Unsupervised Generative Learning and Native Explanatory Frameworks

A framework of native concepts emergent in unsupervised generative training under the constraints of redundancy reduction and generative accuracy was observed and investigated with an unsupervised generative neural network model and real-world image data. Characteristics of concept distributions in the latent representations were measured and possibility of effective learning with identified density structure demonstrated. We discuss the potential of using frameworks of native information clusters in the effective latent representations of learning models as a natural platform for explanation of learning processes in machine and biological systems based on the relations and criteria of native similarity that form in the process of generative learning under certain constraints. This approach can be general, intuitive and independent of specific architectures and types of data.