Semi-supervised robust deep neural networks for multi-label image classification. (April 2020)

• Record Type: Journal Article
• Authors: Cevikalp, Hakan; Benligiray, Burak; Gerek, Omer Nezih
• Is Part Of: Pattern recognition. Volume 100(2020:Apr.)
• Publication Date: 2020-04
• Subjects: Multi-label classification -- Semi-supervised learning -- Ramp loss -- Image classification -- Deep learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2019.107164
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Large-scale data includes many noisily labeled and unlabeled examples. With traditional methods, incorrect labels cause a very large loss, which misleads the training process. We propose to use the robust ramp function for multi-label image classification, which allows us to utilize examples with noisy labels successfully. In addition, the added robustness makes the method more robust against errors made during propagating labels to unlabeled examples, which allows robust semi-supervised training. Abstract: This paper introduces a robust method for semi-supervised training of deep neural networks for multi-label image classification. To this end, a ramp loss is utilized since it is more robust against noisy and incomplete image labels compared to the classic hinge loss. The proposed method allows for learning from both labeled and unlabeled data in a semi-supervised setting. This is achieved by propagating labels from the labeled images to their unlabeled neighbors in the feature space. Using a robust loss function becomes crucial here, as the initial label propagations may include many errors, which degrades the performance of non-robust loss functions. In contrast, the proposed robust ramp loss restricts extreme penalties from the samples with incorrect labels, and the label assignment improves in each iteration and contributes to the learning process. The proposed method achieves state-of-the-art results in semi-supervised learning experiments on the CIFAR-10 and STL-10 datasets, and comparable results to the state-of the-art in supervised learning experiments on the NUS-WIDE and MS-COCO datasets. Experimental results also verify that our proposed method is more robust against noisy image labels as expected.