Machine Perception and Learning from the Evolutionary Viewpoint: The Hyperball Algorithms

We present a set of highly parallel (fine grain), general pattern classification algorithms aimed at mimicking pattern identification, classification and learning skills of animals. We cover first an idealized case of supervised learning from an infallible expert, followed by the realistic cases of learning from fallible experts as well as the case of autonomous (i.e. self-supervised) learning. To ensure wide range of applicability in our algorithms we use only the basic concepts of mathematics: set theory and theory of metric spaces. The proposed methodology allows for creation of expert systems capable of continuous learning and forgetting of less useful facts, and so to maintain their adaptation to an environment evolving sufficiently slowly. As proof of concept we offer a sketch of automated deep sky observation in search for unknown objects together with a detailed solution to the problem of automated mineral identification in petrographic thin sections. These illustration problems seem intractable using traditional means.