Formula-driven Supervised Learning with Recursive Tiling Patterns

Can convolutional neural networks pre-trained without natural images be used to assist natural image understanding? Formula-Driven Supervised Learning (FDSL) automatically generates image patterns and their category labels by assigning a well-organized formula. Due to the characteristics of not using natural images in pre-training phase, FDSL is expected to develop a trustworthy vision-based system in terms of human-annotation-free, fairer and more transparent datasets. In this paper, we propose TileDB which consists of recursive tiling patterns in the whole image and evaluates the family of FDSL such as the datasets consist of Perlin noise and Bezier curves. Experimental results show that our proposed TileDB pre-trained model performs much better than models trained from scratch, surpasses a similar self-supervised learning (SSL), and performs similarly to the models pre-trained with 100k-order natural image datasets such as ImageNet-100 and Places-30. By comparing to the FractalDB pre-trained model, the TileDB pre-trained model achieves better performances in a compact dataset (< 1,000 categories). Moreover, the image representation trained on TileDB can extract similar features to the ImageNet pre-trained model even though the training images are non-trivially different.