Biologically Feasible Generative Neural Architectures and Evolutionary Learning in Simple Visual Environments

Abstract

Informative representations are essential in the learning processes of artificial and biological systems due to the potential to identify characteristic patterns, general types or concepts in the sensory environments. In this work, we examined generative representations of several sets of images, such as basic geometric shapes and handwritten digits, obtained with biologically feasible generative neural models in the process of unsupervised generative learning. A neural architecture based on bi-directional synaptic connection, equivalent in training and processing of sensory inputs to a near-symmetrical feed-forward generative neural network was described. It was demonstrated that conceptual representations with good decoupling of concept regions can be produced with generative models of minimal complexity; and that incremental variations of generative architecture fully compatible with the objectives of biological feasibility and learning success can produce architectures with improved ability to learn data of increasing conceptual complexity, including realistic images such as handwritten digits. The results of this work indicate that incremental adaptation of neural architectures with the incentive to improve learning success can be a natural pathway for the emergence of an ability of successful conceptual modeling of sensory environments of increasing complexity.