Towards Unsupervised Transfer Learning
Event details
| Date | 30.08.2023 |
| Hour | 14:15 › 17:15 |
| Speaker | Artem Gadetskii |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Martin Jaggi
Thesis advisor: Prof. Maria Brbic
Co-examiner: Prof. François Fleuret
Abstract
Transfer learning is a fundamental paradigm in
machine learning that leverages the pretraining of deep neural
networks on large-scale datasets to transfer knowledge to a downstream
task with limited resources. A well-established approach
involves fine-tuning the weights of a linear classifier on top of
frozen representations in a supervised manner. This approach
has demonstrated success in various domains, such as few-shot
learning, domain adaptation, and domain generalization. Recently,
foundation models trained on extensive data have achieved
significant advancements across many modalities, exhibiting
exceptional representation learning capabilities and enabling
zero-shot predictions. However, existing approaches presume the
availability of at least minimal supervision for downstream tasks,
preventing their application to more challenging fully-unsupervised
settings. To overcome this limitation, deep clustering methods can
be applied to solve the task of interest without requiring human
annotations. Yet, the performance of such unsupervised methods
is lagging far behind. To bridge the gap between supervised
and unsupervised learning we propose to consider a general
unsupervised transfer learning setting pushing the limits of the
recent breakthroughs in representation learning achieved by
foundational models.
Background papers
Exam president: Prof. Martin Jaggi
Thesis advisor: Prof. Maria Brbic
Co-examiner: Prof. François Fleuret
Abstract
Transfer learning is a fundamental paradigm in
machine learning that leverages the pretraining of deep neural
networks on large-scale datasets to transfer knowledge to a downstream
task with limited resources. A well-established approach
involves fine-tuning the weights of a linear classifier on top of
frozen representations in a supervised manner. This approach
has demonstrated success in various domains, such as few-shot
learning, domain adaptation, and domain generalization. Recently,
foundation models trained on extensive data have achieved
significant advancements across many modalities, exhibiting
exceptional representation learning capabilities and enabling
zero-shot predictions. However, existing approaches presume the
availability of at least minimal supervision for downstream tasks,
preventing their application to more challenging fully-unsupervised
settings. To overcome this limitation, deep clustering methods can
be applied to solve the task of interest without requiring human
annotations. Yet, the performance of such unsupervised methods
is lagging far behind. To bridge the gap between supervised
and unsupervised learning we propose to consider a general
unsupervised transfer learning setting pushing the limits of the
recent breakthroughs in representation learning achieved by
foundational models.
Background papers
- Learning transferable visual models from natural language supervision. Alec Radford et al. International Conference on Machine Learning 2021. (link)
- SCAN: Learning to classify images without labels. Wouter Van Gansbeke et al. European Conference on Computer Vision 2020. (link)
- Learning to discover novel visual categories via deep transfer clustering. Kai Han et al. International Conference on Computer Vision 2019 (link)
Practical information
- General public
- Free