engleski

Identification of 1D-Signal Types Using Unsupervised Deep Learning

During de novo genome assembly process, certain types of sequenced reads can cause problems during genome reconstruction. Goal of this thesis is to learn more about possible types of reads and classification of those reads using unsupervised learning. Coverage graphs of reads are generated using read overlaps and those coverage graphs are further analysed. Autoencoder is used to compress the signal, i.e. the coverage graph, and clustering algorithm is then applied to the compressed data. Variational and denoising autoencoders along with k-means and spectral clustering algorithms are used. Visualisation of found clusters is performed along with semantic analysis. Signal classification quality using unsupervised learning is estimated.

bioinformatics, unsupervised learning, deep learning, autoencoders

nije evidentirano