Generalizing and Decoding SVM Classification of Spatiotemporal Data

This presentation of my dissertation progress consists of two main parts.

After giving a general overview, I will present new insights into Support Vector Machine (SVM) Classification. The Balanced Relative Margin Machine (BRMM) will be introduced as a more general model, which connects among others Fisher's Discriminant Classification (FDA/LDA), SVMs and Support Vector Regression (SVR). Furthermore, connections between the different solution algorithms and their online variants will be given. Applications to incremental learning and hyperparameter optimization will be shown.

The second part focuses on the understanding of the resulting classifier and will show a way to interpret the classification in combination with the preprocessing, especially on spatio-temporal data. This can be used for visualization, feature or sensor selection and improved incremental learning.

Both concepts are or will be integrated into pySPACE to improve this signal processing and classification environment.