Hi Marc,
Autoencoders are often used in unsupervised learning tasks where the goal is to learn a compressed, encoded representation of the input data. The training process doesn't require labeled data; instead, the network learns by trying to reproduce its input at its output. Hence, the traditional validation and testing processes, which rely on comparing the model's predictions to actual labels, aren't applicable in the same way they are for supervised learning tasks. This could be one fo the main reasons behind the default "dividetrain" setting in the trainAutoencoder function as it is an auto-training function.
You can define and train a custom autoencoder according to your task. Refer to the following example which shows how to train a deep learning variational autoencoder (VAE) to generate images: https://www.mathworks.com/help/deeplearning/ug/train-a-variational-autoencoder-vae-to-generate-images.html
