Assessing Generalization Capability of Convolutional Neural Networks

Anonymous

Nov 03, 2017 (modified: Nov 03, 2017) ICLR 2018 Conference Blind Submission readers: everyone Show Bibtex
  • Abstract: In this paper, we assess the capability of Convolutional Neural Networks (CNNs) in generalizing to images with different distribution than the training data, specifically images with similar shapes but different colors. We show that, although CNNs do not intrinsically classify objects based on their shapes, they can learn to do so when trained with enough number of images with the same shape and different colors. In experiments, we use original and negative images of training data as such images. Through systematic experiments, we investigate the role of training data, model architecture, initialization and regularization techniques on model generalization to negative images. We conclude that although CNNs can memorize any training data, they only learn and generalize the structures.
  • Keywords: Deep learning, Convolutional Neural Networks, Generalization

Loading