Robust classifier training with small data sets
(Suitable for MSc thesis)
While big data fuels the machine learning progress based on deep neural networks, many real-life problems come with limited data sets only. This is specifically the case for some medical domains, where the number of patients to learn from is small and a single recording of a patient is limited to a few hundred data points at most.
Machine learning models trained on and applied for such small data sets can profit from data augmentation and transfer learning. In this project, a novel strategy shall be evaluated for the robustification of covariance matrices, which are the basis for several subspace projection methods such as the linear discriminant analysis (LDA).
Required:
- solid math background, specifically linear algebra
Filed under:
student project