Transfer learning based on cross-domain homophily influences
Methods, computer program products, and systems are presented. The methods can include, for instance: generating a plurality of deep transfer learning networks. Further, the methods can include, for instance: encoding one or more transfer layers.
1. A computer implemented method comprising:
generating a plurality of deep transfer learning networks based on a plurality of first exemplars and a plurality of second exemplars;
encoding one or more transfer layers to a chromosome for genetic operators, where the one or more transfer layers are to be transferred from a source deep transfer learning network corresponding to a source domain to a target deep transfer learning network corresponding to a target domain;
diversifying concurrently both the source deep transfer learning network and the target deep transfer learning network by use of the genetic operators; and
producing the target deep transfer learning network that integrates a result from the diversifying.
2. The computer implemented method of claim 1 , wherein a first subset of the first exemplars represents the source domain, a second subset of the first exemplars represents the target domain, and the second exemplars represent both the source domain and the target domain.
3. The computer implemented method of claim 1 , wherein the plurality of deep transfer learning networks include both the source deep transfer learning network and the target deep transfer learning network.
4. The computer implemented method of claim 1 , wherein the producing the target deep transfer learning network that integrates a result from the diversifying includes producing the target deep transfer learning network that integrates a result from the diversifying on the source deep transfer learning network and that passes a predefined fitness threshold condition for the target deep transfer learning network.
5. The computer implemented method of claim 4 , wherein the predefined fitness threshold condition is determined based on the accuracy of the target deep transfer learning network as determined from an average between ground truths specific to the target domain and ground truths applicable for both the source domain and the target domain.
6. The computer implemented method of claim 1 , further comprising:
prior to the encoding, training the source deep transfer learning network with a first subset of the first exemplars and the second exemplars; and
prior to the encoding, and concurrently with the training the source deep transfer learning network, training the target deep transfer learning network with a second subset of the first exemplars and the second exemplars.
7. The computer implemented method of claim 1 , further comprising:
prior to the encoding, measuring a homophily influence of the source deep transfer learning network to the target deep transfer learning network.
8. The computer implemented method of claim 1 , the diversifying comprising:
mutating randomly the transfer layers of the source deep transfer learning network from the encoding.
9. The computer implemented method of claim 1 , the diversifying comprising:
ascertaining that respective transfer layers of first and second source deep transfer learning networks have their respective weights migrated over to the target deep transfer learning network for respectively passing a fitness threshold condition for the source deep transfer learning network; and
crossing over the transfer layers of the target deep transfer learning network with the transfer layers of the first and second source deep transfer learning networks.
10. The computer implemented method of claim 1 , the diversifying comprising:
ascertaining that respective transfer layers of more than three source deep transfer learning networks have their respective weights migrated over to the target deep transfer learning network for respectively passing a fitness threshold condition for the source deep transfer learning network;
selecting first and second source deep transfer learning networks from the more than three source deep transfer learning network from the ascertaining; and
crossing over the transfer layers of the target deep transfer learning network with the transfer layers of the first and second source deep transfer learning networks from the selecting.
11. A computer program product comprising:
a computer readable storage medium readable by one or more processor and storing instructions for execution by the one or more processor for performing a method comprising:
generating a plurality of deep transfer learning networks based on a plurality of first exemplars and a plurality of second exemplars;
encoding one or more transfer layers to a chromosome for genetic operators, where the one or more transfer layers are to be transferred from a source deep transfer learning network corresponding to a source domain to a target deep transfer learning network corresponding to a target domain;
diversifying concurrently both the source deep transfer learning network and the target deep transfer learning network by use of the genetic operators; and
producing the target deep transfer learning network that integrates a result from the diversifying.
12. A system comprising:
a memory;
one or more processor in communication with the memory; and
program instructions executable by the one or more processor via the memory to perform a method comprising:
generating a plurality of deep transfer learning networks based on a plurality of first exemplars and a plurality of second exemplars;
encoding one or more transfer layers to a chromosome for genetic operators, where the one or more transfer layers are to be transferred from a source deep transfer learning network corresponding to a source domain to a target deep transfer learning network corresponding to a target domain;
diversifying concurrently both the source deep transfer learning network and the target deep transfer learning network by use of the genetic operators; and
producing the target deep transfer learning network that integrates a result from the diversifying.
13. The system of claim 12 , wherein a first subset of the first exemplars represents the source domain, a second subset of the first exemplars represents the target domain, and the second exemplars represent both the source domain and the target domain.
14. The system of claim 12 , wherein the plurality of the deep transfer learning networks include both the source deep transfer learning network and the target deep transfer learning network.
15. The system of claim 12 , wherein the producing the target deep transfer learning network that integrates a result from the diversifying includes producing the target deep transfer learning network that integrates a result from the diversifying on the source deep transfer learning network and that passes a predefined fitness threshold condition for the target deep transfer learning network.
16. A computer implemented method comprising:
generating a plurality of deep transfer learning networks based on a plurality of first exemplars and a plurality of second exemplars;
training a source deep transfer learning network with a subset of the first exemplars;
encoding one or more transfer layers to a chromosome for genetic operators, where the one or more transfer layers are to be transferred from a source deep transfer learning network corresponding to a source domain to a target deep transfer learning network corresponding to a target domain;
mutating randomly the transfer layers of the source deep transfer learning network from the encoding; and
ascertaining that a fitness threshold condition for the source deep transfer learning network is satisfied.
17. The computer implemented method of claim 16 , further comprising:
prior to the encoding, measuring a homophily influence of the source deep transfer learning network to the target deep transfer learning network.
18. The computer implemented method of claim 16 , wherein a first subset of the first exemplars represents the source domain, a second subset of the first exemplars represents the target domain, and the second exemplars represent both the source domain and the target domain.
19. The computer implemented method of claim 16 , wherein the plurality of deep transfer learning networks include both the source deep transfer learning network and the target deep transfer learning network.
20. The computer implemented method of claim 16 , wherein the method includes migrating weights of respective transfer layers of first and second or more source deep transfer learning networks over to the target deep transfer learning network respectively.
21. A computer implemented method comprising:
generating a plurality of deep transfer learning networks based on a plurality of first exemplars and a plurality of second exemplars;
training a target deep transfer learning network with a subset of the first exemplars and the second exemplars;
encoding one or more transfer layers to a chromosome for genetic operators, where the one or more transfer layers are to be transferred from a source deep transfer learning network corresponding to a source domain to a target deep transfer learning network corresponding to a target domain;
ascertaining that respective transfer layers of first and second source deep transfer learning networks have their respective weights migrated over to the target deep transfer learning network; and
crossing over the transfer layers of the target deep transfer learning network with the transfer layers of the first and second source deep transfer learning networks from the ascertaining.
22. The computer implemented method of claim 21 , wherein the method includes producing, the target deep transfer learning network that integrates a result from the crossing over on the target deep transfer learning network and that passes a predefined fitness threshold condition for the target deep transfer learning network.
23. The computer implemented method of claim 21 , wherein the plurality of deep transfer learning networks include both the source deep transfer learning network and the target deep transfer learning network.
24. The computer implemented method of claim 21 , further comprising:
prior to the crossing over, randomly selecting the first and second source deep transfer learning network from a pool of source deep transfer learning networks that had migrated over respective weights of the transfer layers.
25. The computer implemented method of claim 21 , wherein a first subset of the first exemplars represents the source domain, a second subset of the first exemplars represents the target domain, and the second exemplars represent both the source domain and the target domain.