| Abstract [eng] |
Master thesis researches the problem of interpreting deep neural networks for image recognition. Particularly interpretation methods which are model agnostic, because they can be applied to broad array of machine learning models. Literature analysis chapter looks into relevant interpretation methods and motivate choosing of this subject. LIME “Local Interpretable Model-agnostic Explanations” method [RSG16a] was chosen for further analysis because of its relevancy and wide spread usage. Thesis analysis two parts of this method: segmenting an image into useful segments and extending LIME's submodular pick algorithm to be useful with image data. Thesis evaluates six possible image segmentation algorithms and makes recommendations which ones are most beneficial to use with LIME. Submodular pick algorithm was extended by adding clustering step to make it useful for image data. Also a new alternative function which values Submodular pick tries to to maximize was proposed. This alternative function solves some problematic aspects of the original function. |
