Schematic illustrating our pipeline for producing data summaries for a given astrophysical parameter combination (adapted from Prelogović & Mesinger 2023). Starting from a realization of the cosmological signal simulated with 21cmFAST, we remove the mean at each redshift and add a telescope noise realization corresponding to a 1000 h observation with SKA1-Low. The resulting light cone is compressed into different summaries. Explicitly defined summaries are marked in blue, while neural network summaries are marked in red. This realization is computed at the fiducial parameter set, which was used only for IMNN training and visualizations of the summaries. See the main text for more details.