Representative Instances Selection
Event details
| Date | 01.07.2016 |
| Hour | 14:00 › 16:00 |
| Speaker | Aleksei Triastcyn |
| Location | |
| Category | Conferences - Seminars |
EDIC Candidacy Exam
Exam President: Prof. Karl Aberer
Thesis Director: Prof. Boi Faltings
Co-examiner: Prof. Matthias Grossglauser
Background papers:
Machine Teaching: An Inverse Problem to Machine Learning and an Approach Toward Optimal Education.
Submodularity in Data Subset Selection and Active Learning.
Leveraging for Big Data Regression.
Abstract
We are looking at the problem of finding the most representative examples in a dataset. Exponential growth of data nowadays makes
it difficult to process, store, or transfer information. Thus, it would be beneficial to select a small amount of representative examples
that can be used to train machine learning models without compromising the quality, be transferred over network easily or stored for
prolonged periods of time. In this proposal, we consider three different perspectives: machine teaching, submodular data subset
selection, and leverage score sampling. We report the main ideas of each approach, as well as their shortcomings. Finally, we outline
the future research possibilities, potential solutions to existing drawbacks, and possible applications, which include human teaching,
security, and multi-agent systems.
Exam President: Prof. Karl Aberer
Thesis Director: Prof. Boi Faltings
Co-examiner: Prof. Matthias Grossglauser
Background papers:
Machine Teaching: An Inverse Problem to Machine Learning and an Approach Toward Optimal Education.
Submodularity in Data Subset Selection and Active Learning.
Leveraging for Big Data Regression.
Abstract
We are looking at the problem of finding the most representative examples in a dataset. Exponential growth of data nowadays makes
it difficult to process, store, or transfer information. Thus, it would be beneficial to select a small amount of representative examples
that can be used to train machine learning models without compromising the quality, be transferred over network easily or stored for
prolonged periods of time. In this proposal, we consider three different perspectives: machine teaching, submodular data subset
selection, and leverage score sampling. We report the main ideas of each approach, as well as their shortcomings. Finally, we outline
the future research possibilities, potential solutions to existing drawbacks, and possible applications, which include human teaching,
security, and multi-agent systems.
Practical information
- General public
- Free
Contact
- Cecilia Chapuis EDIC