Applications of machine learning in dynamic settings
Stephanie Brandl · DepositOnce · 2021
The field of machine learning is deeply intertwined with the dynamics that determine technical progress in our lives. It benefits invaluably from the increasing amount of available data and the access to computing power. Simultaneously, it drives those changes by continuously improving methodologies and thus extending the possible applications in our everyday lives. In order to keep up with those changes, machine learning models need to tackle the obstacles that arise in dynamic settings. This thesis contributes solutions to those challenges in the fields of computational neuroscience and natural language processing. The first contribution researches possible challenges when moving brain-computer interfaces out of the lab and into the real world. We have recorded data from 16 participants conducting a motor imagery task while handling secondary distraction tasks that simulate everyday life situations. We investigate the artifacts that contaminate the data which makes it difficult to successfully classify within the standard pipeline. We propose two approaches that tackle those difficulties and significantly improved classification results. In the second contribution, we propose a new version of Source Power Co-Modulation: Fourier-SPoC (f-SPoC) extracts brain components maximally correlating with a regular target function. We apply f-SPoC to the magnetoencephalogram (MEG) from a passive beat listening task to find neural correlates that show rhythmic entrainment with the regular beat. The resulting components show very regular patterns of peaks and troughs in the spectrogram on the group average and even on some individuals. The third contribution provides two new methods: Word2Vec with Structure Constraint (W2V&Constr) and Word2Vec with Structure Prediction (W2V&Pred) with which we can learn dynamic word embeddings of high quality based on the underlying structure in the dataset (W2V&Constr) and are even able to predict the structure of the dataset (W2V&Pred). We apply both methods to three datasets of different structures where we achieve significant improvements on both the embedding accuracy and the structure score compared to the baseline and capture interesting insights with respect to outliers and the underlying data structure. Furthermore, we apply W2V&Pred on a German dataset in an explorative experiment to find connections between different authors where the underlying structure is not known a-priori.