Harnessing meta-learning via probabilistic modelling and trajectory optimisation

Simple item page
dc.contributor.authorNguyen, Cuong Cao
dc.contributor.schoolSchool of Computer Scienceen
dc.date.issued2021
dc.description.abstractMeta-learning has recently flourished as one of the most promising transfer learning techniques that can adapt quickly to a new task, even if that task consists of a limited number of training examples. The main idea of meta-learning is to use a meta-parameter to model the shared structure between many observed tasks and utilise the knowledge gained from such modelling to facilitate the learning for unobserved tasks. Despite steady progress with many remarkable state-of-the-art results, existing meta-learning algorithms are often fragile due to the lack of studies in prediction uncertainty and generalisation for unseen tasks. In addition, little is known about how tasks are related to each other, potentially leading to sub-optimal solutions due to the assumption that tasks are evenly distributed – which is hardly true in practice. This thesis, therefore, aims to address such problems through the lenses of probabilistic modelling and optimisation. In particular, the thesis proposes to (i) integrate variational inference into meta-learning that considers the epistemic uncertainty into the modelling to reduce calibration errors and overfitting induced by meta-learning models, (ii) derive a PAC-Bayes upper-bound of errors evaluated on both seen and unseen tasks to enable the study of theoretical generalisation in meta-learning and use that bound to formulate a loss function applied in the training of different meta-learning methods, (iii) model tasks via a variant of Gaussian latent Dirichlet allocation and utilise the newly-obtained representation for task selection to make training more efficient, and (iv) adopt trajectory optimisation from optimal control to determine the re-weighting factor of each training task to optimise the training process of meta-learning. The results of these studies improve the robustness and provide an insightful understanding of meta-learning, and thus, enable further development of practical meta-learning approaches.en
dc.description.dissertationThesis (Ph.D.) -- University of Adelaide, School of Computer Science, 2022en
dc.identifier.urihttps://hdl.handle.net/2440/135234
dc.language.isoenen
dc.provenanceThis electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legalsen
dc.subjectMachine learningen
dc.subjectMeta-learningen
dc.titleHarnessing meta-learning via probabilistic modelling and trajectory optimisationen
dc.typeThesisen

Files

Original bundle
Now showing 1 - 1 of 1
No Thumbnail Available
Name:
Nguyen2022_PhD.pdf
Size:
2.09 MB
Format:
Adobe Portable Document Format
Description:
Download

Collections

Research Theses