Domain Adaptation by Preserving Topology.

Date of Submission

December 2018

Date of Award

Winter 12-12-2019

Institute Name (Publisher)

Indian Statistical Institute

Document Type

Master's Dissertation

Degree Name

Master of Technology

Subject Name

Computer Science

Department

Electronics and Communication Sciences Unit (ECSU-Kolkata)

Supervisor

Pal, Nikhil Ranjan (ECSU-Kolkata; ISI)

Abstract (Summary of the Work)

Domain adaptation is highly researched area among machine learning experts. In domain adaptation we use a domain with enough class label (source domain) and try to predict class label of a different data which doesn’t have any class label (target domain) . Both source and target domain share the same features space. Many approaches are there but one popular approach is to reduce the distance between source and target domain data distributions. There are many algorithm that tries to project the source and target data into a latent space so that the distance between data distribution of source and target domain reduces. But when we try to project data into a latent space the shape of data may change . Most domain adaptation algorithm does not try to preserve the shape of data explicitly. Projection of data into a latent space may change the shape of data or precisely the topology of the data. If we can preserve the topology of the data when we are projecting the data into a latent space then we may achieve better accuracy . In this thesis we have developed a method so that we can preserve the topology of the data at the time of projecting it into a latent space . For projection in to a latent space we have used auto-encoder that create encoded version of data at hidden layers. Hidden layer representation of an autoencoder is a projection of data into a latent space . The distance between data distribution of hidden layer representation of source and target data reduces . In an auto-encoder ,the shape of the data may not be preserved i.e auto-encoder is not forced to preserve any topological property of the data. In this thesis we have added a extra constraint in a auto-encoder so that auto-encoder can preserve topological property of the data . Preserving topology of data can enhance the ability of the classifier that is trained on source data to correctly classify target data .

Comments

ProQuest Collection ID: http://gateway.proquest.com/openurl?url_ver=Z39.88-2004&rft_val_fmt=info:ofi/fmt:kev:mtx:dissertation&res_dat=xri:pqm&rft_dat=xri:pqdiss:28843242

Control Number

ISI-DISS-2018-384

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

DOI

http://dspace.isical.ac.in:8080/jspui/handle/10263/6950

This document is currently not available here.

Share

COinS