Nils was working on "Synthetic Data Generation of Resistive Pulse Signals from Biological Nanopores" supervised by Julian Hoßbach. We congratulate him on this wonderful achievement and invite you to attend the defense.
Abstract: Machine learning applied to nanopore sensing data can be used to detect and classify post-translational modifications of proteins, but its performance is limited by a shortage of labeled training data, as measurements are expensive, time-consuming, and vary considerably. This thesis presents a method to address this data shortage by generating synthetic data from the statistics of existing measurements. White Gaussian noise is scaled with the power spectral density estimated from a measurement, producing synthetic blockade events that retain the spectral properties of the real data. The effect of this synthetic data on classification performance is then evaluated using masked autoencoder pretraining. In particular, the experiments investigate whether the synthetic data carries useful information, how its benefit depends on the dataset size, and whether class-specific synthetic data can help mitigate class imbalance. In nearly all scenarios, the synthetic data improves classification accuracy.
Best of luck Nils on future endeavors!
