Pre-training Without Natural Images

Is it possible to use convolutional neural networks pre-trained without any natural images assist image understanding? The paper proposes a novel concept, Formula-driven Supervised Learning. We automatically generate patterns and their category labels by assigning fractals, which are based on law existing in the background knowledge of real world. Theoretically, generated instead pre-training phase allows us an infinite scale dataset labeled images. Although models with proposed Fractal DataBase (FractalDB), database images, does not necessarily outperform human annotated datasets at all settings, we able partially surpass accuracy ImageNet/Places models. representation FractalDB captures unique feature visualization layers attentions.