Dark matter maps, and their temporal evolution, carry information about the laws of physics and can be used to measure parameters of cosmological models. Typically, simple human-designed features, such as the power spectra, are used to analyse these maps. However, as the dark matter maps contain very rich and complex patterns, such as halos, filaments, and voids, and the simple statistics do not extract all the information contained in these patterns. I used the convolutional neural networks to automatically design relevant features and combine them in an optimal way to maximise the information gain from dark matter maps. I present the first AI-based analysis of dark matter maps, using the KiDS-450 dataset, which results in 40% more precise measurements than with the standard approach. I discuss the feasibility of the AI-based analysis for future cosmological datasets.