Label noise in supervised and unsupervised classification
(Suitable for MSc thesis and MSc project)
Supervised classification methods require labeled data to train a classification model. In practice, however, training data typically contains a certain amount of wrongly labeled examples. In the domain of brain-computer interfaces, however, recently approaches for unsupervised classification has been proposed, which exploits specific domain knowledge in order to train a classification model from scratch (random initialization) without any labels, see https://ieeexplore.ieee.org/document/8335845
In this project, it shall be analyzed, how sensitive different supervised and unsupervised ML models are w.r.t. label noise.
Required:
- solid machine learning background
- solid math background (specifically linear algebra)
- Fluent in python and with ML toolboxes
Filed under:
student project