In present work we explore the applicability of autoencoders (AEs) as a tool to extract metric properties of a physical space from the neural activity signal of the hippocampal CA1 place cells, while the object (mouse) was exploring novel environments, flat arenas of different shapes. We show that an AE as shallow as 3+3 layers (3 layers both in the encoder and the decoder) can be trained to encode a cognitive map in its latent space—by introducing a simple additional loss of (supervised) metric reconstruction. The introduction of this loss term significantly changes the importance distribution of the AE encoder’s last-layer neurons, making their statistical properties similar to those found in natural neural networks.