Supervised target data hashing with domain adaptation
In many real-world applications, the ability of knowledge transfer is required everywhere, which leads to a dramatic increase of the research interests targeting on cross domain learning. When a large amount of labeled data is required for training a model, one feasible solution is to reuse knowledg...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Final Year Project |
| Language: | English |
| Published: |
2018
|
| Subjects: | |
| Online Access: | http://hdl.handle.net/10356/73945 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Nanyang Technological University |
| Language: | English |
| Summary: | In many real-world applications, the ability of knowledge transfer is required everywhere, which leads to a dramatic increase of the research interests targeting on cross domain learning. When a large amount of labeled data is required for training a model, one feasible solution is to reuse knowledge learned from another dataset which has plenty of information provided and has some similarities with the source dataset. However, simply using the knowledge obtained from source data and putting it directly on the target domain might not be able to generate good performance, especially when the data distributions of the two domains have large differences. Thus, it is necessary to introduce a domain adaptation method that could reduce the effect of domain shift.
Furthermore, with the explosive growth of data, images of large volume and high dimension are pervasive, which leads to challenges like storage capability as well as data retrieval efficiency. To tackle this concern, hashing has emerged as one of the most popular solutions. It is a technique that transforms high dimensional data into precisely compact binary values, which is being widely used in various of computer vision tasks.
The objective of this project is to integrate domain adaptation with hash learning such that target data retrieval can be performed precisely. A novel model architecture is introduced in this report. It consists of three major parts, a Basic Feature Extractor, a Shared Code Generator and a Specific Code Generator. The Shared Code Generator is used to reduce domain disparity and outputs code that is extracted from both domains, while the Specific Code Generator is only designed for target data in order to capture more information on the target domain. Various experiments are carried out and based on empirical studies, the proposed framework achieves great performance. |
|---|