Pre-training without Natural Images
We would like to replace Supervised/Self-supervised Learning!
Is it possible to use convolutional neural networks pre-trainedwithout any natural images to assist natural image understanding? Thepaper proposes a novel concept, Formula-driven Supervised Learning.We automatically generate image patterns and their category labels byassigning fractals, which are based on a natural law existing in the back-ground knowledge of the real world. Theoretically, the use of automati-cally generated images instead of natural images in the pre-training phaseallows us to generate an infinite scale dataset of labeled images. Althoughthe models pre-trained with the proposed Fractal DataBase (FractalDB), a database without natural images, does not necessarily outperform mod-els pre-trained with human annotated datasets at all settings, we are ableto partially surpass the accuracy of ImageNet/Places pre-trained mod-els. The image representation with the proposed FractalDB captures aunique feature in the visualization of convolutional layers and attentions.
ACCV 2020 Best Paper Honorable Mention Award
Hirokatsu Kataoka (AIST), Kazushige Okayasu (AIST), Asato Matsumoto (AIST), Eisuke Yamagata (TITech), Ryosuke Yamada (AIST), Nakamasa Inoue (TITech), Akio Nakamura (TDU), Yutaka Satoh (AIST)