Learned and Native Concepts in Latent Representations of Terrain Images

Abstract

Research into generative representations of complex data is a rapidly expanding field in machine learning. In this work we propose and evaluate a process of production and analysis of informative low-dimensional latent representations of real-world images with neural network models of unsupervised generative learning. A model of convolutional autoencoder based on VGG-16 architecture was used to produce low-dimensional generative representations of two datasets of aerial images and the characteristics of distributions of several classes of images of terrain studied. An analysis of latent distributions of terrain classes demonstrated a landscape of compact density structures for most studied classes with good separation of concept regions. The results of this work can be used in developing methods of effective learning in problems and environments with a deficit of labeled data based on the concept-sensitive structure in the latent representations that emerges in the process of unsupervised generative self-learning.