to the video data so as to change the posture of the person.
8. The machine learning method according to claim 7 ,
wherein, in the generating, data similar to the real data in the second task is generated as the pseudo data.
9. The machine learning method according to claim 7 ,
wherein, in the generating, a loss value calculated at the time of second transfer learning is acquired, and the pseudo data is generated while adjusting the degree of perturbation to be given to the joint based on the acquired loss value.
10. The machine learning method according to claim 7 ,
wherein, in the generating, a loss value calculated at the time of first transfer learning is acquired, and the pseudo data is generated while adjusting the degree of perturbation to be given to the joint based on the acquired loss value.
11. The machine learning method according to claim 7 , further comprising:
acquiring a loss value calculated at the time of second transfer learning, selecting a specific piece of pseudo data from the generated pseudo data based on the acquired loss value, and outputting only the selected piece of pseudo data for the constructing.
12. The machine learning method according to claim 7 , further comprising:
acquiring a loss value calculated at the time of first transfer learning, selecting a specific piece of pseudo data from the generated pseudo data based on the acquired loss value, and outputting only the selected piece of pseudo data for the constructing.
13. A non-transitory computer-readable recording medium that includes a program recorded thereon for constructing, using a computer, a second identification dictionary to be used in a second task from a first identification dictionary to be used in a first task, by transfer learning, the program including instructions that cause the computer to carry out:
generating pseudo data by processing one of real data in the first task and real data in the second task or both pieces of real data; and
constructing a third identification dictionary by performing, using the first identification dictionary, first transfer learning using the pseudo data as training data, and furthermore, constructing the second identification dictionary by performing, using the third identification dictionary, second transfer learning using the real data in the second task as training data,
wherein, in a case where the first task is a task for identifying a motion of a person, and the second task is a task for identifying a motion that is different from the motion to be identified by the first task,
in the generating, the pseudo data is generated by specifying, from video data of a person that is the real data in the second task, a joint of the person, and executing image processing in which a perturbation is given to the specified joint with respect to the video data so as to change the posture of the person.
14. The non-transitory computer-readable recording medium according to claim 13 ,
wherein, in the generating, data similar to the real data in the second task is generated as the pseudo data.
15. The non-transitory computer-readable recording medium according to claim 13 ,
wherein, in the generating, a loss value calculated at the time of second transfer learning is acquired, and the pseudo data is generated while adjusting the degree of perturbation to be given to the joint based on the acquired loss value.
16. The non-transitory computer-readable recording medium according to claim 13 ,
wherein, in the generating, a loss value calculated at the time of first transfer learning is acquired, and the pseudo data is generated while adjusting the degree of perturbation to be given to the joint based on the acquired loss value.
17. The non-transitory computer-readable recording medium according to claim 13 ,
wherein the program further includes instructions that cause the computer to carry out:
acquiring a loss value calculated at the time of second transfer learning, selecting a specific piece of pseudo data from the generated pseudo data based on the acquired loss value, and outputting only the selected piece of pseudo data for the constructing.