Training Convolutional Neural Networks with Natural Principle

Abstract

The paper introduces a framework for Formula-Driven Supervised Learning (FDSL) which automatically generates image patterns and their image labels for creating a large-scale image dataset. We mainly focus on Fractal DataBase (FractalDB) which consists of fractal geometry existing around our real world. That is, one of the important natural principle enables to train convolutional neural networks without any natural images taken by a camera and human-annotated labels in a pre-training phase.