Presentation is loading. Please wait.

Presentation is loading. Please wait.

Chang WangChang Wang, Sridhar mahadevanSridhar mahadevan.

Similar presentations


Presentation on theme: "Chang WangChang Wang, Sridhar mahadevanSridhar mahadevan."— Presentation transcript:

1 Chang WangChang Wang, Sridhar mahadevanSridhar mahadevan

2  Problem Introduction  Problem Definition  Who cares  Previous work & challenges  Contribution  A glance at methods

3  Example problem  Input: Three collections of documents in  English (sufficient labels)  Italian (sufficient labels)  Arabic (few labels).  Target: Assign labels to the Arabic documents.  A way: find a common feature space for 3 domains  Shared labels, (sports, military)  No shared documents. (no instance correspondence)  No words translations are available.

4 English docs Italian Docs Arabic docs Shared label set: {sports, military} No corresponding instances or words Question : Can we construct a common feature space so that we can use English docs and Italian docs to help classify Arabic docs? docword1word2…label 102sports 220military 310 401sports docparola1parola2…etichetta 120sports 202military 301 410sports doc1 كلمة 2 كلمة … ملصق 102sports 220military 310? 401?

5 English docs Italian docs Arabic docs docword1word2…label 102sports 220military 310 401sports docparola1parola2…etichetta 120sports 202military 301 410sports doc1 كلمة 2 كلمة … ملصق 102sports 220military 310? 401? docfeature1feature2…label 102sports 220military 310 401sports 502 620military 710 801sports 902 1020military 1110? 1201? Common feature space

6  Given K input datasets in different domains, with different features, but all of the datasets shared the same label set.  Source domain have sufficient labeled instances.  Target domain have few labeled instances.  Question: Can we construct a common feature space?  So all instances in different domain can be mapped to the same feature space, so that we can perform learning task?

7 Source k Source 1 Target Common feature space Learning

8

9  Search engine  classify docs, rank docs, find docs topics  Businessman  Customer clustering  Biologist  Match protein

10  Target domain have little labels  No instance correspondence  Source domain and target domain have different feature space

11  Most work assumes that the source domain and the target domain have the same features.  Manifold regularization  Do not leverage source domain information  Transfer learning based on manifold alignment: use both label and unlabeled instance to learn mapping  require small amount of instance correspondence.

12  Transfer learning perspective  Can work on different feature space  Cope with multiple input domain  Can combine with existing domain adaption methods  Manifold alignment perspective  Need no instance correspondence  Use label to learn alignment

13

14 English docs Italian docs Arabic docs docword1word2…label 102sports 220military 310 401sports docparola1parola2…etichetta 120sports 202military 301 410sports doc1 كلمة 2 كلمة … ملصق 102sports 220military 310? 401? docfeature1feature2…label 102sports 220military 310 401sports 502 620military 710 801sports 902 1020military 1110? 1201? Common feature space

15


Download ppt "Chang WangChang Wang, Sridhar mahadevanSridhar mahadevan."

Similar presentations


Ads by Google