Feature-Based Transfer Learning in Novel Systems

In recent years, the transfer learning framework has gained increasing interest in the machine learning community. Fundamentally, this framework aims to train a new system called target domain using existing knowledge from one or more previous system(s) called source domain(s). By extending the theory of standard machine learning techniques, this framework allows us to solve many challenging problems directly and intuitively. A domain is a specialized area, application or system developed to achieve specific goals. A novel system is one that is significantly different from previous approaches used to achieve similar goals. It receives a different set of inputs but is expected to produce similar outputs. On the other hand, a non-stationary system is one that is that receives the same set of inputs – albeit the inputs shift or drift depending on circumstances. As a result, models trained initially on one set of data distributions do not perform well on a different set of data distributions. Therefore, traditional machine techniques are not sufficient to train novel or non-stationary systems. In addition, in many problem areas, training the target system from scratch could be time-consuming and expensive. By effectively recycling knowledge from the source system, these limitations could be alleviated. For knowledge to be shared the systems must be related – they must achieve similar goals. Generally speaking, transfer learning consists of two steps. Firstly, an adaptive step called domain adaptation is applied to reduce data distribution differences between the source domain and the target domain. Secondly, a knowledge transfer step is applied to model and transfer some expert knowledge from the source domain to the target domain. Several techniques have already been proposed to deal with transfer learning problems. In this research, we are particularly interested in a form of transfer learning called feature-based transfer where the domain adaptation step is based on changing the feature spaces of the source and target domains. In the original feature space, the source and target domains are different. By mapping the data to a new space, their data distributions are made similar so that traditional machine learning techniques can then be applied to model the knowledge in the source domain and transfer it over to the target domain. Motivated by a real-world application, we present a weakly supervised feature-based transfer algorithm called G2DA. This approach uses the Generalized Discriminant Analysis (GDA) to generate potential spaces for domain adaptation. In particular, G2DA uses tokens during domain adaptation to ensure that novel target systems are aligned correctly.

Name thesis:degree_name

Doctor of Philosophy

Grantor

Texas Tech University

Year dc:date.issued

2012

Author dc:creator

• Shahriar, Muneem

Advisor dc:contributor.advisor

• Sari-Sarraf, Hamed

Committee members dc:contributor.committeemember

• Hequet, Eric
• Sridharan, Mohan
• Pal, Ranadip

Handle dc:identifier.uri

https://hdl.handle.net/2346/106621

OAI identifier oai:identifier

oai:ttu-ir.tdl.org:2346/106621