Synergy between traditional classification and classification based on negative features in deep convolutional neural ne
- PDF / 469,143 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 13 Downloads / 167 Views
(0123456789().,-volV)(0123456789(). ,- volV)
ORIGINAL ARTICLE
Synergy between traditional classification and classification based on negative features in deep convolutional neural networks Nemanja Milosˇevic´1
•
Milosˇ Rackovic´1
Received: 5 June 2020 / Accepted: 2 November 2020 Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract In recent times, convolutional neural networks became an irreplaceable tool in many different machine learning applications, especially in image classification. On the other hand, new research about robustness and susceptibility of these models to different adversarial attacks has emerged. With the rise in usage and widespread adoption of these models, it is very important to make them suitable for critical applications. In our previous work, we experimented with a new type of learning applicable to all convolutional neural networks: classification based on missing (low-impact) features. In the case of partial inputs/image occlusion, we have shown that our new method creates models that are more robust and perform better when compared to traditional models of the same architecture. In this paper, we explore an interesting characteristic of our newly developed models in that while we see a general increase in validation accuracy, we also lose some important knowledge. We propose one solution to overcome this problem and validate our assumptions against CIFAR-10 image classification dataset. Keywords Neural networks Machine learning Convolutional neural networks Machine learning robustness Computer vision
1 Introduction and motivation Convolutional neural networks are widely used algorithms for image processing [10, 22]. These algorithms can be used for many tasks like image classification, image segmentation, object detection, etc., and they have become an irreplaceable tool in many systems (e.g., face recognition software, self-driving cars [6]). As the adoption of these algorithms rises, they are used in critical systems where robustness and interpretability become important factors [4, 7, 15]. For robustness specifically, in our previous work [25], we introduced a new family of convolutional neural networks (referred as negative or inverse networks) which try to classify instances based on the negative features of the input samples. We define negative features as features & Nemanja Milosˇevic´ [email protected] Milosˇ Rackovic´ [email protected] 1
Department of Mathematics and Informatics, Faculty of Sciences, University of Novi Sad, Novi Sad, Serbia
we know exist (from many other samples in the dataset) but which are not present in the example we are currently observing. These features can also be seen as low-impact, low-importance features. In our work, we already tested and empirically proved that our way of feature representation is suitable for training neural network models without any loss in accuracy. Moreover, we tested our models in one difficult scenario where algorithm robustness is important (object occlusion/partial input) and obs
Data Loading...
