Interpreting and Controlling Latent Space Parameters of Autoencoders for TAIGA-HiSCORE Data

We propose using the latent space (LS) of an autoencoder (AE) as a compressed representation of input data in gamma astronomy. The AE learns compressed representations directly from data and its LS dimensionality N is tunable, enabling optimization for specific tasks. To ensure interpretability and control, we train two dedicated networks: one maps the LS to physical parameters, while the other performs the inverse mapping. The approach was validated on TAIGA-HiSCORE simulated data, demonstrating promising results.