Can Jellyfish Dream? Conceptual Representations in Unsupervised Generative Learning

Representations play essential role in learning of artificial and biologic systems by identifying characteristic patterns in the observed environment. In this work we examine unsupervised latent representations of image data with a collective of generative neural network models. A convolutional autoencoder with strong redundancy reduction was used to create low-dimensional latent representations of a dataset of geometrical shapes. The structure of the resulting latent representations was studied comprehensively with several methods, including density clustering, direct visualization, generative probing and scanning. It was demonstrated that latent representations have high level of consistency between individual learners and a well-defined geometrical and topological structure correlated with principal patterns in the observable data. The results of this work support the hypothesis that conceptual latent representations can emerge naturally in unsupervised generative learning under certain essential constraints and can be a natural platform for emergence of some intelligent functions and behaviors.