Abstract <p>One of the main tasks of studying gamma-ray sources using Imaging Atmospheric Cherenkov telescopes (IACTs) is to reconstruct their energy spectrum. In the traditional method of gamma-ray energy reconstruction, the physical characteristics are reconstructed by analyzing the Hillas parameters obtained from EAS images on the IACTs and being derivatives of the statistical moments of the image. We present an alternative method for processing IACT data based on the formation of another set of parameters—essential features extracted from EAS images using neural networks—autoencoders. The parameters of the latent space of autoencoders, which contain information about the image in a compressed form, are used as essential features. In this paper, the problem of reconstructing the energy of a primary gamma quantum using deep learning methods (both a direct method of reconstruction directly from images using convolutional neural networks and by extracting essential features using an autoencoder) is considered, which are compared with the traditional method based on the analysis of Hillas parameters. TAIGA-IACT Monte Carlo data were used as the analyzed data. The dependence of the quality of gamma-ray energy restoration on the latent space dimensionality was investigated. Comparison of traditional approach with neural network methods showed that both the convolutional network and autoencoders reconstruct gamma-ray energy with an accuracy that is not inferior to the classical method at low energies and have a significantly smaller energy reconstruction error at high energies.</p>