On the Limitation of Convolutional Neural Networks in Recognizing Negative Images
This reveals a limitation in CNN training methods for computer vision, potentially impacting robustness in real-world applications, though it is an incremental analysis of existing models.
The paper investigates whether CNNs can learn semantic concepts by testing them on negative images, which humans can classify correctly, and finds that accuracy drops significantly when trained on regular images and tested on negatives, suggesting poor generalization.
Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance on a variety of computer vision tasks, particularly visual classification problems, where new algorithms reported to achieve or even surpass the human performance. In this paper, we examine whether CNNs are capable of learning the semantics of training data. To this end, we evaluate CNNs on negative images, since they share the same structure and semantics as regular images and humans can classify them correctly. Our experimental results indicate that when training on regular images and testing on negative images, the model accuracy is significantly lower than when it is tested on regular images. This leads us to the conjecture that current training methods do not effectively train models to generalize the concepts. We then introduce the notion of semantic adversarial examples - transformed inputs that semantically represent the same objects, but the model does not classify them correctly - and present negative images as one class of such inputs.