In data management, indices are used to retrieve data in faster than linear complexity. A representative of such index could be B-tree, which is used for classical, attribute data; to handle more complex data such as high-dimensional descriptors extracted from raw images or videos where no attributes are used to store or retrieve the data, index structures such as M-tree and M-index have been developed. In 2018, a paper called The Case for Learned Index Structures has been published, arguing for benefits of replacing the standard index structures for attribute data with deep-learning models. Recently, a small research team at FI MU has applied the approach to complex data in a supervised manner. The subject of this thesis is to implement an Unsupervised Learned Index Structure, and to examine its searching performance with comparison to M-tree, M-index and LMI on a number of different datasets (CoPhIR and Profiset). Unsupervised LMI will be implemented in Python, utilizing the Tensorflow and scikit-learn.