Description:

It is highly desirable to represent a motion dataset with a prototype-based (PB) model, such that the motion categories can be described appropriately by one or more representatives with interpretable characteristics. Accordingly, I develop prototype-based models for the representation of motion datasets using dictionary learning (DL) designs. In a DL framework, one can consider each dictionary atom as a prototype which represents the data samples via a reconstruction-based relationship. Nevertheless, these atoms do not provide the general characteristics which are expected from a PB model.

Therefore, in the particular dictionary learning algorithm proposed in this Ph.D. work, I focused on the following aspects of prototype-based learning:

• Flexibility: The optimal number of prototypes per class is learned by the algorithm and depends on the complexity and structure of classes in the dataset.
• Interpretability: The prototypes are learned to be interpretable by the classes they belong to while they are still flexible to represent the overlapping of different classes efficiently.
• Discriminability: In a discriminative prototype-based framework, the prototypes are trained to distinguish the members of the classes to which they belong from other classes.
• Multiple-representations/feature selection: Assume that the data is represented by multiple modalities or if each the distribution of the data in each dimension is considered individually. The PB framework finds an efficient (weighted) combination of these representations to fulfill above concerns more optimally.

The above concerns are addressed together or individually in the DL frameworks of my Ph.D. project.