Generalizing Domain Adaptation: Relaxing Task Assumptions & an Alternative Cost Function for

Postgraduate Thesis uoadl:2884539 260 Read counter

Unit:
Κατεύθυνση / ειδίκευση Θεωρητική Πληροφορική (ΘΕΩ)
Πληροφορική
Deposit date:
2019-11-01
Year:
2019
Author:
Pikramenos Georgios
Supervisors info:
Σταύρος Περαντώνης,
ΕΡΕΥΝΗΤΗΣ Α,
Ινστιτούτο Πληροφορικής και Τηλεπικοινωνιών,
ΕΚΕΦΕ ΔΗΜΟΚΡΙΤΟΣ
Original Title:
Γενικεύοντας την Προσαρμoγή Πεδίου: Χαλάρωση των Υποθέσεων & μία Εναλλακτική Συνάρτηση Κόστους για Μεθόδους με Αντιμαχόμενα Δίκτυα
Languages:
English
Translated title:
Generalizing Domain Adaptation: Relaxing Task Assumptions & an Alternative Cost Function for
Summary:
Supervised learning techniques typically work under the assumption that train and test
datasets are drawn from the same distribution. As such, training useful models with con-
ventional techniques for a supervised learning task, requires us to obtain at least some
labeled data for our problem of interest. Extensive research efforts in the fields of ML
and in particular DL have yielded powerful methods for tackling supervised problems and
developments in big data systems have made it possible to gather raw data at unprece-
dented rates. These advances have seen, to a large extent, the bottleneck of the learning
procedure shift from modelling/training to obtaining labels for training data. We argue that
in order to fully utilize our technologies, we need to bypass this bottleneck by creating
models that are robust under train and test data distribution discrepancies. DA is a frame-
work that addresses the aforementioned issues and under certain assumptions, provides
tools to resolve them. In this thesis we discuss improvements on current techniques for
DA that rely on adversarial neural networks. We introduce a new cost function for such
methods inspired by progress in generative adversarial networks (GANs) and the field of
optimal transport. Finally, we propose a novel problem setup, termed two-way partial do-
main adaptation, which relaxes the assumptions made in traditional DA and we present a
first algorithm to tackle problems in this setup.
Main subject category:
Technology - Computer science
Keywords:
Transfer Learning, Domain Adaptation, Adversarial Neural Networks
Index:
Yes
Number of index pages:
5
Contains images:
Yes
Number of references:
32
Number of pages:
84
GeorgePikramenosDissertation.pdf (1 MB) Open in new window