Comparison of different autoencoder variations vs state-of-the-art methods on Sim16 (9 clusters), the optimal value of neighbours for Isomap was found as 20 through grid search.

The metrics indicate that several variations of autoencoders and the Isomap algorithm are able to outperform conventional algorithms such as PCA or ICA for the Sim16 dataset that contains a moderate number of clusters. Moreover, two of autoencoder variants outperform Isomap with regard to the external metrics implying that these variants offer more separable features for the clustering algorithm, the values indicating the highest performance for each metric have been bolded.